diff --git a/MindIE/MindIE-Torch/built-in/foundation/CogVideoX-5b/README.md b/MindIE/MultiModal/CogVideoX-5B/README.md similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/CogVideoX-5b/README.md rename to MindIE/MultiModal/CogVideoX-5B/README.md diff --git a/MindIE/MindIE-Torch/built-in/foundation/CogVideoX-5b/cogvideox_5b/__init__.py b/MindIE/MultiModal/CogVideoX-5B/cogvideox_5b/__init__.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/CogVideoX-5b/cogvideox_5b/__init__.py rename to MindIE/MultiModal/CogVideoX-5B/cogvideox_5b/__init__.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/CogVideoX-5b/cogvideox_5b/models/__init__.py b/MindIE/MultiModal/CogVideoX-5B/cogvideox_5b/models/__init__.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/CogVideoX-5b/cogvideox_5b/models/__init__.py rename to MindIE/MultiModal/CogVideoX-5B/cogvideox_5b/models/__init__.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/CogVideoX-5b/cogvideox_5b/models/activations.py b/MindIE/MultiModal/CogVideoX-5B/cogvideox_5b/models/activations.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/CogVideoX-5b/cogvideox_5b/models/activations.py rename to MindIE/MultiModal/CogVideoX-5B/cogvideox_5b/models/activations.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/CogVideoX-5b/cogvideox_5b/models/attention.py b/MindIE/MultiModal/CogVideoX-5B/cogvideox_5b/models/attention.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/CogVideoX-5b/cogvideox_5b/models/attention.py rename to MindIE/MultiModal/CogVideoX-5B/cogvideox_5b/models/attention.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/CogVideoX-5b/cogvideox_5b/models/attention_processor.py b/MindIE/MultiModal/CogVideoX-5B/cogvideox_5b/models/attention_processor.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/CogVideoX-5b/cogvideox_5b/models/attention_processor.py rename to MindIE/MultiModal/CogVideoX-5B/cogvideox_5b/models/attention_processor.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/CogVideoX-5b/cogvideox_5b/models/embeddings.py b/MindIE/MultiModal/CogVideoX-5B/cogvideox_5b/models/embeddings.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/CogVideoX-5b/cogvideox_5b/models/embeddings.py rename to MindIE/MultiModal/CogVideoX-5B/cogvideox_5b/models/embeddings.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/CogVideoX-5b/cogvideox_5b/models/normalization.py b/MindIE/MultiModal/CogVideoX-5B/cogvideox_5b/models/normalization.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/CogVideoX-5b/cogvideox_5b/models/normalization.py rename to MindIE/MultiModal/CogVideoX-5B/cogvideox_5b/models/normalization.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/CogVideoX-5b/cogvideox_5b/models/transformers/__init__.py b/MindIE/MultiModal/CogVideoX-5B/cogvideox_5b/models/transformers/__init__.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/CogVideoX-5b/cogvideox_5b/models/transformers/__init__.py rename to MindIE/MultiModal/CogVideoX-5B/cogvideox_5b/models/transformers/__init__.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/CogVideoX-5b/cogvideox_5b/models/transformers/cogvideox_transformer_3d.py b/MindIE/MultiModal/CogVideoX-5B/cogvideox_5b/models/transformers/cogvideox_transformer_3d.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/CogVideoX-5b/cogvideox_5b/models/transformers/cogvideox_transformer_3d.py rename to MindIE/MultiModal/CogVideoX-5B/cogvideox_5b/models/transformers/cogvideox_transformer_3d.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/CogVideoX-5b/cogvideox_5b/pipelines/__init__.py b/MindIE/MultiModal/CogVideoX-5B/cogvideox_5b/pipelines/__init__.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/CogVideoX-5b/cogvideox_5b/pipelines/__init__.py rename to MindIE/MultiModal/CogVideoX-5B/cogvideox_5b/pipelines/__init__.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/CogVideoX-5b/cogvideox_5b/pipelines/pipeline_cogvideox.py b/MindIE/MultiModal/CogVideoX-5B/cogvideox_5b/pipelines/pipeline_cogvideox.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/CogVideoX-5b/cogvideox_5b/pipelines/pipeline_cogvideox.py rename to MindIE/MultiModal/CogVideoX-5B/cogvideox_5b/pipelines/pipeline_cogvideox.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/CogVideoX-5b/cogvideox_5b/pipelines/pipeline_output.py b/MindIE/MultiModal/CogVideoX-5B/cogvideox_5b/pipelines/pipeline_output.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/CogVideoX-5b/cogvideox_5b/pipelines/pipeline_output.py rename to MindIE/MultiModal/CogVideoX-5B/cogvideox_5b/pipelines/pipeline_output.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/CogVideoX-5b/cogvideox_5b/utils/__init__.py b/MindIE/MultiModal/CogVideoX-5B/cogvideox_5b/utils/__init__.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/CogVideoX-5b/cogvideox_5b/utils/__init__.py rename to MindIE/MultiModal/CogVideoX-5B/cogvideox_5b/utils/__init__.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/CogVideoX-5b/cogvideox_5b/utils/parallel_mgr.py b/MindIE/MultiModal/CogVideoX-5B/cogvideox_5b/utils/parallel_mgr.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/CogVideoX-5b/cogvideox_5b/utils/parallel_mgr.py rename to MindIE/MultiModal/CogVideoX-5B/cogvideox_5b/utils/parallel_mgr.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/CogVideoX-5b/cogvideox_5b/utils/parallel_state.py b/MindIE/MultiModal/CogVideoX-5B/cogvideox_5b/utils/parallel_state.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/CogVideoX-5b/cogvideox_5b/utils/parallel_state.py rename to MindIE/MultiModal/CogVideoX-5B/cogvideox_5b/utils/parallel_state.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/CogVideoX-5b/inference.py b/MindIE/MultiModal/CogVideoX-5B/inference.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/CogVideoX-5b/inference.py rename to MindIE/MultiModal/CogVideoX-5B/inference.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/CogVideoX-5b/pta_plugin/CMakeLists.txt b/MindIE/MultiModal/CogVideoX-5B/pta_plugin/CMakeLists.txt similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/CogVideoX-5b/pta_plugin/CMakeLists.txt rename to MindIE/MultiModal/CogVideoX-5B/pta_plugin/CMakeLists.txt diff --git a/MindIE/MindIE-Torch/built-in/foundation/CogVideoX-5b/pta_plugin/build.sh b/MindIE/MultiModal/CogVideoX-5B/pta_plugin/build.sh similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/CogVideoX-5b/pta_plugin/build.sh rename to MindIE/MultiModal/CogVideoX-5B/pta_plugin/build.sh diff --git a/MindIE/MindIE-Torch/built-in/foundation/CogVideoX-5b/pta_plugin/extension_ops.cpp b/MindIE/MultiModal/CogVideoX-5B/pta_plugin/extension_ops.cpp similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/CogVideoX-5b/pta_plugin/extension_ops.cpp rename to MindIE/MultiModal/CogVideoX-5B/pta_plugin/extension_ops.cpp diff --git a/MindIE/MindIE-Torch/built-in/foundation/CogVideoX-5b/pta_plugin/test/test_rope.py b/MindIE/MultiModal/CogVideoX-5B/pta_plugin/test/test_rope.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/CogVideoX-5b/pta_plugin/test/test_rope.py rename to MindIE/MultiModal/CogVideoX-5B/pta_plugin/test/test_rope.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/CogVideoX-5b/requirements.txt b/MindIE/MultiModal/CogVideoX-5B/requirements.txt similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/CogVideoX-5b/requirements.txt rename to MindIE/MultiModal/CogVideoX-5B/requirements.txt diff --git a/MindIE/MindIE-Torch/built-in/foundation/cogview3/README.md b/MindIE/MultiModal/CogView3-Plus-3B/README.md similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/cogview3/README.md rename to MindIE/MultiModal/CogView3-Plus-3B/README.md diff --git a/MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/__init__.py b/MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/__init__.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/__init__.py rename to MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/__init__.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/layers/__init__.py b/MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/layers/__init__.py similarity index 99% rename from MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/layers/__init__.py rename to MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/layers/__init__.py index 09760b9fd0e5838eba0865adbac38a31d59768a1..602ad432a01cefe3824b16f4ee90ce56f18c3aab 100644 --- a/MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/layers/__init__.py +++ b/MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/layers/__init__.py @@ -1,3 +1,3 @@ -from .normalization import CogView3PlusAdaLayerNormZeroTextImage, AdaLayerNormContinuous -from .embeddings import CogView3CombinedTimestepSizeEmbeddings, CogView3PlusPatchEmbed +from .normalization import CogView3PlusAdaLayerNormZeroTextImage, AdaLayerNormContinuous +from .embeddings import CogView3CombinedTimestepSizeEmbeddings, CogView3PlusPatchEmbed from .linear import QKVLinear \ No newline at end of file diff --git a/MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/layers/embeddings.py b/MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/layers/embeddings.py similarity index 97% rename from MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/layers/embeddings.py rename to MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/layers/embeddings.py index 129384dffc0fb92090012e46a8d89e7f4d0a1d6c..fc2d3101eb92b43f95ef51a4e732b5747286f8e9 100644 --- a/MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/layers/embeddings.py +++ b/MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/layers/embeddings.py @@ -1,304 +1,304 @@ -# Copyright 2024 The HuggingFace Team. All rights reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# https://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -import math -from typing import Optional - -import torch -from torch import nn -from diffusers.models.activations import get_activation - - -def get_timestep_embedding( - timesteps: torch.Tensor, - embedding_dim: int, - flip_sin_to_cos: bool = False, - downscale_freq_shift: float = 1, - max_period: int = 10000, -): - half_dim = embedding_dim // 2 - exponent = -math.log(max_period) * torch.arange( - start=0, end=half_dim, dtype=torch.float32, device=timesteps.device - ) - exponent = exponent / (half_dim - downscale_freq_shift) - - emb = torch.exp(exponent) - emb = timesteps[:, None].float() * emb[None, :] - - # concat sine and cosine embeddings - emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=-1) - - # flip sine and cosine embeddings - if flip_sin_to_cos: - emb = torch.cat([emb[:, half_dim:], emb[:, :half_dim]], dim=-1) - - # zero pad - if embedding_dim % 2 == 1: - emb = torch.nn.functional.pad(emb, (0, 1, 0, 0)) - return emb - - -def get_2d_sincos_pos_embed( - embed_dim, - grid_size, - interpolation_scale=1.0, - base_size=16, -): - if isinstance(grid_size, int): - grid_size = (grid_size, grid_size) - - grid_h = ( - torch.arange(grid_size[0], dtype=torch.float32) - / (grid_size[0] / base_size) - / interpolation_scale - ) - grid_w = ( - torch.arange(grid_size[1], dtype=torch.float32) - / (grid_size[1] / base_size) - / interpolation_scale - ) - grid = torch.meshgrid(grid_w, grid_h, indexing="xy") # here w goes first - grid = torch.stack(grid, dim=0) - - grid = grid.reshape([2, 1, grid_size[1], grid_size[0]]) - pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid) - return pos_embed - - -def get_2d_sincos_pos_embed_from_grid(embed_dim, grid): - r""" - This function generates 2D sinusoidal positional embeddings from a grid. - - Args: - embed_dim (`int`): The embedding dimension. - grid (`torch.Tensor`): Grid of positions with shape `(H * W,)`. - - Returns: - `torch.Tensor`: The 2D sinusoidal positional embeddings with shape `(H * W, embed_dim)` - """ - if embed_dim % 2 != 0: - raise ValueError("embed_dim must be divisible by 2") - - # use half of dimensions to encode grid_h - emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0]) # (H*W, D/2) - emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1]) # (H*W, D/2) - - emb = torch.concat([emb_h, emb_w], dim=1) # (H*W, D) - return emb - - -def get_1d_sincos_pos_embed_from_grid(embed_dim, pos): - """ - This function generates 1D positional embeddings from a grid. - - Args: - embed_dim (`int`): The embedding dimension `D` - pos (`torch.Tensor`): 1D tensor of positions with shape `(M,)` - - Returns: - `torch.Tensor`: Sinusoidal positional embeddings of shape `(M, D)`. - """ - if embed_dim % 2 != 0: - raise ValueError("embed_dim must be divisible by 2") - - omega = torch.arange(embed_dim // 2, device=pos.device, dtype=torch.float64) - omega /= embed_dim / 2.0 - omega = 1.0 / 10000**omega # (D/2,) - - pos = pos.reshape(-1) # (M,) - out = torch.outer(pos, omega) # (M, D/2), outer product - - emb_sin = torch.sin(out) # (M, D/2) - emb_cos = torch.cos(out) # (M, D/2) - - emb = torch.concat([emb_sin, emb_cos], dim=1) # (M, D) - return emb - - -class Timesteps(nn.Module): - def __init__(self, num_channels: int, flip_sin_to_cos: bool, downscale_freq_shift: float): - super().__init__() - self.num_channels = num_channels - self.flip_sin_to_cos = flip_sin_to_cos - self.downscale_freq_shift = downscale_freq_shift - - def forward(self, timesteps): - t_emb = get_timestep_embedding( - timesteps, - self.num_channels, - flip_sin_to_cos=self.flip_sin_to_cos, - downscale_freq_shift=self.downscale_freq_shift, - ) - return t_emb - - -class TimestepEmbedding(nn.Module): - def __init__( - self, - in_channels: int, - time_embed_dim: int, - act_fn: str = "silu", - out_dim: int = None, - post_act_fn: Optional[str] = None, - cond_proj_dim=None, - sample_proj_bias=True, - ): - super().__init__() - - self.linear_1 = nn.Linear(in_channels, time_embed_dim, sample_proj_bias) - - if cond_proj_dim is not None: - self.cond_proj = nn.Linear(cond_proj_dim, in_channels, bias=False) - else: - self.cond_proj = None - - self.act = get_activation(act_fn) - - if out_dim is not None: - time_embed_dim_out = out_dim - else: - time_embed_dim_out = time_embed_dim - self.linear_2 = nn.Linear(time_embed_dim, time_embed_dim_out, sample_proj_bias) - - if post_act_fn is None: - self.post_act = None - else: - self.post_act = get_activation(post_act_fn) - - def forward(self, sample, condition=None): - if condition is not None: - sample = sample + self.cond_proj(condition) - sample = self.linear_1(sample) - - if self.act is not None: - sample = self.act(sample) - - sample = self.linear_2(sample) - - if self.post_act is not None: - sample = self.post_act(sample) - return sample - - -class PixArtAlphaTextProjection(nn.Module): - """ - Projects caption embeddings. Also handles dropout for classifier-free guidance. - """ - - def __init__(self, in_features, hidden_size, out_features=None, act_fn="gelu_tanh"): - super().__init__() - if out_features is None: - out_features = hidden_size - self.linear_1 = nn.Linear(in_features=in_features, out_features=hidden_size, bias=True) - if act_fn == "gelu_tanh": - self.act_1 = nn.GELU(approximate="tanh") - elif act_fn == "silu": - self.act_1 = nn.SiLU() - else: - raise ValueError(f"Unknown activation function: {act_fn}") - self.linear_2 = nn.Linear(in_features=hidden_size, out_features=out_features, bias=True) - - def forward(self, caption): - hidden_states = self.linear_1(caption) - hidden_states = self.act_1(hidden_states) - hidden_states = self.linear_2(hidden_states) - return hidden_states - - -class CogView3CombinedTimestepSizeEmbeddings(nn.Module): - def __init__(self, embedding_dim: int, condition_dim: int, pooled_projection_dim: int, timesteps_dim: int = 256): - super().__init__() - - self.time_proj = Timesteps(num_channels=timesteps_dim, flip_sin_to_cos=True, downscale_freq_shift=0) - self.condition_proj = Timesteps(num_channels=condition_dim, flip_sin_to_cos=True, downscale_freq_shift=0) - self.timestep_embedder = TimestepEmbedding(in_channels=timesteps_dim, time_embed_dim=embedding_dim) - self.condition_embedder = PixArtAlphaTextProjection(pooled_projection_dim, embedding_dim, act_fn="silu") - - def forward( - self, - timestep: torch.Tensor, - original_size: torch.Tensor, - target_size: torch.Tensor, - crop_coords: torch.Tensor, - hidden_dtype: torch.dtype, - ) -> torch.Tensor: - timesteps_proj = self.time_proj(timestep) - - original_size_proj = self.condition_proj(original_size.flatten()).view(original_size.size(0), -1) - crop_coords_proj = self.condition_proj(crop_coords.flatten()).view(crop_coords.size(0), -1) - target_size_proj = self.condition_proj(target_size.flatten()).view(target_size.size(0), -1) - - condition_proj = torch.cat([original_size_proj, crop_coords_proj, target_size_proj], dim=1) - - timesteps_emb = self.timestep_embedder(timesteps_proj.to(dtype=hidden_dtype)) # (B, embedding_dim) - condition_emb = self.condition_embedder(condition_proj.to(dtype=hidden_dtype)) # (B, embedding_dim) - - conditioning = timesteps_emb + condition_emb - return conditioning - - -class CogView3PlusPatchEmbed(nn.Module): - def __init__( - self, - in_channels: int = 16, - hidden_size: int = 2560, - patch_size: int = 2, - text_hidden_size: int = 4096, - pos_embed_max_size: int = 128, - ): - super().__init__() - self.in_channels = in_channels - self.hidden_size = hidden_size - self.patch_size = patch_size - self.text_hidden_size = text_hidden_size - self.pos_embed_max_size = pos_embed_max_size - # Linear projection for image patches - self.proj = nn.Linear(in_channels * patch_size**2, hidden_size) - - # Linear projection for text embeddings - self.text_proj = nn.Linear(text_hidden_size, hidden_size) - - pos_embed = get_2d_sincos_pos_embed( - hidden_size, pos_embed_max_size, base_size=pos_embed_max_size - ) - pos_embed = pos_embed.reshape(pos_embed_max_size, pos_embed_max_size, hidden_size) - self.register_buffer("pos_embed", pos_embed.float(), persistent=False) - - def forward(self, hidden_states: torch.Tensor, encoder_hidden_states: torch.Tensor) -> torch.Tensor: - batch_size, channel, height, width = hidden_states.shape - - if height % self.patch_size != 0 or width % self.patch_size != 0: - raise ValueError("Height and width must be divisible by patch size") - - height = height // self.patch_size - width = width // self.patch_size - hidden_states = hidden_states.view(batch_size, channel, height, self.patch_size, width, self.patch_size) - hidden_states = hidden_states.permute(0, 2, 4, 1, 3, 5).contiguous() - hidden_states = hidden_states.view(batch_size, height * width, channel * self.patch_size * self.patch_size) - - # Project the patches - hidden_states = self.proj(hidden_states) - encoder_hidden_states = self.text_proj(encoder_hidden_states) - hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1) - - # Calculate text_length - text_length = encoder_hidden_states.shape[1] - - image_pos_embed = self.pos_embed[:height, :width].reshape(height * width, -1) - text_pos_embed = torch.zeros( - (text_length, self.hidden_size), dtype=image_pos_embed.dtype, device=image_pos_embed.device - ) - pos_embed = torch.cat([text_pos_embed, image_pos_embed], dim=0)[None, ...] - +# Copyright 2024 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# https://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import math +from typing import Optional + +import torch +from torch import nn +from diffusers.models.activations import get_activation + + +def get_timestep_embedding( + timesteps: torch.Tensor, + embedding_dim: int, + flip_sin_to_cos: bool = False, + downscale_freq_shift: float = 1, + max_period: int = 10000, +): + half_dim = embedding_dim // 2 + exponent = -math.log(max_period) * torch.arange( + start=0, end=half_dim, dtype=torch.float32, device=timesteps.device + ) + exponent = exponent / (half_dim - downscale_freq_shift) + + emb = torch.exp(exponent) + emb = timesteps[:, None].float() * emb[None, :] + + # concat sine and cosine embeddings + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=-1) + + # flip sine and cosine embeddings + if flip_sin_to_cos: + emb = torch.cat([emb[:, half_dim:], emb[:, :half_dim]], dim=-1) + + # zero pad + if embedding_dim % 2 == 1: + emb = torch.nn.functional.pad(emb, (0, 1, 0, 0)) + return emb + + +def get_2d_sincos_pos_embed( + embed_dim, + grid_size, + interpolation_scale=1.0, + base_size=16, +): + if isinstance(grid_size, int): + grid_size = (grid_size, grid_size) + + grid_h = ( + torch.arange(grid_size[0], dtype=torch.float32) + / (grid_size[0] / base_size) + / interpolation_scale + ) + grid_w = ( + torch.arange(grid_size[1], dtype=torch.float32) + / (grid_size[1] / base_size) + / interpolation_scale + ) + grid = torch.meshgrid(grid_w, grid_h, indexing="xy") # here w goes first + grid = torch.stack(grid, dim=0) + + grid = grid.reshape([2, 1, grid_size[1], grid_size[0]]) + pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid) + return pos_embed + + +def get_2d_sincos_pos_embed_from_grid(embed_dim, grid): + r""" + This function generates 2D sinusoidal positional embeddings from a grid. + + Args: + embed_dim (`int`): The embedding dimension. + grid (`torch.Tensor`): Grid of positions with shape `(H * W,)`. + + Returns: + `torch.Tensor`: The 2D sinusoidal positional embeddings with shape `(H * W, embed_dim)` + """ + if embed_dim % 2 != 0: + raise ValueError("embed_dim must be divisible by 2") + + # use half of dimensions to encode grid_h + emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0]) # (H*W, D/2) + emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1]) # (H*W, D/2) + + emb = torch.concat([emb_h, emb_w], dim=1) # (H*W, D) + return emb + + +def get_1d_sincos_pos_embed_from_grid(embed_dim, pos): + """ + This function generates 1D positional embeddings from a grid. + + Args: + embed_dim (`int`): The embedding dimension `D` + pos (`torch.Tensor`): 1D tensor of positions with shape `(M,)` + + Returns: + `torch.Tensor`: Sinusoidal positional embeddings of shape `(M, D)`. + """ + if embed_dim % 2 != 0: + raise ValueError("embed_dim must be divisible by 2") + + omega = torch.arange(embed_dim // 2, device=pos.device, dtype=torch.float64) + omega /= embed_dim / 2.0 + omega = 1.0 / 10000**omega # (D/2,) + + pos = pos.reshape(-1) # (M,) + out = torch.outer(pos, omega) # (M, D/2), outer product + + emb_sin = torch.sin(out) # (M, D/2) + emb_cos = torch.cos(out) # (M, D/2) + + emb = torch.concat([emb_sin, emb_cos], dim=1) # (M, D) + return emb + + +class Timesteps(nn.Module): + def __init__(self, num_channels: int, flip_sin_to_cos: bool, downscale_freq_shift: float): + super().__init__() + self.num_channels = num_channels + self.flip_sin_to_cos = flip_sin_to_cos + self.downscale_freq_shift = downscale_freq_shift + + def forward(self, timesteps): + t_emb = get_timestep_embedding( + timesteps, + self.num_channels, + flip_sin_to_cos=self.flip_sin_to_cos, + downscale_freq_shift=self.downscale_freq_shift, + ) + return t_emb + + +class TimestepEmbedding(nn.Module): + def __init__( + self, + in_channels: int, + time_embed_dim: int, + act_fn: str = "silu", + out_dim: int = None, + post_act_fn: Optional[str] = None, + cond_proj_dim=None, + sample_proj_bias=True, + ): + super().__init__() + + self.linear_1 = nn.Linear(in_channels, time_embed_dim, sample_proj_bias) + + if cond_proj_dim is not None: + self.cond_proj = nn.Linear(cond_proj_dim, in_channels, bias=False) + else: + self.cond_proj = None + + self.act = get_activation(act_fn) + + if out_dim is not None: + time_embed_dim_out = out_dim + else: + time_embed_dim_out = time_embed_dim + self.linear_2 = nn.Linear(time_embed_dim, time_embed_dim_out, sample_proj_bias) + + if post_act_fn is None: + self.post_act = None + else: + self.post_act = get_activation(post_act_fn) + + def forward(self, sample, condition=None): + if condition is not None: + sample = sample + self.cond_proj(condition) + sample = self.linear_1(sample) + + if self.act is not None: + sample = self.act(sample) + + sample = self.linear_2(sample) + + if self.post_act is not None: + sample = self.post_act(sample) + return sample + + +class PixArtAlphaTextProjection(nn.Module): + """ + Projects caption embeddings. Also handles dropout for classifier-free guidance. + """ + + def __init__(self, in_features, hidden_size, out_features=None, act_fn="gelu_tanh"): + super().__init__() + if out_features is None: + out_features = hidden_size + self.linear_1 = nn.Linear(in_features=in_features, out_features=hidden_size, bias=True) + if act_fn == "gelu_tanh": + self.act_1 = nn.GELU(approximate="tanh") + elif act_fn == "silu": + self.act_1 = nn.SiLU() + else: + raise ValueError(f"Unknown activation function: {act_fn}") + self.linear_2 = nn.Linear(in_features=hidden_size, out_features=out_features, bias=True) + + def forward(self, caption): + hidden_states = self.linear_1(caption) + hidden_states = self.act_1(hidden_states) + hidden_states = self.linear_2(hidden_states) + return hidden_states + + +class CogView3CombinedTimestepSizeEmbeddings(nn.Module): + def __init__(self, embedding_dim: int, condition_dim: int, pooled_projection_dim: int, timesteps_dim: int = 256): + super().__init__() + + self.time_proj = Timesteps(num_channels=timesteps_dim, flip_sin_to_cos=True, downscale_freq_shift=0) + self.condition_proj = Timesteps(num_channels=condition_dim, flip_sin_to_cos=True, downscale_freq_shift=0) + self.timestep_embedder = TimestepEmbedding(in_channels=timesteps_dim, time_embed_dim=embedding_dim) + self.condition_embedder = PixArtAlphaTextProjection(pooled_projection_dim, embedding_dim, act_fn="silu") + + def forward( + self, + timestep: torch.Tensor, + original_size: torch.Tensor, + target_size: torch.Tensor, + crop_coords: torch.Tensor, + hidden_dtype: torch.dtype, + ) -> torch.Tensor: + timesteps_proj = self.time_proj(timestep) + + original_size_proj = self.condition_proj(original_size.flatten()).view(original_size.size(0), -1) + crop_coords_proj = self.condition_proj(crop_coords.flatten()).view(crop_coords.size(0), -1) + target_size_proj = self.condition_proj(target_size.flatten()).view(target_size.size(0), -1) + + condition_proj = torch.cat([original_size_proj, crop_coords_proj, target_size_proj], dim=1) + + timesteps_emb = self.timestep_embedder(timesteps_proj.to(dtype=hidden_dtype)) # (B, embedding_dim) + condition_emb = self.condition_embedder(condition_proj.to(dtype=hidden_dtype)) # (B, embedding_dim) + + conditioning = timesteps_emb + condition_emb + return conditioning + + +class CogView3PlusPatchEmbed(nn.Module): + def __init__( + self, + in_channels: int = 16, + hidden_size: int = 2560, + patch_size: int = 2, + text_hidden_size: int = 4096, + pos_embed_max_size: int = 128, + ): + super().__init__() + self.in_channels = in_channels + self.hidden_size = hidden_size + self.patch_size = patch_size + self.text_hidden_size = text_hidden_size + self.pos_embed_max_size = pos_embed_max_size + # Linear projection for image patches + self.proj = nn.Linear(in_channels * patch_size**2, hidden_size) + + # Linear projection for text embeddings + self.text_proj = nn.Linear(text_hidden_size, hidden_size) + + pos_embed = get_2d_sincos_pos_embed( + hidden_size, pos_embed_max_size, base_size=pos_embed_max_size + ) + pos_embed = pos_embed.reshape(pos_embed_max_size, pos_embed_max_size, hidden_size) + self.register_buffer("pos_embed", pos_embed.float(), persistent=False) + + def forward(self, hidden_states: torch.Tensor, encoder_hidden_states: torch.Tensor) -> torch.Tensor: + batch_size, channel, height, width = hidden_states.shape + + if height % self.patch_size != 0 or width % self.patch_size != 0: + raise ValueError("Height and width must be divisible by patch size") + + height = height // self.patch_size + width = width // self.patch_size + hidden_states = hidden_states.view(batch_size, channel, height, self.patch_size, width, self.patch_size) + hidden_states = hidden_states.permute(0, 2, 4, 1, 3, 5).contiguous() + hidden_states = hidden_states.view(batch_size, height * width, channel * self.patch_size * self.patch_size) + + # Project the patches + hidden_states = self.proj(hidden_states) + encoder_hidden_states = self.text_proj(encoder_hidden_states) + hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1) + + # Calculate text_length + text_length = encoder_hidden_states.shape[1] + + image_pos_embed = self.pos_embed[:height, :width].reshape(height * width, -1) + text_pos_embed = torch.zeros( + (text_length, self.hidden_size), dtype=image_pos_embed.dtype, device=image_pos_embed.device + ) + pos_embed = torch.cat([text_pos_embed, image_pos_embed], dim=0)[None, ...] + return (hidden_states + pos_embed).to(hidden_states.dtype) \ No newline at end of file diff --git a/MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/layers/linear.py b/MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/layers/linear.py similarity index 97% rename from MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/layers/linear.py rename to MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/layers/linear.py index 57fe8d55dc4d030b1def6f8e7a97e9049a69fdc9..d242d17c2e83b7d27d86f1132a736951963b71bf 100644 --- a/MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/layers/linear.py +++ b/MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/layers/linear.py @@ -1,48 +1,48 @@ -#!/usr/bin/env python -# coding=utf-8 -# Copyright 2024 Huawei Technologies Co., Ltd -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# https://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - - -import torch -import torch.nn as nn - - -class QKVLinear(nn.Module): - def __init__(self, attention_dim, hidden_size, qkv_bias=True, device=None, dtype=None): - super(QKVLinear, self).__init__() - self.attention_dim = attention_dim - self.hidden_size = hidden_size - self.qkv_bias = qkv_bias - - factory_kwargs = {"device": device, "dtype": dtype} - - self.weight = nn.Parameter(torch.empty([self.attention_dim, 3 * self.hidden_size], **factory_kwargs)) - if self.qkv_bias: - self.bias = nn.Parameter(torch.empty([3 * self.hidden_size], **factory_kwargs)) - - def forward(self, hidden_states): - - if not self.qkv_bias: - qkv = torch.matmul(hidden_states, self.weight) - else: - qkv = torch.addmm( - self.bias, - hidden_states.view(hidden_states.size(0) * hidden_states.size(1), hidden_states.size(2)), - self.weight, - beta=1, - alpha=1 - ) - +#!/usr/bin/env python +# coding=utf-8 +# Copyright 2024 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# https://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +import torch +import torch.nn as nn + + +class QKVLinear(nn.Module): + def __init__(self, attention_dim, hidden_size, qkv_bias=True, device=None, dtype=None): + super(QKVLinear, self).__init__() + self.attention_dim = attention_dim + self.hidden_size = hidden_size + self.qkv_bias = qkv_bias + + factory_kwargs = {"device": device, "dtype": dtype} + + self.weight = nn.Parameter(torch.empty([self.attention_dim, 3 * self.hidden_size], **factory_kwargs)) + if self.qkv_bias: + self.bias = nn.Parameter(torch.empty([3 * self.hidden_size], **factory_kwargs)) + + def forward(self, hidden_states): + + if not self.qkv_bias: + qkv = torch.matmul(hidden_states, self.weight) + else: + qkv = torch.addmm( + self.bias, + hidden_states.view(hidden_states.size(0) * hidden_states.size(1), hidden_states.size(2)), + self.weight, + beta=1, + alpha=1 + ) + return qkv \ No newline at end of file diff --git a/MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/layers/normalization.py b/MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/layers/normalization.py similarity index 97% rename from MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/layers/normalization.py rename to MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/layers/normalization.py index 1ec0a5b15c93a8f3635abb42b4588e42c0daef9b..c12b70c9b12f59b81b1a6b5f3080e4008b96de48 100644 --- a/MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/layers/normalization.py +++ b/MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/layers/normalization.py @@ -1,177 +1,177 @@ -# coding=utf-8 -# Copyright 2024 HuggingFace Inc. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# https://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -import numbers -from typing import Optional, Tuple -from dataclasses import dataclass - -import torch -import torch.nn as nn -import torch.nn.functional as F - - -class RMSNorm(nn.Module): - def __init__(self, dim, eps: float, elementwise_affine: bool = True, bias: bool = False): - super().__init__() - - self.eps = eps - self.elementwise_affine = elementwise_affine - - if isinstance(dim, numbers.Integral): - dim = (dim,) - - self.dim = torch.Size(dim) - - self.weight = None - self.bias = None - - if elementwise_affine: - self.weight = nn.Parameter(torch.ones(dim)) - if bias: - self.bias = nn.Parameter(torch.zeros(dim)) - - def forward(self, hidden_states): - input_dtype = hidden_states.dtype - variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True) - hidden_states = hidden_states * torch.rsqrt(variance + self.eps) - - if self.weight is not None: - # convert into half-precision if necessary - if self.weight.dtype in [torch.float16, torch.bfloat16]: - hidden_states = hidden_states.to(self.weight.dtype) - hidden_states = hidden_states * self.weight - if self.bias is not None: - hidden_states = hidden_states + self.bias - else: - hidden_states = hidden_states.to(input_dtype) - - return hidden_states - - -@dataclass -class ChunkParam: - gate_msa: torch.Tensor - shift_mlp: torch.Tensor - scale_mlp: torch.Tensor - gate_mlp: torch.Tensor - context: torch.Tensor - c_gate_msa: torch.Tensor - c_shift_mlp: torch.Tensor - c_scale_mlp: torch.Tensor - c_gate_mlp: torch.Tensor - - -class CogView3PlusAdaLayerNormZeroTextImage(nn.Module): - r""" - Norm layer adaptive layer norm zero (adaLN-Zero). - - Parameters: - embedding_dim (`int`): The size of each embedding vector. - num_embeddings (`int`): The size of the embeddings dictionary. - """ - - def __init__(self, embedding_dim: int, dim: int): - super().__init__() - - self.silu = nn.SiLU() - self.linear = nn.Linear(embedding_dim, 12 * dim, bias=True) - self.norm_x = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-5) - self.norm_c = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-5) - - def forward( - self, - x: torch.Tensor, - context: torch.Tensor, - emb: Optional[torch.Tensor] = None, - ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]: - emb = self.linear(self.silu(emb)) - ( - shift_msa, - scale_msa, - gate_msa, - shift_mlp, - scale_mlp, - gate_mlp, - c_shift_msa, - c_scale_msa, - c_gate_msa, - c_shift_mlp, - c_scale_mlp, - c_gate_mlp, - ) = emb.chunk(12, dim=1) - normed_x = self.norm_x(x) - normed_context = self.norm_c(context) - x = normed_x * (1 + scale_msa[:, None]) + shift_msa[:, None] - context = normed_context * (1 + c_scale_msa[:, None]) + c_shift_msa[:, None] - return x, ChunkParam( - gate_msa, shift_mlp, scale_mlp, gate_mlp, context, c_gate_msa, c_shift_mlp, c_scale_mlp, c_gate_mlp - ) - - -class FP32LayerNorm(nn.LayerNorm): - def forward(self, inputs: torch.Tensor) -> torch.Tensor: - origin_dtype = inputs.dtype - return F.layer_norm( - inputs.float(), - self.normalized_shape, - self.weight.float() if self.weight is not None else None, - self.bias.float() if self.bias is not None else None, - self.eps, - ).to(origin_dtype) - - -class LpNorm(nn.Module): - def __init__(self, p: int = 2, dim: int = -1, eps: float = 1e-12): - super().__init__() - - self.p = p - self.dim = dim - self.eps = eps - - def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: - return F.normalize(hidden_states, p=self.p, dim=self.dim, eps=self.eps) - - -class AdaLayerNormContinuous(nn.Module): - def __init__( - self, - embedding_dim: int, - conditioning_embedding_dim: int, - # NOTE: It is a bit weird that the norm layer can be configured to have scale and shift parameters - # because the output is immediately scaled and shifted by the projected conditioning embeddings. - # Note that AdaLayerNorm does not let the norm layer have scale and shift parameters. - # However, this is how it was implemented in the original code, and it's rather likely you should - # set `elementwise_affine` to False. - elementwise_affine=True, - eps=1e-5, - bias=True, - norm_type="layer_norm", - ): - super().__init__() - self.silu = nn.SiLU() - self.linear = nn.Linear(conditioning_embedding_dim, embedding_dim * 2, bias=bias) - if norm_type == "layer_norm": - self.norm = nn.LayerNorm(embedding_dim, eps, elementwise_affine, bias) - elif norm_type == "rms_norm": - self.norm = RMSNorm(embedding_dim, eps, elementwise_affine) - else: - raise ValueError(f"unknown norm_type {norm_type}") - - def forward(self, x: torch.Tensor, conditioning_embedding: torch.Tensor) -> torch.Tensor: - # convert back to the original dtype in case `conditioning_embedding`` is upcasted to float32 (needed for hunyuanDiT) - emb = self.linear(self.silu(conditioning_embedding).to(x.dtype)) - scale, shift = torch.chunk(emb, 2, dim=1) - x = self.norm(x) * (1 + scale)[:, None, :] + shift[:, None, :] +# coding=utf-8 +# Copyright 2024 HuggingFace Inc. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# https://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import numbers +from typing import Optional, Tuple +from dataclasses import dataclass + +import torch +import torch.nn as nn +import torch.nn.functional as F + + +class RMSNorm(nn.Module): + def __init__(self, dim, eps: float, elementwise_affine: bool = True, bias: bool = False): + super().__init__() + + self.eps = eps + self.elementwise_affine = elementwise_affine + + if isinstance(dim, numbers.Integral): + dim = (dim,) + + self.dim = torch.Size(dim) + + self.weight = None + self.bias = None + + if elementwise_affine: + self.weight = nn.Parameter(torch.ones(dim)) + if bias: + self.bias = nn.Parameter(torch.zeros(dim)) + + def forward(self, hidden_states): + input_dtype = hidden_states.dtype + variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True) + hidden_states = hidden_states * torch.rsqrt(variance + self.eps) + + if self.weight is not None: + # convert into half-precision if necessary + if self.weight.dtype in [torch.float16, torch.bfloat16]: + hidden_states = hidden_states.to(self.weight.dtype) + hidden_states = hidden_states * self.weight + if self.bias is not None: + hidden_states = hidden_states + self.bias + else: + hidden_states = hidden_states.to(input_dtype) + + return hidden_states + + +@dataclass +class ChunkParam: + gate_msa: torch.Tensor + shift_mlp: torch.Tensor + scale_mlp: torch.Tensor + gate_mlp: torch.Tensor + context: torch.Tensor + c_gate_msa: torch.Tensor + c_shift_mlp: torch.Tensor + c_scale_mlp: torch.Tensor + c_gate_mlp: torch.Tensor + + +class CogView3PlusAdaLayerNormZeroTextImage(nn.Module): + r""" + Norm layer adaptive layer norm zero (adaLN-Zero). + + Parameters: + embedding_dim (`int`): The size of each embedding vector. + num_embeddings (`int`): The size of the embeddings dictionary. + """ + + def __init__(self, embedding_dim: int, dim: int): + super().__init__() + + self.silu = nn.SiLU() + self.linear = nn.Linear(embedding_dim, 12 * dim, bias=True) + self.norm_x = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-5) + self.norm_c = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-5) + + def forward( + self, + x: torch.Tensor, + context: torch.Tensor, + emb: Optional[torch.Tensor] = None, + ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]: + emb = self.linear(self.silu(emb)) + ( + shift_msa, + scale_msa, + gate_msa, + shift_mlp, + scale_mlp, + gate_mlp, + c_shift_msa, + c_scale_msa, + c_gate_msa, + c_shift_mlp, + c_scale_mlp, + c_gate_mlp, + ) = emb.chunk(12, dim=1) + normed_x = self.norm_x(x) + normed_context = self.norm_c(context) + x = normed_x * (1 + scale_msa[:, None]) + shift_msa[:, None] + context = normed_context * (1 + c_scale_msa[:, None]) + c_shift_msa[:, None] + return x, ChunkParam( + gate_msa, shift_mlp, scale_mlp, gate_mlp, context, c_gate_msa, c_shift_mlp, c_scale_mlp, c_gate_mlp + ) + + +class FP32LayerNorm(nn.LayerNorm): + def forward(self, inputs: torch.Tensor) -> torch.Tensor: + origin_dtype = inputs.dtype + return F.layer_norm( + inputs.float(), + self.normalized_shape, + self.weight.float() if self.weight is not None else None, + self.bias.float() if self.bias is not None else None, + self.eps, + ).to(origin_dtype) + + +class LpNorm(nn.Module): + def __init__(self, p: int = 2, dim: int = -1, eps: float = 1e-12): + super().__init__() + + self.p = p + self.dim = dim + self.eps = eps + + def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: + return F.normalize(hidden_states, p=self.p, dim=self.dim, eps=self.eps) + + +class AdaLayerNormContinuous(nn.Module): + def __init__( + self, + embedding_dim: int, + conditioning_embedding_dim: int, + # NOTE: It is a bit weird that the norm layer can be configured to have scale and shift parameters + # because the output is immediately scaled and shifted by the projected conditioning embeddings. + # Note that AdaLayerNorm does not let the norm layer have scale and shift parameters. + # However, this is how it was implemented in the original code, and it's rather likely you should + # set `elementwise_affine` to False. + elementwise_affine=True, + eps=1e-5, + bias=True, + norm_type="layer_norm", + ): + super().__init__() + self.silu = nn.SiLU() + self.linear = nn.Linear(conditioning_embedding_dim, embedding_dim * 2, bias=bias) + if norm_type == "layer_norm": + self.norm = nn.LayerNorm(embedding_dim, eps, elementwise_affine, bias) + elif norm_type == "rms_norm": + self.norm = RMSNorm(embedding_dim, eps, elementwise_affine) + else: + raise ValueError(f"unknown norm_type {norm_type}") + + def forward(self, x: torch.Tensor, conditioning_embedding: torch.Tensor) -> torch.Tensor: + # convert back to the original dtype in case `conditioning_embedding`` is upcasted to float32 (needed for hunyuanDiT) + emb = self.linear(self.silu(conditioning_embedding).to(x.dtype)) + scale, shift = torch.chunk(emb, 2, dim=1) + x = self.norm(x) * (1 + scale)[:, None, :] + shift[:, None, :] return x \ No newline at end of file diff --git a/MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/models/__init__.py b/MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/models/__init__.py similarity index 99% rename from MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/models/__init__.py rename to MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/models/__init__.py index b3c595bfccdd80de85b0883dc5246eb86c19d436..ae8f24f59afbe24b7ec2640ed3e338b042475482 100644 --- a/MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/models/__init__.py +++ b/MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/models/__init__.py @@ -1,2 +1,2 @@ -from .transformer_cogview3plus import CogView3PlusTransformer2DModel +from .transformer_cogview3plus import CogView3PlusTransformer2DModel from .modeling_utils import ModelMixin \ No newline at end of file diff --git a/MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/models/activations.py b/MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/models/activations.py similarity index 97% rename from MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/models/activations.py rename to MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/models/activations.py index 5bb3783ae4bb753995bcf9ddf5e23f3ee05f6ef1..b7d7cec29dfdef1b71c327e2baf12b464c1e9cf6 100644 --- a/MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/models/activations.py +++ b/MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/models/activations.py @@ -1,163 +1,163 @@ -# coding=utf-8 -# Copyright 2024 HuggingFace Inc. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# https://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -import torch -import torch.nn.functional as F -from torch import nn - -from diffusers.utils import deprecate -from diffusers.utils.import_utils import is_torch_npu_available - -if is_torch_npu_available(): - import torch_npu - -ACTIVATION_FUNCTIONS = { - "swish": nn.SiLU(), - "silu": nn.SiLU(), - "mish": nn.Mish(), - "gelu": nn.GELU(), - "relu": nn.ReLU(), -} - - -def get_activation(act_fn: str) -> nn.Module: - """Helper function to get activation function from string. - - Args: - act_fn (str): Name of activation function. - - Returns: - nn.Module: Activation function. - """ - - act_fn = act_fn.lower() - if act_fn in ACTIVATION_FUNCTIONS: - return ACTIVATION_FUNCTIONS[act_fn] - else: - raise ValueError(f"Unsupported activation function: {act_fn}") - - -class FP32SiLU(nn.Module): - r""" - SiLU activation function with input upcasted to torch.float32. - """ - - def __init__(self): - super().__init__() - - def forward(self, inputs: torch.Tensor) -> torch.Tensor: - return F.silu(inputs.float(), inplace=False).to(inputs.dtype) - - -class GELU(nn.Module): - r""" - GELU activation function with tanh approximation support with `approximate="tanh"`. - - Parameters: - dim_in (`int`): The number of channels in the input. - dim_out (`int`): The number of channels in the output. - approximate (`str`, *optional*, defaults to `"none"`): If `"tanh"`, use tanh approximation. - bias (`bool`, defaults to True): Whether to use a bias in the linear layer. - """ - - def __init__(self, dim_in: int, dim_out: int, approximate: str = "none", bias: bool = True): - super().__init__() - self.proj = nn.Linear(dim_in, dim_out, bias=bias) - self.approximate = approximate - - def gelu(self, gate: torch.Tensor) -> torch.Tensor: - return F.gelu(gate, approximate=self.approximate) - - def forward(self, hidden_states): - hidden_states = self.proj(hidden_states) - hidden_states = self.gelu(hidden_states) - return hidden_states - - -class GEGLU(nn.Module): - r""" - Parameters: - dim_in (`int`): The number of channels in the input. - dim_out (`int`): The number of channels in the output. - bias (`bool`, defaults to True): Whether to use a bias in the linear layer. - """ - - def __init__(self, dim_in: int, dim_out: int, bias: bool = True): - super().__init__() - self.proj = nn.Linear(dim_in, dim_out * 2, bias=bias) - - def gelu(self, gate: torch.Tensor) -> torch.Tensor: - return F.gelu(gate) - - def forward(self, hidden_states, *args, **kwargs): - if len(args) > 0 or kwargs.get("scale", None) is not None: - deprecation_message = "The `scale` argument is deprecated and will be ignored. Please remove it, as passing it will raise an error in the future. `scale` should directly be passed while calling the underlying pipeline component i.e., via `cross_attention_kwargs`." - deprecate("scale", "1.0.0", deprecation_message) - hidden_states = self.proj(hidden_states) - if is_torch_npu_available(): - # using torch_npu.npu_geglu can run faster and save memory on NPU. - return torch_npu.npu_geglu(hidden_states, dim=-1, approximate=1)[0] - else: - hidden_states, gate = hidden_states.chunk(2, dim=-1) - return hidden_states * self.gelu(gate) - - -class SwiGLU(nn.Module): - r""" - Parameters: - dim_in (`int`): The number of channels in the input. - dim_out (`int`): The number of channels in the output. - bias (`bool`, defaults to True): Whether to use a bias in the linear layer. - """ - - def __init__(self, dim_in: int, dim_out: int, bias: bool = True): - super().__init__() - - self.proj = nn.Linear(dim_in, dim_out * 2, bias=bias) - self.activation = nn.SiLU() - - def forward(self, hidden_states): - hidden_states = self.proj(hidden_states) - hidden_states, gate = hidden_states.chunk(2, dim=-1) - return hidden_states * self.activation(gate) - - -class ApproximateGELU(nn.Module): - r""" - Parameters: - dim_in (`int`): The number of channels in the input. - dim_out (`int`): The number of channels in the output. - bias (`bool`, defaults to True): Whether to use a bias in the linear layer. - """ - - def __init__(self, dim_in: int, dim_out: int, bias: bool = True): - super().__init__() - self.proj = nn.Linear(dim_in, dim_out, bias=bias) - - def forward(self, x: torch.Tensor) -> torch.Tensor: - x = self.proj(x) - return x * torch.sigmoid(1.702 * x) - - -class LinearActivation(nn.Module): - def __init__(self, dim_in: int, dim_out: int, bias: bool = True, activation: str = "silu"): - super().__init__() - - self.proj = nn.Linear(dim_in, dim_out, bias=bias) - self.activation = get_activation(activation) - - def forward(self, hidden_states): - hidden_states = self.proj(hidden_states) +# coding=utf-8 +# Copyright 2024 HuggingFace Inc. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# https://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import torch +import torch.nn.functional as F +from torch import nn + +from diffusers.utils import deprecate +from diffusers.utils.import_utils import is_torch_npu_available + +if is_torch_npu_available(): + import torch_npu + +ACTIVATION_FUNCTIONS = { + "swish": nn.SiLU(), + "silu": nn.SiLU(), + "mish": nn.Mish(), + "gelu": nn.GELU(), + "relu": nn.ReLU(), +} + + +def get_activation(act_fn: str) -> nn.Module: + """Helper function to get activation function from string. + + Args: + act_fn (str): Name of activation function. + + Returns: + nn.Module: Activation function. + """ + + act_fn = act_fn.lower() + if act_fn in ACTIVATION_FUNCTIONS: + return ACTIVATION_FUNCTIONS[act_fn] + else: + raise ValueError(f"Unsupported activation function: {act_fn}") + + +class FP32SiLU(nn.Module): + r""" + SiLU activation function with input upcasted to torch.float32. + """ + + def __init__(self): + super().__init__() + + def forward(self, inputs: torch.Tensor) -> torch.Tensor: + return F.silu(inputs.float(), inplace=False).to(inputs.dtype) + + +class GELU(nn.Module): + r""" + GELU activation function with tanh approximation support with `approximate="tanh"`. + + Parameters: + dim_in (`int`): The number of channels in the input. + dim_out (`int`): The number of channels in the output. + approximate (`str`, *optional*, defaults to `"none"`): If `"tanh"`, use tanh approximation. + bias (`bool`, defaults to True): Whether to use a bias in the linear layer. + """ + + def __init__(self, dim_in: int, dim_out: int, approximate: str = "none", bias: bool = True): + super().__init__() + self.proj = nn.Linear(dim_in, dim_out, bias=bias) + self.approximate = approximate + + def gelu(self, gate: torch.Tensor) -> torch.Tensor: + return F.gelu(gate, approximate=self.approximate) + + def forward(self, hidden_states): + hidden_states = self.proj(hidden_states) + hidden_states = self.gelu(hidden_states) + return hidden_states + + +class GEGLU(nn.Module): + r""" + Parameters: + dim_in (`int`): The number of channels in the input. + dim_out (`int`): The number of channels in the output. + bias (`bool`, defaults to True): Whether to use a bias in the linear layer. + """ + + def __init__(self, dim_in: int, dim_out: int, bias: bool = True): + super().__init__() + self.proj = nn.Linear(dim_in, dim_out * 2, bias=bias) + + def gelu(self, gate: torch.Tensor) -> torch.Tensor: + return F.gelu(gate) + + def forward(self, hidden_states, *args, **kwargs): + if len(args) > 0 or kwargs.get("scale", None) is not None: + deprecation_message = "The `scale` argument is deprecated and will be ignored. Please remove it, as passing it will raise an error in the future. `scale` should directly be passed while calling the underlying pipeline component i.e., via `cross_attention_kwargs`." + deprecate("scale", "1.0.0", deprecation_message) + hidden_states = self.proj(hidden_states) + if is_torch_npu_available(): + # using torch_npu.npu_geglu can run faster and save memory on NPU. + return torch_npu.npu_geglu(hidden_states, dim=-1, approximate=1)[0] + else: + hidden_states, gate = hidden_states.chunk(2, dim=-1) + return hidden_states * self.gelu(gate) + + +class SwiGLU(nn.Module): + r""" + Parameters: + dim_in (`int`): The number of channels in the input. + dim_out (`int`): The number of channels in the output. + bias (`bool`, defaults to True): Whether to use a bias in the linear layer. + """ + + def __init__(self, dim_in: int, dim_out: int, bias: bool = True): + super().__init__() + + self.proj = nn.Linear(dim_in, dim_out * 2, bias=bias) + self.activation = nn.SiLU() + + def forward(self, hidden_states): + hidden_states = self.proj(hidden_states) + hidden_states, gate = hidden_states.chunk(2, dim=-1) + return hidden_states * self.activation(gate) + + +class ApproximateGELU(nn.Module): + r""" + Parameters: + dim_in (`int`): The number of channels in the input. + dim_out (`int`): The number of channels in the output. + bias (`bool`, defaults to True): Whether to use a bias in the linear layer. + """ + + def __init__(self, dim_in: int, dim_out: int, bias: bool = True): + super().__init__() + self.proj = nn.Linear(dim_in, dim_out, bias=bias) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + x = self.proj(x) + return x * torch.sigmoid(1.702 * x) + + +class LinearActivation(nn.Module): + def __init__(self, dim_in: int, dim_out: int, bias: bool = True, activation: str = "silu"): + super().__init__() + + self.proj = nn.Linear(dim_in, dim_out, bias=bias) + self.activation = get_activation(activation) + + def forward(self, hidden_states): + hidden_states = self.proj(hidden_states) return self.activation(hidden_states) \ No newline at end of file diff --git a/MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/models/attention.py b/MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/models/attention.py similarity index 97% rename from MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/models/attention.py rename to MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/models/attention.py index a7a559ff2f529de83c4620249cd37c3afe59fefb..946d829c6c733d03e3d4970a7899368190b93a1f 100644 --- a/MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/models/attention.py +++ b/MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/models/attention.py @@ -1,87 +1,87 @@ -# Copyright 2024 The HuggingFace Team. All rights reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# https://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -from typing import Optional - -import torch -from torch import nn - -from diffusers.utils import deprecate, logging -from .activations import GEGLU, GELU, ApproximateGELU, LinearActivation, SwiGLU - - -logger = logging.get_logger(__name__) - - -class FeedForward(nn.Module): - r""" - A feed-forward layer. - - Parameters: - dim (`int`): The number of channels in the input. - dim_out (`int`, *optional*): The number of channels in the output. If not given, defaults to `dim`. - mult (`int`, *optional*, defaults to 4): The multiplier to use for the hidden dimension. - dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. - activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward. - final_dropout (`bool` *optional*, defaults to False): Apply a final dropout. - bias (`bool`, defaults to True): Whether to use a bias in the linear layer. - """ - - def __init__( - self, - dim: int, - dim_out: Optional[int] = None, - mult: int = 4, - dropout: float = 0.0, - activation_fn: str = "geglu", - final_dropout: bool = False, - inner_dim=None, - bias: bool = True, - ): - super().__init__() - if inner_dim is None: - inner_dim = int(dim * mult) - dim_out = dim_out if dim_out is not None else dim - - if activation_fn == "gelu": - act_fn = GELU(dim, inner_dim, bias=bias) - if activation_fn == "gelu-approximate": - act_fn = GELU(dim, inner_dim, approximate="tanh", bias=bias) - elif activation_fn == "geglu": - act_fn = GEGLU(dim, inner_dim, bias=bias) - elif activation_fn == "geglu-approximate": - act_fn = ApproximateGELU(dim, inner_dim, bias=bias) - elif activation_fn == "swiglu": - act_fn = SwiGLU(dim, inner_dim, bias=bias) - elif activation_fn == "linear-silu": - act_fn = LinearActivation(dim, inner_dim, bias=bias, activation="silu") - - self.net = nn.ModuleList([]) - # project in - self.net.append(act_fn) - # project dropout - self.net.append(nn.Dropout(dropout)) - # project out - self.net.append(nn.Linear(inner_dim, dim_out, bias=bias)) - # FF as used in Vision Transformer, MLP-Mixer, etc. have a final dropout - if final_dropout: - self.net.append(nn.Dropout(dropout)) - - def forward(self, hidden_states: torch.Tensor, *args, **kwargs) -> torch.Tensor: - if len(args) > 0 or kwargs.get("scale", None) is not None: - deprecation_message = "The `scale` argument is deprecated and will be ignored. Please remove it, as passing it will raise an error in the future. `scale` should directly be passed while calling the underlying pipeline component i.e., via `cross_attention_kwargs`." - deprecate("scale", "1.0.0", deprecation_message) - for module in self.net: - hidden_states = module(hidden_states) +# Copyright 2024 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# https://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from typing import Optional + +import torch +from torch import nn + +from diffusers.utils import deprecate, logging +from .activations import GEGLU, GELU, ApproximateGELU, LinearActivation, SwiGLU + + +logger = logging.get_logger(__name__) + + +class FeedForward(nn.Module): + r""" + A feed-forward layer. + + Parameters: + dim (`int`): The number of channels in the input. + dim_out (`int`, *optional*): The number of channels in the output. If not given, defaults to `dim`. + mult (`int`, *optional*, defaults to 4): The multiplier to use for the hidden dimension. + dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. + activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward. + final_dropout (`bool` *optional*, defaults to False): Apply a final dropout. + bias (`bool`, defaults to True): Whether to use a bias in the linear layer. + """ + + def __init__( + self, + dim: int, + dim_out: Optional[int] = None, + mult: int = 4, + dropout: float = 0.0, + activation_fn: str = "geglu", + final_dropout: bool = False, + inner_dim=None, + bias: bool = True, + ): + super().__init__() + if inner_dim is None: + inner_dim = int(dim * mult) + dim_out = dim_out if dim_out is not None else dim + + if activation_fn == "gelu": + act_fn = GELU(dim, inner_dim, bias=bias) + if activation_fn == "gelu-approximate": + act_fn = GELU(dim, inner_dim, approximate="tanh", bias=bias) + elif activation_fn == "geglu": + act_fn = GEGLU(dim, inner_dim, bias=bias) + elif activation_fn == "geglu-approximate": + act_fn = ApproximateGELU(dim, inner_dim, bias=bias) + elif activation_fn == "swiglu": + act_fn = SwiGLU(dim, inner_dim, bias=bias) + elif activation_fn == "linear-silu": + act_fn = LinearActivation(dim, inner_dim, bias=bias, activation="silu") + + self.net = nn.ModuleList([]) + # project in + self.net.append(act_fn) + # project dropout + self.net.append(nn.Dropout(dropout)) + # project out + self.net.append(nn.Linear(inner_dim, dim_out, bias=bias)) + # FF as used in Vision Transformer, MLP-Mixer, etc. have a final dropout + if final_dropout: + self.net.append(nn.Dropout(dropout)) + + def forward(self, hidden_states: torch.Tensor, *args, **kwargs) -> torch.Tensor: + if len(args) > 0 or kwargs.get("scale", None) is not None: + deprecation_message = "The `scale` argument is deprecated and will be ignored. Please remove it, as passing it will raise an error in the future. `scale` should directly be passed while calling the underlying pipeline component i.e., via `cross_attention_kwargs`." + deprecate("scale", "1.0.0", deprecation_message) + for module in self.net: + hidden_states = module(hidden_states) return hidden_states \ No newline at end of file diff --git a/MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/models/attention_processor.py b/MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/models/attention_processor.py similarity index 97% rename from MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/models/attention_processor.py rename to MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/models/attention_processor.py index c197a989b7c837ec1f6c588dd03633268c62175f..d36e9265a394ef66be3d1098fb1abe9d686f1482 100644 --- a/MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/models/attention_processor.py +++ b/MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/models/attention_processor.py @@ -1,348 +1,348 @@ -# Copyright 2024 The HuggingFace Team. All rights reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# https://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -import inspect -from typing import Optional - -import torch -import torch.nn.functional as F -from torch import nn -import torch_npu - -from diffusers.utils import logging -from diffusers.utils.torch_utils import maybe_allow_in_graph - -from ..layers import QKVLinear - -logger = logging.get_logger(__name__) # pylint: disable=invalid-name - - -@maybe_allow_in_graph -class Attention(nn.Module): - def __init__( - self, - query_dim: int, - cross_attention_dim: Optional[int] = None, - heads: int = 8, - kv_heads: Optional[int] = None, - dim_head: int = 64, - dropout: float = 0.0, - bias: bool = False, - upcast_attention: bool = False, - upcast_softmax: bool = False, - cross_attention_norm: Optional[str] = None, - cross_attention_norm_num_groups: int = 32, - qk_norm: Optional[str] = None, - added_kv_proj_dim: Optional[int] = None, - added_proj_bias: Optional[bool] = True, - norm_num_groups: Optional[int] = None, - out_bias: bool = True, - scale_qk: bool = True, - only_cross_attention: bool = False, - eps: float = 1e-5, - rescale_output_factor: float = 1.0, - residual_connection: bool = False, - _from_deprecated_attn_block: bool = False, - processor: Optional["AttnProcessor"] = None, - out_dim: int = None, - out_context_dim: int = None, - context_pre_only=None, - pre_only=False, - elementwise_affine: bool = True, - is_causal: bool = False, - ): - super().__init__() - - # To prevent circular import. - from ..layers.normalization import FP32LayerNorm, LpNorm, RMSNorm - - self.inner_dim = out_dim if out_dim is not None else dim_head * heads - self.inner_kv_dim = self.inner_dim if kv_heads is None else dim_head * kv_heads - self.query_dim = query_dim - self.use_bias = bias - self.is_cross_attention = cross_attention_dim is not None - self.cross_attention_dim = cross_attention_dim if cross_attention_dim is not None else query_dim - self.upcast_attention = upcast_attention - self.upcast_softmax = upcast_softmax - self.rescale_output_factor = rescale_output_factor - self.residual_connection = residual_connection - self.dropout = dropout - self.fused_projections = False - self.out_dim = out_dim if out_dim is not None else query_dim - self.out_context_dim = out_context_dim if out_context_dim is not None else query_dim - self.context_pre_only = context_pre_only - self.pre_only = pre_only - self.is_causal = is_causal - - # we make use of this private variable to know whether this class is loaded - # with an deprecated state dict so that we can convert it on the fly - self._from_deprecated_attn_block = _from_deprecated_attn_block - - self.scale_qk = scale_qk - self.scale = dim_head**-0.5 if self.scale_qk else 1.0 - - self.heads = out_dim // dim_head if out_dim is not None else heads - # for slice_size > 0 the attention score computation - # is split across the batch axis to save memory - # You can set slice_size with `set_attention_slice` - self.sliceable_head_dim = heads - - self.added_kv_proj_dim = added_kv_proj_dim - self.only_cross_attention = only_cross_attention - - if self.added_kv_proj_dim is None and self.only_cross_attention: - raise ValueError( - "`only_cross_attention` can only be set to True if `added_kv_proj_dim` is not None. Make sure to set either `only_cross_attention=False` or define `added_kv_proj_dim`." - ) - - if norm_num_groups is not None: - self.group_norm = nn.GroupNorm(num_channels=query_dim, num_groups=norm_num_groups, eps=eps, affine=True) - else: - self.group_norm = None - - self.spatial_norm = None - - if qk_norm is None: - self.norm_q = None - self.norm_k = None - elif qk_norm == "layer_norm": - self.norm_q = nn.LayerNorm(dim_head, eps=eps, elementwise_affine=elementwise_affine) - self.norm_k = nn.LayerNorm(dim_head, eps=eps, elementwise_affine=elementwise_affine) - elif qk_norm == "fp32_layer_norm": - self.norm_q = FP32LayerNorm(dim_head, elementwise_affine=False, bias=False, eps=eps) - self.norm_k = FP32LayerNorm(dim_head, elementwise_affine=False, bias=False, eps=eps) - elif qk_norm == "layer_norm_across_heads": - # Lumina applies qk norm across all heads - self.norm_q = nn.LayerNorm(dim_head * heads, eps=eps) - self.norm_k = nn.LayerNorm(dim_head * kv_heads, eps=eps) - elif qk_norm == "rms_norm": - self.norm_q = RMSNorm(dim_head, eps=eps) - self.norm_k = RMSNorm(dim_head, eps=eps) - elif qk_norm == "rms_norm_across_heads": - # LTX applies qk norm across all heads - self.norm_q = RMSNorm(dim_head * heads, eps=eps) - self.norm_k = RMSNorm(dim_head * kv_heads, eps=eps) - elif qk_norm == "l2": - self.norm_q = LpNorm(p=2, dim=-1, eps=eps) - self.norm_k = LpNorm(p=2, dim=-1, eps=eps) - else: - raise ValueError(f"unknown qk_norm: {qk_norm}. Should be None,'layer_norm','fp32_layer_norm','rms_norm'") - - if cross_attention_norm is None: - self.norm_cross = None - elif cross_attention_norm == "layer_norm": - self.norm_cross = nn.LayerNorm(self.cross_attention_dim) - elif cross_attention_norm == "group_norm": - if self.added_kv_proj_dim is not None: - # The given `encoder_hidden_states` are initially of shape - # (batch_size, seq_len, added_kv_proj_dim) before being projected - # to (batch_size, seq_len, cross_attention_dim). The norm is applied - # before the projection, so we need to use `added_kv_proj_dim` as - # the number of channels for the group norm. - norm_cross_num_channels = added_kv_proj_dim - else: - norm_cross_num_channels = self.cross_attention_dim - - self.norm_cross = nn.GroupNorm( - num_channels=norm_cross_num_channels, num_groups=cross_attention_norm_num_groups, eps=1e-5, affine=True - ) - else: - raise ValueError( - f"unknown cross_attention_norm: {cross_attention_norm}. Should be None, 'layer_norm' or 'group_norm'" - ) - - self.to_qkv = QKVLinear(self.inner_dim, query_dim) - - self.added_proj_bias = added_proj_bias - if self.added_kv_proj_dim is not None: - self.add_k_proj = nn.Linear(added_kv_proj_dim, self.inner_kv_dim, bias=added_proj_bias) - self.add_v_proj = nn.Linear(added_kv_proj_dim, self.inner_kv_dim, bias=added_proj_bias) - if self.context_pre_only is not None: - self.add_q_proj = nn.Linear(added_kv_proj_dim, self.inner_dim, bias=added_proj_bias) - else: - self.add_q_proj = None - self.add_k_proj = None - self.add_v_proj = None - - if not self.pre_only: - self.to_out = nn.ModuleList([]) - self.to_out.append(nn.Linear(self.inner_dim, self.out_dim, bias=out_bias)) - self.to_out.append(nn.Dropout(dropout)) - else: - self.to_out = None - - if self.context_pre_only is not None and not self.context_pre_only: - self.to_add_out = nn.Linear(self.inner_dim, self.out_context_dim, bias=out_bias) - else: - self.to_add_out = None - - if qk_norm is not None and added_kv_proj_dim is not None: - if qk_norm == "fp32_layer_norm": - self.norm_added_q = FP32LayerNorm(dim_head, elementwise_affine=False, bias=False, eps=eps) - self.norm_added_k = FP32LayerNorm(dim_head, elementwise_affine=False, bias=False, eps=eps) - elif qk_norm == "rms_norm": - self.norm_added_q = RMSNorm(dim_head, eps=eps) - self.norm_added_k = RMSNorm(dim_head, eps=eps) - else: - raise ValueError( - f"unknown qk_norm: {qk_norm}. Should be one of `None,'layer_norm','fp32_layer_norm','rms_norm'`" - ) - else: - self.norm_added_q = None - self.norm_added_k = None - - self.set_processor(processor) - - def set_processor(self, processor: "AttnProcessor") -> None: - r""" - Set the attention processor to use. - - Args: - processor (`AttnProcessor`): - The attention processor to use. - """ - if ( - hasattr(self, "processor") - and isinstance(self.processor, torch.nn.Module) - and not isinstance(processor, torch.nn.Module) - ): - logger.info(f"You are removing possibly trained weights of {self.processor} with {processor}") - self._modules.pop("processor") - - self.processor = processor - - def forward( - self, - hidden_states: torch.Tensor, - encoder_hidden_states: Optional[torch.Tensor] = None, - attention_mask: Optional[torch.Tensor] = None, - **cross_attention_kwargs, - ) -> torch.Tensor: - attn_parameters = set(inspect.signature(self.processor.__call__).parameters.keys()) - quiet_attn_parameters = {"ip_adapter_masks", "ip_hidden_states"} - unused_kwargs = [ - k for k, _ in cross_attention_kwargs.items() if k not in attn_parameters and k not in quiet_attn_parameters - ] - if len(unused_kwargs) > 0: - logger.warning( - f"cross_attention_kwargs {unused_kwargs} are not expected by {self.processor.__class__.__name__} and will be ignored." - ) - cross_attention_kwargs = {k: w for k, w in cross_attention_kwargs.items() if k in attn_parameters} - - return self.processor( - self, - hidden_states, - encoder_hidden_states=encoder_hidden_states, - attention_mask=attention_mask, - **cross_attention_kwargs, - ) - - def prepare_attention_mask( - self, attention_mask: torch.Tensor, target_length: int, batch_size: int, out_dim: int = 3 - ) -> torch.Tensor: - head_size = self.heads - if attention_mask is None: - return attention_mask - - current_length: int = attention_mask.shape[-1] - if current_length != target_length: - if attention_mask.device.type == "mps": - padding_shape = (attention_mask.shape[0], attention_mask.shape[1], target_length) - padding = torch.zeros(padding_shape, dtype=attention_mask.dtype, device=attention_mask.device) - attention_mask = torch.cat([attention_mask, padding], dim=2) - else: - attention_mask = F.pad(attention_mask, (0, target_length), value=0.0) - - if out_dim == 3: - if attention_mask.shape[0] < batch_size * head_size: - attention_mask = attention_mask.repeat_interleave(head_size, dim=0) - elif out_dim == 4: - attention_mask = attention_mask.unsqueeze(1) - attention_mask = attention_mask.repeat_interleave(head_size, dim=1) - - return attention_mask - - -class CogVideoXAttnProcessor2_0: - r""" - Processor for implementing scaled dot-product attention for the CogVideoX model. It applies a rotary embedding on - query and key vectors, but does not include spatial normalization. - """ - - def __init__(self): - if not hasattr(F, "scaled_dot_product_attention"): - raise ImportError("CogVideoXAttnProcessor requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.") - - def __call__( - self, - attn: Attention, - hidden_states: torch.Tensor, - encoder_hidden_states: torch.Tensor, - attention_mask: Optional[torch.Tensor] = None, - ) -> torch.Tensor: - text_seq_length = encoder_hidden_states.size(1) - - hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1) - - batch_size, sequence_length, _ = ( - hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape - ) - - if attention_mask is not None: - attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) - attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1]) - - B, S, _ = hidden_states.shape - qkv = attn.to_qkv(hidden_states) - inner_dim = qkv.shape[-1] // 3 - head_dim = inner_dim // attn.heads - qkv_shape = (B, S, 3, attn.heads, head_dim) - query, key, value = qkv.view(qkv_shape).permute(2, 0, 3, 1, 4).contiguous().unbind(0) - - if attn.norm_q is not None: - query = attn.norm_q(query) - if attn.norm_k is not None: - key = attn.norm_k(key) - - B, N, S, D = query.shape - dim = 48 - pad_shape = [B, N, S, D] - pad_shape[-1] = dim - pad_shape[-1] - pad = torch.zeros(pad_shape, dtype=query.dtype, device=query.device) - query = torch.cat([query, pad], dim=-1) - key = torch.cat([key, pad], dim=-1) - value = torch.cat([value, pad], dim=-1) - hidden_states = torch_npu.npu_prompt_flash_attention( - query, - key, - value, - input_layout='BNSD', - scale_value=D**-0.5, - pre_tokens=65535, - next_tokens=65535, - num_heads=N - ) - hidden_states = hidden_states[:, :, :, :D] - - hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim) - - # linear proj - hidden_states = attn.to_out[0](hidden_states) - # dropout - hidden_states = attn.to_out[1](hidden_states) - - encoder_hidden_states, hidden_states = hidden_states.split( - [text_seq_length, hidden_states.size(1) - text_seq_length], dim=1 - ) +# Copyright 2024 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# https://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Optional + +import torch +import torch.nn.functional as F +from torch import nn +import torch_npu + +from diffusers.utils import logging +from diffusers.utils.torch_utils import maybe_allow_in_graph + +from ..layers import QKVLinear + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +@maybe_allow_in_graph +class Attention(nn.Module): + def __init__( + self, + query_dim: int, + cross_attention_dim: Optional[int] = None, + heads: int = 8, + kv_heads: Optional[int] = None, + dim_head: int = 64, + dropout: float = 0.0, + bias: bool = False, + upcast_attention: bool = False, + upcast_softmax: bool = False, + cross_attention_norm: Optional[str] = None, + cross_attention_norm_num_groups: int = 32, + qk_norm: Optional[str] = None, + added_kv_proj_dim: Optional[int] = None, + added_proj_bias: Optional[bool] = True, + norm_num_groups: Optional[int] = None, + out_bias: bool = True, + scale_qk: bool = True, + only_cross_attention: bool = False, + eps: float = 1e-5, + rescale_output_factor: float = 1.0, + residual_connection: bool = False, + _from_deprecated_attn_block: bool = False, + processor: Optional["AttnProcessor"] = None, + out_dim: int = None, + out_context_dim: int = None, + context_pre_only=None, + pre_only=False, + elementwise_affine: bool = True, + is_causal: bool = False, + ): + super().__init__() + + # To prevent circular import. + from ..layers.normalization import FP32LayerNorm, LpNorm, RMSNorm + + self.inner_dim = out_dim if out_dim is not None else dim_head * heads + self.inner_kv_dim = self.inner_dim if kv_heads is None else dim_head * kv_heads + self.query_dim = query_dim + self.use_bias = bias + self.is_cross_attention = cross_attention_dim is not None + self.cross_attention_dim = cross_attention_dim if cross_attention_dim is not None else query_dim + self.upcast_attention = upcast_attention + self.upcast_softmax = upcast_softmax + self.rescale_output_factor = rescale_output_factor + self.residual_connection = residual_connection + self.dropout = dropout + self.fused_projections = False + self.out_dim = out_dim if out_dim is not None else query_dim + self.out_context_dim = out_context_dim if out_context_dim is not None else query_dim + self.context_pre_only = context_pre_only + self.pre_only = pre_only + self.is_causal = is_causal + + # we make use of this private variable to know whether this class is loaded + # with an deprecated state dict so that we can convert it on the fly + self._from_deprecated_attn_block = _from_deprecated_attn_block + + self.scale_qk = scale_qk + self.scale = dim_head**-0.5 if self.scale_qk else 1.0 + + self.heads = out_dim // dim_head if out_dim is not None else heads + # for slice_size > 0 the attention score computation + # is split across the batch axis to save memory + # You can set slice_size with `set_attention_slice` + self.sliceable_head_dim = heads + + self.added_kv_proj_dim = added_kv_proj_dim + self.only_cross_attention = only_cross_attention + + if self.added_kv_proj_dim is None and self.only_cross_attention: + raise ValueError( + "`only_cross_attention` can only be set to True if `added_kv_proj_dim` is not None. Make sure to set either `only_cross_attention=False` or define `added_kv_proj_dim`." + ) + + if norm_num_groups is not None: + self.group_norm = nn.GroupNorm(num_channels=query_dim, num_groups=norm_num_groups, eps=eps, affine=True) + else: + self.group_norm = None + + self.spatial_norm = None + + if qk_norm is None: + self.norm_q = None + self.norm_k = None + elif qk_norm == "layer_norm": + self.norm_q = nn.LayerNorm(dim_head, eps=eps, elementwise_affine=elementwise_affine) + self.norm_k = nn.LayerNorm(dim_head, eps=eps, elementwise_affine=elementwise_affine) + elif qk_norm == "fp32_layer_norm": + self.norm_q = FP32LayerNorm(dim_head, elementwise_affine=False, bias=False, eps=eps) + self.norm_k = FP32LayerNorm(dim_head, elementwise_affine=False, bias=False, eps=eps) + elif qk_norm == "layer_norm_across_heads": + # Lumina applies qk norm across all heads + self.norm_q = nn.LayerNorm(dim_head * heads, eps=eps) + self.norm_k = nn.LayerNorm(dim_head * kv_heads, eps=eps) + elif qk_norm == "rms_norm": + self.norm_q = RMSNorm(dim_head, eps=eps) + self.norm_k = RMSNorm(dim_head, eps=eps) + elif qk_norm == "rms_norm_across_heads": + # LTX applies qk norm across all heads + self.norm_q = RMSNorm(dim_head * heads, eps=eps) + self.norm_k = RMSNorm(dim_head * kv_heads, eps=eps) + elif qk_norm == "l2": + self.norm_q = LpNorm(p=2, dim=-1, eps=eps) + self.norm_k = LpNorm(p=2, dim=-1, eps=eps) + else: + raise ValueError(f"unknown qk_norm: {qk_norm}. Should be None,'layer_norm','fp32_layer_norm','rms_norm'") + + if cross_attention_norm is None: + self.norm_cross = None + elif cross_attention_norm == "layer_norm": + self.norm_cross = nn.LayerNorm(self.cross_attention_dim) + elif cross_attention_norm == "group_norm": + if self.added_kv_proj_dim is not None: + # The given `encoder_hidden_states` are initially of shape + # (batch_size, seq_len, added_kv_proj_dim) before being projected + # to (batch_size, seq_len, cross_attention_dim). The norm is applied + # before the projection, so we need to use `added_kv_proj_dim` as + # the number of channels for the group norm. + norm_cross_num_channels = added_kv_proj_dim + else: + norm_cross_num_channels = self.cross_attention_dim + + self.norm_cross = nn.GroupNorm( + num_channels=norm_cross_num_channels, num_groups=cross_attention_norm_num_groups, eps=1e-5, affine=True + ) + else: + raise ValueError( + f"unknown cross_attention_norm: {cross_attention_norm}. Should be None, 'layer_norm' or 'group_norm'" + ) + + self.to_qkv = QKVLinear(self.inner_dim, query_dim) + + self.added_proj_bias = added_proj_bias + if self.added_kv_proj_dim is not None: + self.add_k_proj = nn.Linear(added_kv_proj_dim, self.inner_kv_dim, bias=added_proj_bias) + self.add_v_proj = nn.Linear(added_kv_proj_dim, self.inner_kv_dim, bias=added_proj_bias) + if self.context_pre_only is not None: + self.add_q_proj = nn.Linear(added_kv_proj_dim, self.inner_dim, bias=added_proj_bias) + else: + self.add_q_proj = None + self.add_k_proj = None + self.add_v_proj = None + + if not self.pre_only: + self.to_out = nn.ModuleList([]) + self.to_out.append(nn.Linear(self.inner_dim, self.out_dim, bias=out_bias)) + self.to_out.append(nn.Dropout(dropout)) + else: + self.to_out = None + + if self.context_pre_only is not None and not self.context_pre_only: + self.to_add_out = nn.Linear(self.inner_dim, self.out_context_dim, bias=out_bias) + else: + self.to_add_out = None + + if qk_norm is not None and added_kv_proj_dim is not None: + if qk_norm == "fp32_layer_norm": + self.norm_added_q = FP32LayerNorm(dim_head, elementwise_affine=False, bias=False, eps=eps) + self.norm_added_k = FP32LayerNorm(dim_head, elementwise_affine=False, bias=False, eps=eps) + elif qk_norm == "rms_norm": + self.norm_added_q = RMSNorm(dim_head, eps=eps) + self.norm_added_k = RMSNorm(dim_head, eps=eps) + else: + raise ValueError( + f"unknown qk_norm: {qk_norm}. Should be one of `None,'layer_norm','fp32_layer_norm','rms_norm'`" + ) + else: + self.norm_added_q = None + self.norm_added_k = None + + self.set_processor(processor) + + def set_processor(self, processor: "AttnProcessor") -> None: + r""" + Set the attention processor to use. + + Args: + processor (`AttnProcessor`): + The attention processor to use. + """ + if ( + hasattr(self, "processor") + and isinstance(self.processor, torch.nn.Module) + and not isinstance(processor, torch.nn.Module) + ): + logger.info(f"You are removing possibly trained weights of {self.processor} with {processor}") + self._modules.pop("processor") + + self.processor = processor + + def forward( + self, + hidden_states: torch.Tensor, + encoder_hidden_states: Optional[torch.Tensor] = None, + attention_mask: Optional[torch.Tensor] = None, + **cross_attention_kwargs, + ) -> torch.Tensor: + attn_parameters = set(inspect.signature(self.processor.__call__).parameters.keys()) + quiet_attn_parameters = {"ip_adapter_masks", "ip_hidden_states"} + unused_kwargs = [ + k for k, _ in cross_attention_kwargs.items() if k not in attn_parameters and k not in quiet_attn_parameters + ] + if len(unused_kwargs) > 0: + logger.warning( + f"cross_attention_kwargs {unused_kwargs} are not expected by {self.processor.__class__.__name__} and will be ignored." + ) + cross_attention_kwargs = {k: w for k, w in cross_attention_kwargs.items() if k in attn_parameters} + + return self.processor( + self, + hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=attention_mask, + **cross_attention_kwargs, + ) + + def prepare_attention_mask( + self, attention_mask: torch.Tensor, target_length: int, batch_size: int, out_dim: int = 3 + ) -> torch.Tensor: + head_size = self.heads + if attention_mask is None: + return attention_mask + + current_length: int = attention_mask.shape[-1] + if current_length != target_length: + if attention_mask.device.type == "mps": + padding_shape = (attention_mask.shape[0], attention_mask.shape[1], target_length) + padding = torch.zeros(padding_shape, dtype=attention_mask.dtype, device=attention_mask.device) + attention_mask = torch.cat([attention_mask, padding], dim=2) + else: + attention_mask = F.pad(attention_mask, (0, target_length), value=0.0) + + if out_dim == 3: + if attention_mask.shape[0] < batch_size * head_size: + attention_mask = attention_mask.repeat_interleave(head_size, dim=0) + elif out_dim == 4: + attention_mask = attention_mask.unsqueeze(1) + attention_mask = attention_mask.repeat_interleave(head_size, dim=1) + + return attention_mask + + +class CogVideoXAttnProcessor2_0: + r""" + Processor for implementing scaled dot-product attention for the CogVideoX model. It applies a rotary embedding on + query and key vectors, but does not include spatial normalization. + """ + + def __init__(self): + if not hasattr(F, "scaled_dot_product_attention"): + raise ImportError("CogVideoXAttnProcessor requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.") + + def __call__( + self, + attn: Attention, + hidden_states: torch.Tensor, + encoder_hidden_states: torch.Tensor, + attention_mask: Optional[torch.Tensor] = None, + ) -> torch.Tensor: + text_seq_length = encoder_hidden_states.size(1) + + hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1) + + batch_size, sequence_length, _ = ( + hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape + ) + + if attention_mask is not None: + attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) + attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1]) + + B, S, _ = hidden_states.shape + qkv = attn.to_qkv(hidden_states) + inner_dim = qkv.shape[-1] // 3 + head_dim = inner_dim // attn.heads + qkv_shape = (B, S, 3, attn.heads, head_dim) + query, key, value = qkv.view(qkv_shape).permute(2, 0, 3, 1, 4).contiguous().unbind(0) + + if attn.norm_q is not None: + query = attn.norm_q(query) + if attn.norm_k is not None: + key = attn.norm_k(key) + + B, N, S, D = query.shape + dim = 48 + pad_shape = [B, N, S, D] + pad_shape[-1] = dim - pad_shape[-1] + pad = torch.zeros(pad_shape, dtype=query.dtype, device=query.device) + query = torch.cat([query, pad], dim=-1) + key = torch.cat([key, pad], dim=-1) + value = torch.cat([value, pad], dim=-1) + hidden_states = torch_npu.npu_prompt_flash_attention( + query, + key, + value, + input_layout='BNSD', + scale_value=D**-0.5, + pre_tokens=65535, + next_tokens=65535, + num_heads=N + ) + hidden_states = hidden_states[:, :, :, :D] + + hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim) + + # linear proj + hidden_states = attn.to_out[0](hidden_states) + # dropout + hidden_states = attn.to_out[1](hidden_states) + + encoder_hidden_states, hidden_states = hidden_states.split( + [text_seq_length, hidden_states.size(1) - text_seq_length], dim=1 + ) return hidden_states, encoder_hidden_states \ No newline at end of file diff --git a/MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/models/model_load_utils.py b/MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/models/model_load_utils.py similarity index 97% rename from MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/models/model_load_utils.py rename to MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/models/model_load_utils.py index 34a462528398c0dc7550f58e67e84f0abe9e853a..1257aad30902bd99567b87609e8c5db2d5b06c09 100644 --- a/MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/models/model_load_utils.py +++ b/MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/models/model_load_utils.py @@ -1,42 +1,42 @@ -#!/usr/bin/env python -# coding=utf-8 -# Copyright(C) 2024. Huawei Technologies Co.,Ltd. All rights reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# https://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -import os -import torch -import safetensors.torch - - -SAFETENSORS_EXTENSION = "safetensors" -EMA_STATE_DICT = "ema_state_dict" -STATE_DICT = "state_dict" -CPU = "cpu" - - -def load_state_dict_sd(model_path): - name = os.path.basename(model_path).split('.')[-1] # get weights name - if name.endswith("ckpt"): - weight = torch.load(model_path, map_location=CPU) - if (EMA_STATE_DICT in weight): - weight = weight[EMA_STATE_DICT] - weight = {key.replace("module.", ""): value for key, value in weight.items()} - elif STATE_DICT in weight: - weight = weight[STATE_DICT] - return weight - elif name == SAFETENSORS_EXTENSION: # diffuser model use same name - return safetensors.torch.load_file(model_path, device=CPU) # first load on cpu - else: - # to support hf shard model weights +#!/usr/bin/env python +# coding=utf-8 +# Copyright(C) 2024. Huawei Technologies Co.,Ltd. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# https://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import os +import torch +import safetensors.torch + + +SAFETENSORS_EXTENSION = "safetensors" +EMA_STATE_DICT = "ema_state_dict" +STATE_DICT = "state_dict" +CPU = "cpu" + + +def load_state_dict_sd(model_path): + name = os.path.basename(model_path).split('.')[-1] # get weights name + if name.endswith("ckpt"): + weight = torch.load(model_path, map_location=CPU) + if (EMA_STATE_DICT in weight): + weight = weight[EMA_STATE_DICT] + weight = {key.replace("module.", ""): value for key, value in weight.items()} + elif STATE_DICT in weight: + weight = weight[STATE_DICT] + return weight + elif name == SAFETENSORS_EXTENSION: # diffuser model use same name + return safetensors.torch.load_file(model_path, device=CPU) # first load on cpu + else: + # to support hf shard model weights return torch.load(model_path, map_location=CPU) # first load on cpu \ No newline at end of file diff --git a/MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/models/modeling_utils.py b/MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/models/modeling_utils.py similarity index 97% rename from MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/models/modeling_utils.py rename to MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/models/modeling_utils.py index aa8e33daaa56df927c7b0195e2cabc9d80a2ffd9..fddf0ade3f0c684781e18757e0efc01ea3a279bf 100644 --- a/MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/models/modeling_utils.py +++ b/MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/models/modeling_utils.py @@ -1,771 +1,771 @@ -# coding=utf-8 -# Copyright 2024 The HuggingFace Inc. team. -# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# https://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -import copy -import itertools -import json -import os -import re -from collections import OrderedDict -from functools import wraps -from typing import Any, List, Optional, Tuple, Union - -import torch -from huggingface_hub.utils import validate_hf_hub_args -from torch import Tensor, nn - -from diffusers import __version__ -from diffusers.quantizers import DiffusersAutoQuantizer -from diffusers.quantizers.quantization_config import QuantizationMethod -from diffusers.utils import ( - CONFIG_NAME, - FLAX_WEIGHTS_NAME, - SAFETENSORS_WEIGHTS_NAME, - WEIGHTS_NAME, - _add_variant, - _get_checkpoint_shard_files, - _get_model_file, - deprecate, - is_accelerate_available, - is_bitsandbytes_version, - logging, -) -from diffusers.utils.hub_utils import PushToHubMixin -from diffusers.models.model_loading_utils import ( - _fetch_index_file, - _fetch_index_file_legacy, - _load_state_dict_into_model, - _merge_sharded_checkpoints, - load_model_dict_into_meta, - load_state_dict, -) - - -logger = logging.get_logger(__name__) - - -_LOW_CPU_MEM_USAGE_DEFAULT = True - - -if is_accelerate_available(): - import accelerate - - -def get_parameter_device(parameter: torch.nn.Module) -> torch.device: - try: - parameters_and_buffers = itertools.chain(parameter.parameters(), parameter.buffers()) - return next(parameters_and_buffers).device - except StopIteration: - # For torch.nn.DataParallel compatibility in PyTorch 1.5 - - def find_tensor_attributes(module: torch.nn.Module) -> List[Tuple[str, Tensor]]: - tuples = [(k, v) for k, v in module.__dict__.items() if torch.is_tensor(v)] - return tuples - - gen = parameter._named_members(get_members_fn=find_tensor_attributes) - first_tuple = next(gen) - return first_tuple[1].device - - -def get_parameter_dtype(parameter: torch.nn.Module) -> torch.dtype: - """ - Returns the first found floating dtype in parameters if there is one, otherwise returns the last dtype it found. - """ - last_dtype = None - for param in parameter.parameters(): - last_dtype = param.dtype - if param.is_floating_point(): - return param.dtype - - for buffer in parameter.buffers(): - last_dtype = buffer.dtype - if buffer.is_floating_point(): - return buffer.dtype - - if last_dtype is not None: - # if no floating dtype was found return whatever the first dtype is - return last_dtype - - # For nn.DataParallel compatibility in PyTorch > 1.5 - def find_tensor_attributes(module: nn.Module) -> List[Tuple[str, Tensor]]: - tuples = [(k, v) for k, v in module.__dict__.items() if torch.is_tensor(v)] - return tuples - - gen = parameter._named_members(get_members_fn=find_tensor_attributes) - last_tuple = None - for current_tuple in gen: - last_tuple = current_tuple - if current_tuple[1].is_floating_point(): - return current_tuple[1].dtype - - if last_tuple is not None: - # fallback to the last dtype - return last_tuple[1].dtype - - -class ModelMixin(torch.nn.Module, PushToHubMixin): - config_name = CONFIG_NAME - _automatically_saved_args = ["_diffusers_version", "_class_name", "_name_or_path"] - _supports_gradient_checkpointing = False - _keys_to_ignore_on_load_unexpected = None - _no_split_modules = None - _keep_in_fp32_modules = None - - def __init__(self): - super().__init__() - - def __getattr__(self, name: str) -> Any: - - is_in_config = "_internal_dict" in self.__dict__ and hasattr(self.__dict__["_internal_dict"], name) - is_attribute = name in self.__dict__ - - if is_in_config and not is_attribute: - deprecation_message = f"Accessing config attribute `{name}` directly via '{type(self).__name__}' object attribute is deprecated. Please access '{name}' over '{type(self).__name__}'s config object instead, e.g. 'unet.config.{name}'." - deprecate("direct config name access", "1.0.0", deprecation_message, standard_warn=False, stacklevel=3) - return self._internal_dict[name] - - return super().__getattr__(name) - - @classmethod - @validate_hf_hub_args - def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], **kwargs): - cache_dir = kwargs.pop("cache_dir", None) - ignore_mismatched_sizes = kwargs.pop("ignore_mismatched_sizes", False) - force_download = kwargs.pop("force_download", False) - from_flax = kwargs.pop("from_flax", False) - proxies = kwargs.pop("proxies", None) - output_loading_info = kwargs.pop("output_loading_info", False) - local_files_only = kwargs.pop("local_files_only", None) - token = kwargs.pop("token", None) - revision = kwargs.pop("revision", None) - torch_dtype = kwargs.pop("torch_dtype", None) - subfolder = kwargs.pop("subfolder", None) - device_map = kwargs.pop("device_map", None) - low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT) - variant = kwargs.pop("variant", None) - use_safetensors = kwargs.pop("use_safetensors", None) - quantization_config = kwargs.pop("quantization_config", None) - - allow_pickle = False - if use_safetensors is None: - use_safetensors = True - allow_pickle = True - - if low_cpu_mem_usage and not is_accelerate_available(): - low_cpu_mem_usage = False - logger.warning( - "Cannot initialize model with low cpu memory usage because `accelerate` was not found in the" - " environment. Defaulting to `low_cpu_mem_usage=False`. It is strongly recommended to install" - " `accelerate` for faster and less memory-intense model loading. You can do so with: \n```\npip" - " install accelerate\n```\n." - ) - - if device_map is not None and not is_accelerate_available(): - raise NotImplementedError( - "Loading and dispatching requires `accelerate`. Please make sure to install accelerate or set" - " `device_map=None`. You can install accelerate with `pip install accelerate`." - ) - - if low_cpu_mem_usage is False and device_map is not None: - raise ValueError( - f"You cannot set `low_cpu_mem_usage` to `False` while using device_map={device_map} for loading and" - " dispatching. Please make sure to set `low_cpu_mem_usage=True`." - ) - - if isinstance(device_map, torch.device): - device_map = {"": device_map} - elif isinstance(device_map, str) and device_map not in ["auto", "balanced", "balanced_low_0", "sequential"]: - try: - device_map = {"": torch.device(device_map)} - except RuntimeError as e: - raise ValueError( - "When passing device_map as a string, the value needs to be a device name (e.g. cpu, cuda:0) or " - f"'auto', 'balanced', 'balanced_low_0', 'sequential' but found {device_map}." - ) from e - elif isinstance(device_map, int): - if device_map < 0: - raise ValueError( - "You can't pass device_map as a negative int. If you want to put the model on the cpu, pass device_map = 'cpu' " - ) - else: - device_map = {"": device_map} - - if device_map is not None: - if low_cpu_mem_usage is None: - low_cpu_mem_usage = True - elif not low_cpu_mem_usage: - raise ValueError("Passing along a `device_map` requires `low_cpu_mem_usage=True`") - - # Load config if we don't provide a configuration - config_path = pretrained_model_name_or_path - - user_agent = { - "diffusers": __version__, - "file_type": "model", - "framework": "pytorch", - } - - # load config - config, unused_kwargs, commit_hash = cls.load_config( - config_path, - cache_dir=cache_dir, - return_unused_kwargs=True, - return_commit_hash=True, - force_download=force_download, - proxies=proxies, - local_files_only=local_files_only, - token=token, - revision=revision, - subfolder=subfolder, - user_agent=user_agent, - **kwargs, - ) - # no in-place modification of the original config. - config = copy.deepcopy(config) - - # determine initial quantization config. - ####################################### - pre_quantized = "quantization_config" in config and config["quantization_config"] is not None - if pre_quantized or quantization_config is not None: - if pre_quantized: - config["quantization_config"] = DiffusersAutoQuantizer.merge_quantization_configs( - config["quantization_config"], quantization_config - ) - else: - config["quantization_config"] = quantization_config - hf_quantizer = DiffusersAutoQuantizer.from_config( - config["quantization_config"], pre_quantized=pre_quantized - ) - else: - hf_quantizer = None - - if hf_quantizer is not None: - is_bnb_quantization_method = hf_quantizer.quantization_config.quant_method.value == "bitsandbytes" - if is_bnb_quantization_method and device_map is not None: - raise NotImplementedError( - "Currently, `device_map` is automatically inferred for quantized bitsandbytes models. Support for providing `device_map` as an input will be added in the future." - ) - - hf_quantizer.validate_environment(torch_dtype=torch_dtype, from_flax=from_flax, device_map=device_map) - torch_dtype = hf_quantizer.update_torch_dtype(torch_dtype) - - # In order to ensure popular quantization methods are supported. Can be disable with `disable_telemetry` - user_agent["quant"] = hf_quantizer.quantization_config.quant_method.value - - # Force-set to `True` for more mem efficiency - if low_cpu_mem_usage is None: - low_cpu_mem_usage = True - logger.info("Set `low_cpu_mem_usage` to True as `hf_quantizer` is not None.") - elif not low_cpu_mem_usage: - raise ValueError("`low_cpu_mem_usage` cannot be False or None when using quantization.") - - # Check if `_keep_in_fp32_modules` is not None - use_keep_in_fp32_modules = (cls._keep_in_fp32_modules is not None) and ( - (torch_dtype == torch.float16) or hasattr(hf_quantizer, "use_keep_in_fp32_modules") - ) - if use_keep_in_fp32_modules: - keep_in_fp32_modules = cls._keep_in_fp32_modules - if not isinstance(keep_in_fp32_modules, list): - keep_in_fp32_modules = [keep_in_fp32_modules] - - if low_cpu_mem_usage is None: - low_cpu_mem_usage = True - logger.info("Set `low_cpu_mem_usage` to True as `_keep_in_fp32_modules` is not None.") - elif not low_cpu_mem_usage: - raise ValueError("`low_cpu_mem_usage` cannot be False when `keep_in_fp32_modules` is True.") - else: - keep_in_fp32_modules = [] - ####################################### - - # Determine if we're loading from a directory of sharded checkpoints. - is_sharded = False - index_file = None - is_local = os.path.isdir(pretrained_model_name_or_path) - index_file_kwargs = { - "is_local": is_local, - "pretrained_model_name_or_path": pretrained_model_name_or_path, - "subfolder": subfolder or "", - "use_safetensors": use_safetensors, - "cache_dir": cache_dir, - "variant": variant, - "force_download": force_download, - "proxies": proxies, - "local_files_only": local_files_only, - "token": token, - "revision": revision, - "user_agent": user_agent, - "commit_hash": commit_hash, - } - index_file = _fetch_index_file(**index_file_kwargs) - # In case the index file was not found we still have to consider the legacy format. - # this becomes applicable when the variant is not None. - if variant is not None and (index_file is None or not os.path.exists(index_file)): - index_file = _fetch_index_file_legacy(**index_file_kwargs) - if index_file is not None and index_file.is_file(): - is_sharded = True - - if is_sharded and from_flax: - raise ValueError("Loading of sharded checkpoints is not supported when `from_flax=True`.") - - # load model - model_file = None - if from_flax: - model_file = _get_model_file( - pretrained_model_name_or_path, - weights_name=FLAX_WEIGHTS_NAME, - cache_dir=cache_dir, - force_download=force_download, - proxies=proxies, - local_files_only=local_files_only, - token=token, - revision=revision, - subfolder=subfolder, - user_agent=user_agent, - commit_hash=commit_hash, - ) - model = cls.from_config(config, **unused_kwargs) - - # Convert the weights - from .modeling_pytorch_flax_utils import load_flax_checkpoint_in_pytorch_model - - model = load_flax_checkpoint_in_pytorch_model(model, model_file) - else: - if is_sharded: - sharded_ckpt_cached_folder, sharded_metadata = _get_checkpoint_shard_files( - pretrained_model_name_or_path, - index_file, - cache_dir=cache_dir, - proxies=proxies, - local_files_only=local_files_only, - token=token, - user_agent=user_agent, - revision=revision, - subfolder=subfolder or "", - ) - if hf_quantizer is not None and is_bnb_quantization_method: - model_file = _merge_sharded_checkpoints(sharded_ckpt_cached_folder, sharded_metadata) - logger.info("Merged sharded checkpoints as `hf_quantizer` is not None.") - is_sharded = False - - elif use_safetensors and not is_sharded: - try: - model_file = _get_model_file( - pretrained_model_name_or_path, - weights_name=_add_variant(SAFETENSORS_WEIGHTS_NAME, variant), - cache_dir=cache_dir, - force_download=force_download, - proxies=proxies, - local_files_only=local_files_only, - token=token, - revision=revision, - subfolder=subfolder, - user_agent=user_agent, - commit_hash=commit_hash, - ) - - except IOError as e: - logger.error(f"An error occurred while trying to fetch {pretrained_model_name_or_path}: {e}") - if not allow_pickle: - raise - logger.warning( - "Defaulting to unsafe serialization. Pass `allow_pickle=False` to raise an error instead." - ) - - if model_file is None and not is_sharded: - model_file = _get_model_file( - pretrained_model_name_or_path, - weights_name=_add_variant(WEIGHTS_NAME, variant), - cache_dir=cache_dir, - force_download=force_download, - proxies=proxies, - local_files_only=local_files_only, - token=token, - revision=revision, - subfolder=subfolder, - user_agent=user_agent, - commit_hash=commit_hash, - ) - - if low_cpu_mem_usage: - # Instantiate model with empty weights - with accelerate.init_empty_weights(): - model = cls.from_config(config, **unused_kwargs) - - if hf_quantizer is not None: - hf_quantizer.preprocess_model( - model=model, device_map=device_map, keep_in_fp32_modules=keep_in_fp32_modules - ) - - # if device_map is None, load the state dict and move the params from meta device to the cpu - if device_map is None and not is_sharded: - # `torch.cuda.current_device()` is fine here when `hf_quantizer` is not None. - # It would error out during the `validate_environment()` call above in the absence of cuda. - if hf_quantizer is None: - param_device = "cpu" - else: - param_device = torch.device(torch.cuda.current_device()) - state_dict = load_state_dict(model_file, variant=variant) - model._convert_deprecated_attention_blocks(state_dict) - - # move the params from meta device to cpu - missing_keys = set(model.state_dict().keys()) - set(state_dict.keys()) - if hf_quantizer is not None: - missing_keys = hf_quantizer.update_missing_keys(model, missing_keys, prefix="") - if len(missing_keys) > 0: - raise ValueError( - f"Cannot load {cls} from {pretrained_model_name_or_path} because the following keys are" - f" missing: \n {', '.join(missing_keys)}. \n Please make sure to pass" - " `low_cpu_mem_usage=False` and `device_map=None` if you want to randomly initialize" - " those weights or else make sure your checkpoint file is correct." - ) - - unexpected_keys = load_model_dict_into_meta( - model, - state_dict, - device=param_device, - dtype=torch_dtype, - model_name_or_path=pretrained_model_name_or_path, - hf_quantizer=hf_quantizer, - keep_in_fp32_modules=keep_in_fp32_modules, - ) - - if cls._keys_to_ignore_on_load_unexpected is not None: - for pat in cls._keys_to_ignore_on_load_unexpected: - unexpected_keys = [k for k in unexpected_keys if re.search(pat, k) is None] - - if len(unexpected_keys) > 0: - logger.warning( - f"Some weights of the model checkpoint were not used when initializing {cls.__name__}: \n {[', '.join(unexpected_keys)]}" - ) - - else: - weights_path = index_file - with open(index_file) as f: - index = json.loads(f.read()) - if "weight_map" in index: - index = index["weight_map"] - weights_path = sorted(list(set(index.values()))) - weights_path = [os.path.join(pretrained_model_name_or_path, f) for f in weights_path] - - model = cls._load_model(model, weights_path, is_sharded) - - loading_info = { - "missing_keys": [], - "unexpected_keys": [], - "mismatched_keys": [], - "error_msgs": [], - } - else: - model = cls.from_config(config, **unused_kwargs) - - state_dict = load_state_dict(model_file, variant=variant) - model._convert_deprecated_attention_blocks(state_dict) - - model, missing_keys, unexpected_keys, mismatched_keys, error_msgs = cls._load_pretrained_model( - model, - state_dict, - model_file, - pretrained_model_name_or_path, - ignore_mismatched_sizes=ignore_mismatched_sizes, - ) - - loading_info = { - "missing_keys": missing_keys, - "unexpected_keys": unexpected_keys, - "mismatched_keys": mismatched_keys, - "error_msgs": error_msgs, - } - - if hf_quantizer is not None: - hf_quantizer.postprocess_model(model) - model.hf_quantizer = hf_quantizer - - if torch_dtype is not None and not isinstance(torch_dtype, torch.dtype): - raise ValueError( - f"{torch_dtype} needs to be of type `torch.dtype`, e.g. `torch.float16`, but is {type(torch_dtype)}." - ) - # When using `use_keep_in_fp32_modules` if we do a global `to()` here, then we will - # completely lose the effectivity of `use_keep_in_fp32_modules`. - elif torch_dtype is not None and hf_quantizer is None and not use_keep_in_fp32_modules: - model = model.to(torch_dtype) - - if hf_quantizer is not None: - # We also make sure to purge `_pre_quantization_dtype` when we serialize - # the model config because `_pre_quantization_dtype` is `torch.dtype`, not JSON serializable. - model.register_to_config(_name_or_path=pretrained_model_name_or_path, _pre_quantization_dtype=torch_dtype) - else: - model.register_to_config(_name_or_path=pretrained_model_name_or_path) - - # Set model in evaluation mode to deactivate DropOut modules by default - model.eval() - if output_loading_info: - return model, loading_info - - return model - - @classmethod - def _load_model(cls, model, weights_path, is_sharded): - if not is_sharded: - state_dict = load_state_dict(weights_path) - model.load_weights(state_dict) - else: - need_key = set(model.state_dict().keys()) - state_dict = {} - cache = {} - for weight_file in weights_path: - state_dict = load_state_dict(weight_file) - state_dict.update(cache) - loadkey_cache = model.load_weights(state_dict, is_sharded) - if loadkey_cache : - if isinstance(loadkey_cache, tuple): - loaded_keys, cache = loadkey_cache - else: - loaded_keys = loadkey_cache - need_key = need_key.symmetric_difference(set(loaded_keys)) - - if len(need_key) > 0: - raise ValueError(f"The weight miss key: {need_key}") - return model - - def load_weights(self, state_dict, shard=False): - with torch.no_grad(): - if not shard: - self.load_state_dict(state_dict) - return {} - else: - self.load_state_dict(state_dict, strict=False, assign=True) - return state_dict.keys() - - # Adapted from `transformers`. - @wraps(torch.nn.Module.cuda) - def cuda(self, *args, **kwargs): - # Checks if the model has been loaded in 4-bit or 8-bit with BNB - if getattr(self, "quantization_method", None) == QuantizationMethod.BITS_AND_BYTES: - if getattr(self, "is_loaded_in_8bit", False): - raise ValueError( - "Calling `cuda()` is not supported for `8-bit` quantized models. " - " Please use the model as it is, since the model has already been set to the correct devices." - ) - elif is_bitsandbytes_version("<", "0.43.2"): - raise ValueError( - "Calling `cuda()` is not supported for `4-bit` quantized models with the installed version of bitsandbytes. " - f"The current device is `{self.device}`. If you intended to move the model, please install bitsandbytes >= 0.43.2." - ) - return super().cuda(*args, **kwargs) - - # Adapted from `transformers`. - @wraps(torch.nn.Module.to) - def to(self, *args, **kwargs): - dtype_present_in_args = "dtype" in kwargs - - if not dtype_present_in_args: - for arg in args: - if isinstance(arg, torch.dtype): - dtype_present_in_args = True - break - - if getattr(self, "is_quantized", False): - if dtype_present_in_args: - raise ValueError( - "Casting a quantized model to a new `dtype` is unsupported. To set the dtype of unquantized layers, please " - "use the `torch_dtype` argument when loading the model using `from_pretrained` or `from_single_file`" - ) - - if getattr(self, "quantization_method", None) == QuantizationMethod.BITS_AND_BYTES: - if getattr(self, "is_loaded_in_8bit", False): - raise ValueError( - "`.to` is not supported for `8-bit` bitsandbytes models. Please use the model as it is, since the" - " model has already been set to the correct devices and casted to the correct `dtype`." - ) - elif is_bitsandbytes_version("<", "0.43.2"): - raise ValueError( - "Calling `to()` is not supported for `4-bit` quantized models with the installed version of bitsandbytes. " - f"The current device is `{self.device}`. If you intended to move the model, please install bitsandbytes >= 0.43.2." - ) - return super().to(*args, **kwargs) - - # Taken from `transformers`. - def half(self, *args): - # Checks if the model is quantized - if getattr(self, "is_quantized", False): - raise ValueError( - "`.half()` is not supported for quantized model. Please use the model as it is, since the" - " model has already been cast to the correct `dtype`." - ) - else: - return super().half(*args) - - # Taken from `transformers`. - def float(self, *args): - # Checks if the model is quantized - if getattr(self, "is_quantized", False): - raise ValueError( - "`.float()` is not supported for quantized model. Please use the model as it is, since the" - " model has already been cast to the correct `dtype`." - ) - else: - return super().float(*args) - - @classmethod - def _load_pretrained_model( - cls, - model, - state_dict: OrderedDict, - pretrained_model_name_or_path: Union[str, os.PathLike], - ignore_mismatched_sizes: bool = False, - ): - # Retrieve missing & unexpected_keys - model_state_dict = model.state_dict() - loaded_keys = list(state_dict.keys()) - - expected_keys = list(model_state_dict.keys()) - - original_loaded_keys = loaded_keys - - missing_keys = list(set(expected_keys) - set(loaded_keys)) - unexpected_keys = list(set(loaded_keys) - set(expected_keys)) - - # Make sure we are able to load base models as well as derived models (with heads) - model_to_load = model - - def _find_mismatched_keys( - state_dict, - model_state_dict, - loaded_keys, - ignore_mismatched_sizes, - ): - mismatched_keys = [] - if ignore_mismatched_sizes: - for checkpoint_key in loaded_keys: - model_key = checkpoint_key - - if ( - model_key in model_state_dict - and state_dict[checkpoint_key].shape != model_state_dict[model_key].shape - ): - mismatched_keys.append( - (checkpoint_key, state_dict[checkpoint_key].shape, model_state_dict[model_key].shape) - ) - del state_dict[checkpoint_key] - return mismatched_keys - - if state_dict is not None: - # Whole checkpoint - mismatched_keys = _find_mismatched_keys( - state_dict, - model_state_dict, - original_loaded_keys, - ignore_mismatched_sizes, - ) - error_msgs = _load_state_dict_into_model(model_to_load, state_dict) - - if len(error_msgs) > 0: - error_msg = "\n\t".join(error_msgs) - if "size mismatch" in error_msg: - error_msg += ( - "\n\tYou may consider adding `ignore_mismatched_sizes=True` in the model `from_pretrained` method." - ) - raise RuntimeError(f"Error(s) in loading state_dict for {model.__class__.__name__}:\n\t{error_msg}") - - if len(unexpected_keys) > 0: - logger.warning( - f"Some weights of the model checkpoint at {pretrained_model_name_or_path} were not used when" - f" initializing {model.__class__.__name__}: {unexpected_keys}\n- This IS expected if you are" - f" initializing {model.__class__.__name__} from the checkpoint of a model trained on another task" - " or with another architecture (e.g. initializing a BertForSequenceClassification model from a" - " BertForPreTraining model).\n- This IS NOT expected if you are initializing" - f" {model.__class__.__name__} from the checkpoint of a model that you expect to be exactly" - " identical (initializing a BertForSequenceClassification model from a" - " BertForSequenceClassification model)." - ) - else: - logger.info(f"All model checkpoint weights were used when initializing {model.__class__.__name__}.\n") - if len(missing_keys) > 0: - logger.warning( - f"Some weights of {model.__class__.__name__} were not initialized from the model checkpoint at" - f" {pretrained_model_name_or_path} and are newly initialized: {missing_keys}\nYou should probably" - " TRAIN this model on a down-stream task to be able to use it for predictions and inference." - ) - elif len(mismatched_keys) == 0: - logger.info( - f"All the weights of {model.__class__.__name__} were initialized from the model checkpoint at" - f" {pretrained_model_name_or_path}.\nIf your task is similar to the task the model of the" - f" checkpoint was trained on, you can already use {model.__class__.__name__} for predictions" - " without further training." - ) - if len(mismatched_keys) > 0: - mismatched_warning = "\n".join( - [ - f"- {key}: found shape {shape1} in the checkpoint and {shape2} in the model instantiated" - for key, shape1, shape2 in mismatched_keys - ] - ) - logger.warning( - f"Some weights of {model.__class__.__name__} were not initialized from the model checkpoint at" - f" {pretrained_model_name_or_path} and are newly initialized because the shapes did not" - f" match:\n{mismatched_warning}\nYou should probably TRAIN this model on a down-stream task to be" - " able to use it for predictions and inference." - ) - - return model, missing_keys, unexpected_keys, mismatched_keys, error_msgs - - @property - def device(self) -> torch.device: - return get_parameter_device(self) - - @property - def dtype(self) -> torch.dtype: - return get_parameter_dtype(self) - - def _convert_deprecated_attention_blocks(self, state_dict: OrderedDict) -> None: - deprecated_attention_block_paths = [] - - def recursive_find_attn_block(name, module): - if hasattr(module, "_from_deprecated_attn_block") and module._from_deprecated_attn_block: - deprecated_attention_block_paths.append(name) - - for sub_name, sub_module in module.named_children(): - sub_name = sub_name if name == "" else f"{name}.{sub_name}" - recursive_find_attn_block(sub_name, sub_module) - - recursive_find_attn_block("", self) - - for path in deprecated_attention_block_paths: - # group_norm path stays the same - - # query -> to_q - if f"{path}.query.weight" in state_dict: - state_dict[f"{path}.to_q.weight"] = state_dict.pop(f"{path}.query.weight") - if f"{path}.query.bias" in state_dict: - state_dict[f"{path}.to_q.bias"] = state_dict.pop(f"{path}.query.bias") - - # key -> to_k - if f"{path}.key.weight" in state_dict: - state_dict[f"{path}.to_k.weight"] = state_dict.pop(f"{path}.key.weight") - if f"{path}.key.bias" in state_dict: - state_dict[f"{path}.to_k.bias"] = state_dict.pop(f"{path}.key.bias") - - # value -> to_v - if f"{path}.value.weight" in state_dict: - state_dict[f"{path}.to_v.weight"] = state_dict.pop(f"{path}.value.weight") - if f"{path}.value.bias" in state_dict: - state_dict[f"{path}.to_v.bias"] = state_dict.pop(f"{path}.value.bias") - - # proj_attn -> to_out.0 - if f"{path}.proj_attn.weight" in state_dict: - state_dict[f"{path}.to_out.0.weight"] = state_dict.pop(f"{path}.proj_attn.weight") - if f"{path}.proj_attn.bias" in state_dict: - state_dict[f"{path}.to_out.0.bias"] = state_dict.pop(f"{path}.proj_attn.bias") +# coding=utf-8 +# Copyright 2024 The HuggingFace Inc. team. +# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# https://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import copy +import itertools +import json +import os +import re +from collections import OrderedDict +from functools import wraps +from typing import Any, List, Optional, Tuple, Union + +import torch +from huggingface_hub.utils import validate_hf_hub_args +from torch import Tensor, nn + +from diffusers import __version__ +from diffusers.quantizers import DiffusersAutoQuantizer +from diffusers.quantizers.quantization_config import QuantizationMethod +from diffusers.utils import ( + CONFIG_NAME, + FLAX_WEIGHTS_NAME, + SAFETENSORS_WEIGHTS_NAME, + WEIGHTS_NAME, + _add_variant, + _get_checkpoint_shard_files, + _get_model_file, + deprecate, + is_accelerate_available, + is_bitsandbytes_version, + logging, +) +from diffusers.utils.hub_utils import PushToHubMixin +from diffusers.models.model_loading_utils import ( + _fetch_index_file, + _fetch_index_file_legacy, + _load_state_dict_into_model, + _merge_sharded_checkpoints, + load_model_dict_into_meta, + load_state_dict, +) + + +logger = logging.get_logger(__name__) + + +_LOW_CPU_MEM_USAGE_DEFAULT = True + + +if is_accelerate_available(): + import accelerate + + +def get_parameter_device(parameter: torch.nn.Module) -> torch.device: + try: + parameters_and_buffers = itertools.chain(parameter.parameters(), parameter.buffers()) + return next(parameters_and_buffers).device + except StopIteration: + # For torch.nn.DataParallel compatibility in PyTorch 1.5 + + def find_tensor_attributes(module: torch.nn.Module) -> List[Tuple[str, Tensor]]: + tuples = [(k, v) for k, v in module.__dict__.items() if torch.is_tensor(v)] + return tuples + + gen = parameter._named_members(get_members_fn=find_tensor_attributes) + first_tuple = next(gen) + return first_tuple[1].device + + +def get_parameter_dtype(parameter: torch.nn.Module) -> torch.dtype: + """ + Returns the first found floating dtype in parameters if there is one, otherwise returns the last dtype it found. + """ + last_dtype = None + for param in parameter.parameters(): + last_dtype = param.dtype + if param.is_floating_point(): + return param.dtype + + for buffer in parameter.buffers(): + last_dtype = buffer.dtype + if buffer.is_floating_point(): + return buffer.dtype + + if last_dtype is not None: + # if no floating dtype was found return whatever the first dtype is + return last_dtype + + # For nn.DataParallel compatibility in PyTorch > 1.5 + def find_tensor_attributes(module: nn.Module) -> List[Tuple[str, Tensor]]: + tuples = [(k, v) for k, v in module.__dict__.items() if torch.is_tensor(v)] + return tuples + + gen = parameter._named_members(get_members_fn=find_tensor_attributes) + last_tuple = None + for current_tuple in gen: + last_tuple = current_tuple + if current_tuple[1].is_floating_point(): + return current_tuple[1].dtype + + if last_tuple is not None: + # fallback to the last dtype + return last_tuple[1].dtype + + +class ModelMixin(torch.nn.Module, PushToHubMixin): + config_name = CONFIG_NAME + _automatically_saved_args = ["_diffusers_version", "_class_name", "_name_or_path"] + _supports_gradient_checkpointing = False + _keys_to_ignore_on_load_unexpected = None + _no_split_modules = None + _keep_in_fp32_modules = None + + def __init__(self): + super().__init__() + + def __getattr__(self, name: str) -> Any: + + is_in_config = "_internal_dict" in self.__dict__ and hasattr(self.__dict__["_internal_dict"], name) + is_attribute = name in self.__dict__ + + if is_in_config and not is_attribute: + deprecation_message = f"Accessing config attribute `{name}` directly via '{type(self).__name__}' object attribute is deprecated. Please access '{name}' over '{type(self).__name__}'s config object instead, e.g. 'unet.config.{name}'." + deprecate("direct config name access", "1.0.0", deprecation_message, standard_warn=False, stacklevel=3) + return self._internal_dict[name] + + return super().__getattr__(name) + + @classmethod + @validate_hf_hub_args + def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], **kwargs): + cache_dir = kwargs.pop("cache_dir", None) + ignore_mismatched_sizes = kwargs.pop("ignore_mismatched_sizes", False) + force_download = kwargs.pop("force_download", False) + from_flax = kwargs.pop("from_flax", False) + proxies = kwargs.pop("proxies", None) + output_loading_info = kwargs.pop("output_loading_info", False) + local_files_only = kwargs.pop("local_files_only", None) + token = kwargs.pop("token", None) + revision = kwargs.pop("revision", None) + torch_dtype = kwargs.pop("torch_dtype", None) + subfolder = kwargs.pop("subfolder", None) + device_map = kwargs.pop("device_map", None) + low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT) + variant = kwargs.pop("variant", None) + use_safetensors = kwargs.pop("use_safetensors", None) + quantization_config = kwargs.pop("quantization_config", None) + + allow_pickle = False + if use_safetensors is None: + use_safetensors = True + allow_pickle = True + + if low_cpu_mem_usage and not is_accelerate_available(): + low_cpu_mem_usage = False + logger.warning( + "Cannot initialize model with low cpu memory usage because `accelerate` was not found in the" + " environment. Defaulting to `low_cpu_mem_usage=False`. It is strongly recommended to install" + " `accelerate` for faster and less memory-intense model loading. You can do so with: \n```\npip" + " install accelerate\n```\n." + ) + + if device_map is not None and not is_accelerate_available(): + raise NotImplementedError( + "Loading and dispatching requires `accelerate`. Please make sure to install accelerate or set" + " `device_map=None`. You can install accelerate with `pip install accelerate`." + ) + + if low_cpu_mem_usage is False and device_map is not None: + raise ValueError( + f"You cannot set `low_cpu_mem_usage` to `False` while using device_map={device_map} for loading and" + " dispatching. Please make sure to set `low_cpu_mem_usage=True`." + ) + + if isinstance(device_map, torch.device): + device_map = {"": device_map} + elif isinstance(device_map, str) and device_map not in ["auto", "balanced", "balanced_low_0", "sequential"]: + try: + device_map = {"": torch.device(device_map)} + except RuntimeError as e: + raise ValueError( + "When passing device_map as a string, the value needs to be a device name (e.g. cpu, cuda:0) or " + f"'auto', 'balanced', 'balanced_low_0', 'sequential' but found {device_map}." + ) from e + elif isinstance(device_map, int): + if device_map < 0: + raise ValueError( + "You can't pass device_map as a negative int. If you want to put the model on the cpu, pass device_map = 'cpu' " + ) + else: + device_map = {"": device_map} + + if device_map is not None: + if low_cpu_mem_usage is None: + low_cpu_mem_usage = True + elif not low_cpu_mem_usage: + raise ValueError("Passing along a `device_map` requires `low_cpu_mem_usage=True`") + + # Load config if we don't provide a configuration + config_path = pretrained_model_name_or_path + + user_agent = { + "diffusers": __version__, + "file_type": "model", + "framework": "pytorch", + } + + # load config + config, unused_kwargs, commit_hash = cls.load_config( + config_path, + cache_dir=cache_dir, + return_unused_kwargs=True, + return_commit_hash=True, + force_download=force_download, + proxies=proxies, + local_files_only=local_files_only, + token=token, + revision=revision, + subfolder=subfolder, + user_agent=user_agent, + **kwargs, + ) + # no in-place modification of the original config. + config = copy.deepcopy(config) + + # determine initial quantization config. + ####################################### + pre_quantized = "quantization_config" in config and config["quantization_config"] is not None + if pre_quantized or quantization_config is not None: + if pre_quantized: + config["quantization_config"] = DiffusersAutoQuantizer.merge_quantization_configs( + config["quantization_config"], quantization_config + ) + else: + config["quantization_config"] = quantization_config + hf_quantizer = DiffusersAutoQuantizer.from_config( + config["quantization_config"], pre_quantized=pre_quantized + ) + else: + hf_quantizer = None + + if hf_quantizer is not None: + is_bnb_quantization_method = hf_quantizer.quantization_config.quant_method.value == "bitsandbytes" + if is_bnb_quantization_method and device_map is not None: + raise NotImplementedError( + "Currently, `device_map` is automatically inferred for quantized bitsandbytes models. Support for providing `device_map` as an input will be added in the future." + ) + + hf_quantizer.validate_environment(torch_dtype=torch_dtype, from_flax=from_flax, device_map=device_map) + torch_dtype = hf_quantizer.update_torch_dtype(torch_dtype) + + # In order to ensure popular quantization methods are supported. Can be disable with `disable_telemetry` + user_agent["quant"] = hf_quantizer.quantization_config.quant_method.value + + # Force-set to `True` for more mem efficiency + if low_cpu_mem_usage is None: + low_cpu_mem_usage = True + logger.info("Set `low_cpu_mem_usage` to True as `hf_quantizer` is not None.") + elif not low_cpu_mem_usage: + raise ValueError("`low_cpu_mem_usage` cannot be False or None when using quantization.") + + # Check if `_keep_in_fp32_modules` is not None + use_keep_in_fp32_modules = (cls._keep_in_fp32_modules is not None) and ( + (torch_dtype == torch.float16) or hasattr(hf_quantizer, "use_keep_in_fp32_modules") + ) + if use_keep_in_fp32_modules: + keep_in_fp32_modules = cls._keep_in_fp32_modules + if not isinstance(keep_in_fp32_modules, list): + keep_in_fp32_modules = [keep_in_fp32_modules] + + if low_cpu_mem_usage is None: + low_cpu_mem_usage = True + logger.info("Set `low_cpu_mem_usage` to True as `_keep_in_fp32_modules` is not None.") + elif not low_cpu_mem_usage: + raise ValueError("`low_cpu_mem_usage` cannot be False when `keep_in_fp32_modules` is True.") + else: + keep_in_fp32_modules = [] + ####################################### + + # Determine if we're loading from a directory of sharded checkpoints. + is_sharded = False + index_file = None + is_local = os.path.isdir(pretrained_model_name_or_path) + index_file_kwargs = { + "is_local": is_local, + "pretrained_model_name_or_path": pretrained_model_name_or_path, + "subfolder": subfolder or "", + "use_safetensors": use_safetensors, + "cache_dir": cache_dir, + "variant": variant, + "force_download": force_download, + "proxies": proxies, + "local_files_only": local_files_only, + "token": token, + "revision": revision, + "user_agent": user_agent, + "commit_hash": commit_hash, + } + index_file = _fetch_index_file(**index_file_kwargs) + # In case the index file was not found we still have to consider the legacy format. + # this becomes applicable when the variant is not None. + if variant is not None and (index_file is None or not os.path.exists(index_file)): + index_file = _fetch_index_file_legacy(**index_file_kwargs) + if index_file is not None and index_file.is_file(): + is_sharded = True + + if is_sharded and from_flax: + raise ValueError("Loading of sharded checkpoints is not supported when `from_flax=True`.") + + # load model + model_file = None + if from_flax: + model_file = _get_model_file( + pretrained_model_name_or_path, + weights_name=FLAX_WEIGHTS_NAME, + cache_dir=cache_dir, + force_download=force_download, + proxies=proxies, + local_files_only=local_files_only, + token=token, + revision=revision, + subfolder=subfolder, + user_agent=user_agent, + commit_hash=commit_hash, + ) + model = cls.from_config(config, **unused_kwargs) + + # Convert the weights + from .modeling_pytorch_flax_utils import load_flax_checkpoint_in_pytorch_model + + model = load_flax_checkpoint_in_pytorch_model(model, model_file) + else: + if is_sharded: + sharded_ckpt_cached_folder, sharded_metadata = _get_checkpoint_shard_files( + pretrained_model_name_or_path, + index_file, + cache_dir=cache_dir, + proxies=proxies, + local_files_only=local_files_only, + token=token, + user_agent=user_agent, + revision=revision, + subfolder=subfolder or "", + ) + if hf_quantizer is not None and is_bnb_quantization_method: + model_file = _merge_sharded_checkpoints(sharded_ckpt_cached_folder, sharded_metadata) + logger.info("Merged sharded checkpoints as `hf_quantizer` is not None.") + is_sharded = False + + elif use_safetensors and not is_sharded: + try: + model_file = _get_model_file( + pretrained_model_name_or_path, + weights_name=_add_variant(SAFETENSORS_WEIGHTS_NAME, variant), + cache_dir=cache_dir, + force_download=force_download, + proxies=proxies, + local_files_only=local_files_only, + token=token, + revision=revision, + subfolder=subfolder, + user_agent=user_agent, + commit_hash=commit_hash, + ) + + except IOError as e: + logger.error(f"An error occurred while trying to fetch {pretrained_model_name_or_path}: {e}") + if not allow_pickle: + raise + logger.warning( + "Defaulting to unsafe serialization. Pass `allow_pickle=False` to raise an error instead." + ) + + if model_file is None and not is_sharded: + model_file = _get_model_file( + pretrained_model_name_or_path, + weights_name=_add_variant(WEIGHTS_NAME, variant), + cache_dir=cache_dir, + force_download=force_download, + proxies=proxies, + local_files_only=local_files_only, + token=token, + revision=revision, + subfolder=subfolder, + user_agent=user_agent, + commit_hash=commit_hash, + ) + + if low_cpu_mem_usage: + # Instantiate model with empty weights + with accelerate.init_empty_weights(): + model = cls.from_config(config, **unused_kwargs) + + if hf_quantizer is not None: + hf_quantizer.preprocess_model( + model=model, device_map=device_map, keep_in_fp32_modules=keep_in_fp32_modules + ) + + # if device_map is None, load the state dict and move the params from meta device to the cpu + if device_map is None and not is_sharded: + # `torch.cuda.current_device()` is fine here when `hf_quantizer` is not None. + # It would error out during the `validate_environment()` call above in the absence of cuda. + if hf_quantizer is None: + param_device = "cpu" + else: + param_device = torch.device(torch.cuda.current_device()) + state_dict = load_state_dict(model_file, variant=variant) + model._convert_deprecated_attention_blocks(state_dict) + + # move the params from meta device to cpu + missing_keys = set(model.state_dict().keys()) - set(state_dict.keys()) + if hf_quantizer is not None: + missing_keys = hf_quantizer.update_missing_keys(model, missing_keys, prefix="") + if len(missing_keys) > 0: + raise ValueError( + f"Cannot load {cls} from {pretrained_model_name_or_path} because the following keys are" + f" missing: \n {', '.join(missing_keys)}. \n Please make sure to pass" + " `low_cpu_mem_usage=False` and `device_map=None` if you want to randomly initialize" + " those weights or else make sure your checkpoint file is correct." + ) + + unexpected_keys = load_model_dict_into_meta( + model, + state_dict, + device=param_device, + dtype=torch_dtype, + model_name_or_path=pretrained_model_name_or_path, + hf_quantizer=hf_quantizer, + keep_in_fp32_modules=keep_in_fp32_modules, + ) + + if cls._keys_to_ignore_on_load_unexpected is not None: + for pat in cls._keys_to_ignore_on_load_unexpected: + unexpected_keys = [k for k in unexpected_keys if re.search(pat, k) is None] + + if len(unexpected_keys) > 0: + logger.warning( + f"Some weights of the model checkpoint were not used when initializing {cls.__name__}: \n {[', '.join(unexpected_keys)]}" + ) + + else: + weights_path = index_file + with open(index_file) as f: + index = json.loads(f.read()) + if "weight_map" in index: + index = index["weight_map"] + weights_path = sorted(list(set(index.values()))) + weights_path = [os.path.join(pretrained_model_name_or_path, f) for f in weights_path] + + model = cls._load_model(model, weights_path, is_sharded) + + loading_info = { + "missing_keys": [], + "unexpected_keys": [], + "mismatched_keys": [], + "error_msgs": [], + } + else: + model = cls.from_config(config, **unused_kwargs) + + state_dict = load_state_dict(model_file, variant=variant) + model._convert_deprecated_attention_blocks(state_dict) + + model, missing_keys, unexpected_keys, mismatched_keys, error_msgs = cls._load_pretrained_model( + model, + state_dict, + model_file, + pretrained_model_name_or_path, + ignore_mismatched_sizes=ignore_mismatched_sizes, + ) + + loading_info = { + "missing_keys": missing_keys, + "unexpected_keys": unexpected_keys, + "mismatched_keys": mismatched_keys, + "error_msgs": error_msgs, + } + + if hf_quantizer is not None: + hf_quantizer.postprocess_model(model) + model.hf_quantizer = hf_quantizer + + if torch_dtype is not None and not isinstance(torch_dtype, torch.dtype): + raise ValueError( + f"{torch_dtype} needs to be of type `torch.dtype`, e.g. `torch.float16`, but is {type(torch_dtype)}." + ) + # When using `use_keep_in_fp32_modules` if we do a global `to()` here, then we will + # completely lose the effectivity of `use_keep_in_fp32_modules`. + elif torch_dtype is not None and hf_quantizer is None and not use_keep_in_fp32_modules: + model = model.to(torch_dtype) + + if hf_quantizer is not None: + # We also make sure to purge `_pre_quantization_dtype` when we serialize + # the model config because `_pre_quantization_dtype` is `torch.dtype`, not JSON serializable. + model.register_to_config(_name_or_path=pretrained_model_name_or_path, _pre_quantization_dtype=torch_dtype) + else: + model.register_to_config(_name_or_path=pretrained_model_name_or_path) + + # Set model in evaluation mode to deactivate DropOut modules by default + model.eval() + if output_loading_info: + return model, loading_info + + return model + + @classmethod + def _load_model(cls, model, weights_path, is_sharded): + if not is_sharded: + state_dict = load_state_dict(weights_path) + model.load_weights(state_dict) + else: + need_key = set(model.state_dict().keys()) + state_dict = {} + cache = {} + for weight_file in weights_path: + state_dict = load_state_dict(weight_file) + state_dict.update(cache) + loadkey_cache = model.load_weights(state_dict, is_sharded) + if loadkey_cache : + if isinstance(loadkey_cache, tuple): + loaded_keys, cache = loadkey_cache + else: + loaded_keys = loadkey_cache + need_key = need_key.symmetric_difference(set(loaded_keys)) + + if len(need_key) > 0: + raise ValueError(f"The weight miss key: {need_key}") + return model + + def load_weights(self, state_dict, shard=False): + with torch.no_grad(): + if not shard: + self.load_state_dict(state_dict) + return {} + else: + self.load_state_dict(state_dict, strict=False, assign=True) + return state_dict.keys() + + # Adapted from `transformers`. + @wraps(torch.nn.Module.cuda) + def cuda(self, *args, **kwargs): + # Checks if the model has been loaded in 4-bit or 8-bit with BNB + if getattr(self, "quantization_method", None) == QuantizationMethod.BITS_AND_BYTES: + if getattr(self, "is_loaded_in_8bit", False): + raise ValueError( + "Calling `cuda()` is not supported for `8-bit` quantized models. " + " Please use the model as it is, since the model has already been set to the correct devices." + ) + elif is_bitsandbytes_version("<", "0.43.2"): + raise ValueError( + "Calling `cuda()` is not supported for `4-bit` quantized models with the installed version of bitsandbytes. " + f"The current device is `{self.device}`. If you intended to move the model, please install bitsandbytes >= 0.43.2." + ) + return super().cuda(*args, **kwargs) + + # Adapted from `transformers`. + @wraps(torch.nn.Module.to) + def to(self, *args, **kwargs): + dtype_present_in_args = "dtype" in kwargs + + if not dtype_present_in_args: + for arg in args: + if isinstance(arg, torch.dtype): + dtype_present_in_args = True + break + + if getattr(self, "is_quantized", False): + if dtype_present_in_args: + raise ValueError( + "Casting a quantized model to a new `dtype` is unsupported. To set the dtype of unquantized layers, please " + "use the `torch_dtype` argument when loading the model using `from_pretrained` or `from_single_file`" + ) + + if getattr(self, "quantization_method", None) == QuantizationMethod.BITS_AND_BYTES: + if getattr(self, "is_loaded_in_8bit", False): + raise ValueError( + "`.to` is not supported for `8-bit` bitsandbytes models. Please use the model as it is, since the" + " model has already been set to the correct devices and casted to the correct `dtype`." + ) + elif is_bitsandbytes_version("<", "0.43.2"): + raise ValueError( + "Calling `to()` is not supported for `4-bit` quantized models with the installed version of bitsandbytes. " + f"The current device is `{self.device}`. If you intended to move the model, please install bitsandbytes >= 0.43.2." + ) + return super().to(*args, **kwargs) + + # Taken from `transformers`. + def half(self, *args): + # Checks if the model is quantized + if getattr(self, "is_quantized", False): + raise ValueError( + "`.half()` is not supported for quantized model. Please use the model as it is, since the" + " model has already been cast to the correct `dtype`." + ) + else: + return super().half(*args) + + # Taken from `transformers`. + def float(self, *args): + # Checks if the model is quantized + if getattr(self, "is_quantized", False): + raise ValueError( + "`.float()` is not supported for quantized model. Please use the model as it is, since the" + " model has already been cast to the correct `dtype`." + ) + else: + return super().float(*args) + + @classmethod + def _load_pretrained_model( + cls, + model, + state_dict: OrderedDict, + pretrained_model_name_or_path: Union[str, os.PathLike], + ignore_mismatched_sizes: bool = False, + ): + # Retrieve missing & unexpected_keys + model_state_dict = model.state_dict() + loaded_keys = list(state_dict.keys()) + + expected_keys = list(model_state_dict.keys()) + + original_loaded_keys = loaded_keys + + missing_keys = list(set(expected_keys) - set(loaded_keys)) + unexpected_keys = list(set(loaded_keys) - set(expected_keys)) + + # Make sure we are able to load base models as well as derived models (with heads) + model_to_load = model + + def _find_mismatched_keys( + state_dict, + model_state_dict, + loaded_keys, + ignore_mismatched_sizes, + ): + mismatched_keys = [] + if ignore_mismatched_sizes: + for checkpoint_key in loaded_keys: + model_key = checkpoint_key + + if ( + model_key in model_state_dict + and state_dict[checkpoint_key].shape != model_state_dict[model_key].shape + ): + mismatched_keys.append( + (checkpoint_key, state_dict[checkpoint_key].shape, model_state_dict[model_key].shape) + ) + del state_dict[checkpoint_key] + return mismatched_keys + + if state_dict is not None: + # Whole checkpoint + mismatched_keys = _find_mismatched_keys( + state_dict, + model_state_dict, + original_loaded_keys, + ignore_mismatched_sizes, + ) + error_msgs = _load_state_dict_into_model(model_to_load, state_dict) + + if len(error_msgs) > 0: + error_msg = "\n\t".join(error_msgs) + if "size mismatch" in error_msg: + error_msg += ( + "\n\tYou may consider adding `ignore_mismatched_sizes=True` in the model `from_pretrained` method." + ) + raise RuntimeError(f"Error(s) in loading state_dict for {model.__class__.__name__}:\n\t{error_msg}") + + if len(unexpected_keys) > 0: + logger.warning( + f"Some weights of the model checkpoint at {pretrained_model_name_or_path} were not used when" + f" initializing {model.__class__.__name__}: {unexpected_keys}\n- This IS expected if you are" + f" initializing {model.__class__.__name__} from the checkpoint of a model trained on another task" + " or with another architecture (e.g. initializing a BertForSequenceClassification model from a" + " BertForPreTraining model).\n- This IS NOT expected if you are initializing" + f" {model.__class__.__name__} from the checkpoint of a model that you expect to be exactly" + " identical (initializing a BertForSequenceClassification model from a" + " BertForSequenceClassification model)." + ) + else: + logger.info(f"All model checkpoint weights were used when initializing {model.__class__.__name__}.\n") + if len(missing_keys) > 0: + logger.warning( + f"Some weights of {model.__class__.__name__} were not initialized from the model checkpoint at" + f" {pretrained_model_name_or_path} and are newly initialized: {missing_keys}\nYou should probably" + " TRAIN this model on a down-stream task to be able to use it for predictions and inference." + ) + elif len(mismatched_keys) == 0: + logger.info( + f"All the weights of {model.__class__.__name__} were initialized from the model checkpoint at" + f" {pretrained_model_name_or_path}.\nIf your task is similar to the task the model of the" + f" checkpoint was trained on, you can already use {model.__class__.__name__} for predictions" + " without further training." + ) + if len(mismatched_keys) > 0: + mismatched_warning = "\n".join( + [ + f"- {key}: found shape {shape1} in the checkpoint and {shape2} in the model instantiated" + for key, shape1, shape2 in mismatched_keys + ] + ) + logger.warning( + f"Some weights of {model.__class__.__name__} were not initialized from the model checkpoint at" + f" {pretrained_model_name_or_path} and are newly initialized because the shapes did not" + f" match:\n{mismatched_warning}\nYou should probably TRAIN this model on a down-stream task to be" + " able to use it for predictions and inference." + ) + + return model, missing_keys, unexpected_keys, mismatched_keys, error_msgs + + @property + def device(self) -> torch.device: + return get_parameter_device(self) + + @property + def dtype(self) -> torch.dtype: + return get_parameter_dtype(self) + + def _convert_deprecated_attention_blocks(self, state_dict: OrderedDict) -> None: + deprecated_attention_block_paths = [] + + def recursive_find_attn_block(name, module): + if hasattr(module, "_from_deprecated_attn_block") and module._from_deprecated_attn_block: + deprecated_attention_block_paths.append(name) + + for sub_name, sub_module in module.named_children(): + sub_name = sub_name if name == "" else f"{name}.{sub_name}" + recursive_find_attn_block(sub_name, sub_module) + + recursive_find_attn_block("", self) + + for path in deprecated_attention_block_paths: + # group_norm path stays the same + + # query -> to_q + if f"{path}.query.weight" in state_dict: + state_dict[f"{path}.to_q.weight"] = state_dict.pop(f"{path}.query.weight") + if f"{path}.query.bias" in state_dict: + state_dict[f"{path}.to_q.bias"] = state_dict.pop(f"{path}.query.bias") + + # key -> to_k + if f"{path}.key.weight" in state_dict: + state_dict[f"{path}.to_k.weight"] = state_dict.pop(f"{path}.key.weight") + if f"{path}.key.bias" in state_dict: + state_dict[f"{path}.to_k.bias"] = state_dict.pop(f"{path}.key.bias") + + # value -> to_v + if f"{path}.value.weight" in state_dict: + state_dict[f"{path}.to_v.weight"] = state_dict.pop(f"{path}.value.weight") + if f"{path}.value.bias" in state_dict: + state_dict[f"{path}.to_v.bias"] = state_dict.pop(f"{path}.value.bias") + + # proj_attn -> to_out.0 + if f"{path}.proj_attn.weight" in state_dict: + state_dict[f"{path}.to_out.0.weight"] = state_dict.pop(f"{path}.proj_attn.weight") + if f"{path}.proj_attn.bias" in state_dict: + state_dict[f"{path}.to_out.0.bias"] = state_dict.pop(f"{path}.proj_attn.bias") diff --git a/MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/models/transformer_cogview3plus.py b/MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/models/transformer_cogview3plus.py similarity index 97% rename from MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/models/transformer_cogview3plus.py rename to MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/models/transformer_cogview3plus.py index f704e22589437bb8533e428f8cf1bbab7375f2ae..37c5961586cc2a16aece9c4949dbbcec6ffd2bb7 100644 --- a/MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/models/transformer_cogview3plus.py +++ b/MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/models/transformer_cogview3plus.py @@ -1,397 +1,397 @@ -# Copyright 2024 The CogView team, Tsinghua University & ZhipuAI and The HuggingFace Team. All rights reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# https://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -from typing import Any, Dict, Union - -import torch -import torch.nn as nn -import numpy as np - -from diffusers.configuration_utils import ConfigMixin, register_to_config -from diffusers.models.attention_processor import AttentionProcessor -from diffusers.utils import logging -from diffusers.models.modeling_outputs import Transformer2DModelOutput - -from .modeling_utils import ModelMixin -from .attention import FeedForward -from .attention_processor import CogVideoXAttnProcessor2_0, Attention -from ..layers import CogView3PlusAdaLayerNormZeroTextImage, AdaLayerNormContinuous -from ..layers import CogView3CombinedTimestepSizeEmbeddings, CogView3PlusPatchEmbed - - -logger = logging.get_logger(__name__) # pylint: disable=invalid-name - - -class CogView3PlusTransformerBlock(nn.Module): - def __init__( - self, - dim: int = 2560, - num_attention_heads: int = 64, - attention_head_dim: int = 40, - time_embed_dim: int = 512, - ): - super().__init__() - - self.norm1 = CogView3PlusAdaLayerNormZeroTextImage(embedding_dim=time_embed_dim, dim=dim) - - self.attn1 = Attention( - query_dim=dim, - heads=num_attention_heads, - dim_head=attention_head_dim, - out_dim=dim, - bias=True, - qk_norm="layer_norm", - elementwise_affine=False, - eps=1e-6, - processor=CogVideoXAttnProcessor2_0(), - ) - - self.norm2 = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-5) - self.norm2_context = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-5) - - self.ff = FeedForward(dim=dim, dim_out=dim, activation_fn="gelu-approximate") - - def forward( - self, - hidden_states: torch.Tensor, - encoder_hidden_states: torch.Tensor, - emb: torch.Tensor, - ) -> torch.Tensor: - text_seq_length = encoder_hidden_states.size(1) - - # norm & modulate - norm_hidden_states, chunk_params = self.norm1(hidden_states, encoder_hidden_states, emb) - - gate_msa = chunk_params.gate_msa - shift_mlp = chunk_params.shift_mlp - scale_mlp = chunk_params.scale_mlp - gate_mlp = chunk_params.gate_mlp - norm_encoder_hidden_states = chunk_params.context - c_gate_msa = chunk_params.c_gate_msa - c_shift_mlp = chunk_params.c_shift_mlp - c_scale_mlp = chunk_params.c_scale_mlp - c_gate_mlp = chunk_params.c_gate_mlp - - # attention - attn_hidden_states, attn_encoder_hidden_states = self.attn1( - hidden_states=norm_hidden_states, encoder_hidden_states=norm_encoder_hidden_states - ) - - hidden_states = hidden_states + gate_msa.unsqueeze(1) * attn_hidden_states - encoder_hidden_states = encoder_hidden_states + c_gate_msa.unsqueeze(1) * attn_encoder_hidden_states - - # norm & modulate - norm_hidden_states = self.norm2(hidden_states) - norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None] - - norm_encoder_hidden_states = self.norm2_context(encoder_hidden_states) - norm_encoder_hidden_states = norm_encoder_hidden_states * (1 + c_scale_mlp[:, None]) + c_shift_mlp[:, None] - - # feed-forward - norm_hidden_states = torch.cat([norm_encoder_hidden_states, norm_hidden_states], dim=1) - ff_output = self.ff(norm_hidden_states) - - hidden_states = hidden_states + gate_mlp.unsqueeze(1) * ff_output[:, text_seq_length:] - encoder_hidden_states = encoder_hidden_states + c_gate_mlp.unsqueeze(1) * ff_output[:, :text_seq_length] - - if hidden_states.dtype == torch.float16: - hidden_states = hidden_states.clip(-65504, 65504) - if encoder_hidden_states.dtype == torch.float16: - encoder_hidden_states = encoder_hidden_states.clip(-65504, 65504) - return hidden_states, encoder_hidden_states - - -class CogView3PlusTransformer2DModel(ModelMixin, ConfigMixin): - _supports_gradient_checkpointing = True - - @register_to_config - def __init__( - self, - patch_size: int = 2, - in_channels: int = 16, - num_layers: int = 30, - attention_head_dim: int = 40, - num_attention_heads: int = 64, - out_channels: int = 16, - text_embed_dim: int = 4096, - time_embed_dim: int = 512, - condition_dim: int = 256, - pos_embed_max_size: int = 128, - use_cache: bool = True, - cache_interval: int = 2, - cache_start: int = 3, - num_cache_layer: int = 13, - cache_start_steps: int = 5, - ): - super().__init__() - self.out_channels = out_channels - self.inner_dim = num_attention_heads * attention_head_dim - self.num_layers = num_layers - - # CogView3 uses 3 additional SDXL-like conditions - original_size, target_size, crop_coords - # Each of these are sincos embeddings of shape 2 * condition_dim - self.pooled_projection_dim = 3 * 2 * condition_dim - - self.patch_embed = CogView3PlusPatchEmbed( - in_channels=in_channels, - hidden_size=self.inner_dim, - patch_size=patch_size, - text_hidden_size=text_embed_dim, - pos_embed_max_size=pos_embed_max_size, - ) - - self.time_condition_embed = CogView3CombinedTimestepSizeEmbeddings( - embedding_dim=time_embed_dim, - condition_dim=condition_dim, - pooled_projection_dim=self.pooled_projection_dim, - timesteps_dim=self.inner_dim, - ) - - self.transformer_blocks = nn.ModuleList( - [ - CogView3PlusTransformerBlock( - dim=self.inner_dim, - num_attention_heads=num_attention_heads, - attention_head_dim=attention_head_dim, - time_embed_dim=time_embed_dim, - ) - for _ in range(num_layers) - ] - ) - - self.norm_out = AdaLayerNormContinuous( - embedding_dim=self.inner_dim, - conditioning_embedding_dim=time_embed_dim, - elementwise_affine=False, - eps=1e-6, - ) - self.proj_out = nn.Linear(self.inner_dim, patch_size * patch_size * self.out_channels, bias=True) - - self.gradient_checkpointing = False - - self.q_weight_cache = None - self.q_bias_cache = None - self.k_weight_cache = None - self.k_bias_cache = None - self.v_weight_cache = None - self.v_bias_cache = None - - self.use_cache = use_cache - self.cache_interval = cache_interval - self.cache_start = cache_start - self.num_cache_layer = num_cache_layer - self.cache_start_steps = cache_start_steps - - self.delta_cache = None - self.delta_encoder_cache = None - - @property - def attn_processors(self) -> Dict[str, AttentionProcessor]: - r""" - Returns: - `dict` of attention processors: A dictionary containing all attention processors used in the model with - indexed by its weight name. - """ - # set recursively - processors = {} - - def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]): - if hasattr(module, "get_processor"): - processors[f"{name}.processor"] = module.get_processor() - - for sub_name, child in module.named_children(): - fn_recursive_add_processors(f"{name}.{sub_name}", child, processors) - - return processors - - for name, module in self.named_children(): - fn_recursive_add_processors(name, module, processors) - - return processors - - def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]): - count = len(self.attn_processors.keys()) - - if isinstance(processor, dict) and len(processor) != count: - raise ValueError( - f"A dict of processors was passed, but the number of processors {len(processor)} does not match the" - f" number of attention layers: {count}. Please make sure to pass {count} processor classes." - ) - - def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor): - if hasattr(module, "set_processor"): - if not isinstance(processor, dict): - module.set_processor(processor) - else: - module.set_processor(processor.pop(f"{name}.processor")) - - for sub_name, child in module.named_children(): - fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor) - - for name, module in self.named_children(): - fn_recursive_attn_processor(name, module, processor) - - def _set_gradient_checkpointing(self, module, value=False): - if hasattr(module, "gradient_checkpointing"): - module.gradient_checkpointing = value - - def forward( - self, - states, - timestep: torch.LongTensor, - original_size: torch.Tensor, - target_size: torch.Tensor, - crop_coords: torch.Tensor, - ) -> Union[torch.Tensor, Transformer2DModelOutput]: - hidden_states = states[0] - encoder_hidden_states = states[1] - height, width = hidden_states.shape[-2:] - text_seq_length = encoder_hidden_states.shape[1] - - hidden_states = self.patch_embed( - hidden_states, encoder_hidden_states - ) # takes care of adding positional embeddings too. - emb = self.time_condition_embed(timestep, original_size, target_size, crop_coords, hidden_states.dtype) - - encoder_hidden_states = hidden_states[:, :text_seq_length] - hidden_states = hidden_states[:, text_seq_length:] - - hidden_states, encoder_hidden_states = self._forward_blocks(hidden_states, encoder_hidden_states, emb, states[2]) - - hidden_states = self.norm_out(hidden_states, emb) - hidden_states = self.proj_out(hidden_states) # (batch_size, height*width, patch_size*patch_size*out_channels) - - # unpatchify - patch_size = self.config.patch_size - height = height // patch_size - width = width // patch_size - - hidden_states = hidden_states.reshape( - shape=(hidden_states.shape[0], height, width, self.out_channels, patch_size, patch_size) - ) - hidden_states = torch.einsum("nhwcpq->nchpwq", hidden_states) - output = hidden_states.reshape( - shape=(hidden_states.shape[0], self.out_channels, height * patch_size, width * patch_size) - ) - - return Transformer2DModelOutput(sample=output) - - # forward blocks in range [start_idx, end_idx), then return input and output - def _forward_blocks_range(self, hidden_states, encoder_hidden_states, emb, start_idx, end_idx, **kwargs): - for _, block in enumerate(self.transformer_blocks[start_idx: end_idx]): - hidden_states, encoder_hidden_states = block( - hidden_states=hidden_states, - encoder_hidden_states=encoder_hidden_states, - emb=emb, - ) - - return hidden_states, encoder_hidden_states - - def _forward_blocks(self, hidden_states, encoder_hidden_states, emb, t_idx): - num_blocks = len(self.transformer_blocks) - - if not self.use_cache or (t_idx < self.cache_start_steps): - hidden_states, encoder_hidden_states = self._forward_blocks_range( - hidden_states, - encoder_hidden_states, - emb, - 0, - num_blocks - ) - else: - # infer [0, cache_start) - hidden_states, encoder_hidden_states = self._forward_blocks_range( - hidden_states, - encoder_hidden_states, - emb, - 0, - self.cache_start - ) - # infer [cache_start, cache_end) - cache_end = np.minimum(self.cache_start + self.num_cache_layer, num_blocks) - hidden_states_before_cache = hidden_states.clone() - encoder_hidden_states_before_cache = encoder_hidden_states.clone() - if t_idx % self.cache_interval == (self.cache_start_steps % self.cache_interval): - hidden_states, encoder_hidden_states = self._forward_blocks_range( - hidden_states, - encoder_hidden_states, - emb, - self.cache_start, - cache_end - ) - self.delta_cache = hidden_states - hidden_states_before_cache - self.delta_encoder_cache = encoder_hidden_states - encoder_hidden_states_before_cache - else: - hidden_states = hidden_states_before_cache + self.delta_cache - encoder_hidden_states = encoder_hidden_states_before_cache + self.delta_encoder_cache - # infer [cache_end, num_blocks) - hidden_states, encoder_hidden_states = self._forward_blocks_range( - hidden_states, - encoder_hidden_states, - emb, - cache_end, - num_blocks - ) - - return hidden_states, encoder_hidden_states - - def load_weights(self, state_dict, shard=False): - with torch.no_grad(): - if not shard: - self.load_state_dict(state_dict) - return {} - else: - weights = state_dict - - for i in range(self.num_layers): - if i != 26: - q_weight = weights.pop(f"transformer_blocks.{i}.attn1.to_q.weight", None) - q_bias = weights.pop(f"transformer_blocks.{i}.attn1.to_q.bias", None) - k_weight = weights.pop(f"transformer_blocks.{i}.attn1.to_k.weight", None) - k_bias = weights.pop(f"transformer_blocks.{i}.attn1.to_k.bias", None) - v_weight = weights.pop(f"transformer_blocks.{i}.attn1.to_v.weight", None) - v_bias = weights.pop(f"transformer_blocks.{i}.attn1.to_v.bias", None) - - # query, key, value的weight和bias权重存在同一个文件中，不会分开存储。 - if q_weight is not None and k_weight is not None and v_weight is not None: - qkv_weight = torch.cat([q_weight, k_weight, v_weight], dim=0).transpose(0, 1).contiguous() - qkv_bias = torch.cat([q_bias, k_bias, v_bias], dim=0).contiguous() - weights[f"transformer_blocks.{i}.attn1.to_qkv.weight"] = qkv_weight - weights[f"transformer_blocks.{i}.attn1.to_qkv.bias"] = qkv_bias - else: - if self.q_weight_cache is None: - self.q_weight_cache = weights.pop(f"transformer_blocks.{i}.attn1.to_q.weight", None) - if self.q_bias_cache is None: - self.q_bias_cache = weights.pop(f"transformer_blocks.{i}.attn1.to_q.bias", None) - if self.k_weight_cache is None: - self.k_weight_cache = weights.pop(f"transformer_blocks.{i}.attn1.to_k.weight", None) - if self.k_bias_cache is None: - self.k_bias_cache = weights.pop(f"transformer_blocks.{i}.attn1.to_k.bias", None) - if self.v_weight_cache is None: - self.v_weight_cache = weights.pop(f"transformer_blocks.{i}.attn1.to_v.weight", None) - if self.v_bias_cache is None: - self.v_bias_cache = weights.pop(f"transformer_blocks.{i}.attn1.to_v.bias", None) - - qk_weight_cache = self.q_weight_cache is not None and self.k_weight_cache is not None - if qk_weight_cache and self.v_weight_cache is not None: - qkv_weight = torch.cat( - [self.q_weight_cache, self.k_weight_cache, self.v_weight_cache], - dim=0 - ).transpose(0, 1).contiguous() - qkv_bias = torch.cat([self.q_bias_cache, self.k_bias_cache, self.v_bias_cache], dim=0).contiguous() - weights[f"transformer_blocks.26.attn1.to_qkv.weight"] = qkv_weight - weights[f"transformer_blocks.26.attn1.to_qkv.bias"] = qkv_bias - - self.load_state_dict(weights, strict=False, assign=True) - return weights.keys() +# Copyright 2024 The CogView team, Tsinghua University & ZhipuAI and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# https://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from typing import Any, Dict, Union + +import torch +import torch.nn as nn +import numpy as np + +from diffusers.configuration_utils import ConfigMixin, register_to_config +from diffusers.models.attention_processor import AttentionProcessor +from diffusers.utils import logging +from diffusers.models.modeling_outputs import Transformer2DModelOutput + +from .modeling_utils import ModelMixin +from .attention import FeedForward +from .attention_processor import CogVideoXAttnProcessor2_0, Attention +from ..layers import CogView3PlusAdaLayerNormZeroTextImage, AdaLayerNormContinuous +from ..layers import CogView3CombinedTimestepSizeEmbeddings, CogView3PlusPatchEmbed + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +class CogView3PlusTransformerBlock(nn.Module): + def __init__( + self, + dim: int = 2560, + num_attention_heads: int = 64, + attention_head_dim: int = 40, + time_embed_dim: int = 512, + ): + super().__init__() + + self.norm1 = CogView3PlusAdaLayerNormZeroTextImage(embedding_dim=time_embed_dim, dim=dim) + + self.attn1 = Attention( + query_dim=dim, + heads=num_attention_heads, + dim_head=attention_head_dim, + out_dim=dim, + bias=True, + qk_norm="layer_norm", + elementwise_affine=False, + eps=1e-6, + processor=CogVideoXAttnProcessor2_0(), + ) + + self.norm2 = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-5) + self.norm2_context = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-5) + + self.ff = FeedForward(dim=dim, dim_out=dim, activation_fn="gelu-approximate") + + def forward( + self, + hidden_states: torch.Tensor, + encoder_hidden_states: torch.Tensor, + emb: torch.Tensor, + ) -> torch.Tensor: + text_seq_length = encoder_hidden_states.size(1) + + # norm & modulate + norm_hidden_states, chunk_params = self.norm1(hidden_states, encoder_hidden_states, emb) + + gate_msa = chunk_params.gate_msa + shift_mlp = chunk_params.shift_mlp + scale_mlp = chunk_params.scale_mlp + gate_mlp = chunk_params.gate_mlp + norm_encoder_hidden_states = chunk_params.context + c_gate_msa = chunk_params.c_gate_msa + c_shift_mlp = chunk_params.c_shift_mlp + c_scale_mlp = chunk_params.c_scale_mlp + c_gate_mlp = chunk_params.c_gate_mlp + + # attention + attn_hidden_states, attn_encoder_hidden_states = self.attn1( + hidden_states=norm_hidden_states, encoder_hidden_states=norm_encoder_hidden_states + ) + + hidden_states = hidden_states + gate_msa.unsqueeze(1) * attn_hidden_states + encoder_hidden_states = encoder_hidden_states + c_gate_msa.unsqueeze(1) * attn_encoder_hidden_states + + # norm & modulate + norm_hidden_states = self.norm2(hidden_states) + norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None] + + norm_encoder_hidden_states = self.norm2_context(encoder_hidden_states) + norm_encoder_hidden_states = norm_encoder_hidden_states * (1 + c_scale_mlp[:, None]) + c_shift_mlp[:, None] + + # feed-forward + norm_hidden_states = torch.cat([norm_encoder_hidden_states, norm_hidden_states], dim=1) + ff_output = self.ff(norm_hidden_states) + + hidden_states = hidden_states + gate_mlp.unsqueeze(1) * ff_output[:, text_seq_length:] + encoder_hidden_states = encoder_hidden_states + c_gate_mlp.unsqueeze(1) * ff_output[:, :text_seq_length] + + if hidden_states.dtype == torch.float16: + hidden_states = hidden_states.clip(-65504, 65504) + if encoder_hidden_states.dtype == torch.float16: + encoder_hidden_states = encoder_hidden_states.clip(-65504, 65504) + return hidden_states, encoder_hidden_states + + +class CogView3PlusTransformer2DModel(ModelMixin, ConfigMixin): + _supports_gradient_checkpointing = True + + @register_to_config + def __init__( + self, + patch_size: int = 2, + in_channels: int = 16, + num_layers: int = 30, + attention_head_dim: int = 40, + num_attention_heads: int = 64, + out_channels: int = 16, + text_embed_dim: int = 4096, + time_embed_dim: int = 512, + condition_dim: int = 256, + pos_embed_max_size: int = 128, + use_cache: bool = True, + cache_interval: int = 2, + cache_start: int = 3, + num_cache_layer: int = 13, + cache_start_steps: int = 5, + ): + super().__init__() + self.out_channels = out_channels + self.inner_dim = num_attention_heads * attention_head_dim + self.num_layers = num_layers + + # CogView3 uses 3 additional SDXL-like conditions - original_size, target_size, crop_coords + # Each of these are sincos embeddings of shape 2 * condition_dim + self.pooled_projection_dim = 3 * 2 * condition_dim + + self.patch_embed = CogView3PlusPatchEmbed( + in_channels=in_channels, + hidden_size=self.inner_dim, + patch_size=patch_size, + text_hidden_size=text_embed_dim, + pos_embed_max_size=pos_embed_max_size, + ) + + self.time_condition_embed = CogView3CombinedTimestepSizeEmbeddings( + embedding_dim=time_embed_dim, + condition_dim=condition_dim, + pooled_projection_dim=self.pooled_projection_dim, + timesteps_dim=self.inner_dim, + ) + + self.transformer_blocks = nn.ModuleList( + [ + CogView3PlusTransformerBlock( + dim=self.inner_dim, + num_attention_heads=num_attention_heads, + attention_head_dim=attention_head_dim, + time_embed_dim=time_embed_dim, + ) + for _ in range(num_layers) + ] + ) + + self.norm_out = AdaLayerNormContinuous( + embedding_dim=self.inner_dim, + conditioning_embedding_dim=time_embed_dim, + elementwise_affine=False, + eps=1e-6, + ) + self.proj_out = nn.Linear(self.inner_dim, patch_size * patch_size * self.out_channels, bias=True) + + self.gradient_checkpointing = False + + self.q_weight_cache = None + self.q_bias_cache = None + self.k_weight_cache = None + self.k_bias_cache = None + self.v_weight_cache = None + self.v_bias_cache = None + + self.use_cache = use_cache + self.cache_interval = cache_interval + self.cache_start = cache_start + self.num_cache_layer = num_cache_layer + self.cache_start_steps = cache_start_steps + + self.delta_cache = None + self.delta_encoder_cache = None + + @property + def attn_processors(self) -> Dict[str, AttentionProcessor]: + r""" + Returns: + `dict` of attention processors: A dictionary containing all attention processors used in the model with + indexed by its weight name. + """ + # set recursively + processors = {} + + def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]): + if hasattr(module, "get_processor"): + processors[f"{name}.processor"] = module.get_processor() + + for sub_name, child in module.named_children(): + fn_recursive_add_processors(f"{name}.{sub_name}", child, processors) + + return processors + + for name, module in self.named_children(): + fn_recursive_add_processors(name, module, processors) + + return processors + + def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]): + count = len(self.attn_processors.keys()) + + if isinstance(processor, dict) and len(processor) != count: + raise ValueError( + f"A dict of processors was passed, but the number of processors {len(processor)} does not match the" + f" number of attention layers: {count}. Please make sure to pass {count} processor classes." + ) + + def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor): + if hasattr(module, "set_processor"): + if not isinstance(processor, dict): + module.set_processor(processor) + else: + module.set_processor(processor.pop(f"{name}.processor")) + + for sub_name, child in module.named_children(): + fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor) + + for name, module in self.named_children(): + fn_recursive_attn_processor(name, module, processor) + + def _set_gradient_checkpointing(self, module, value=False): + if hasattr(module, "gradient_checkpointing"): + module.gradient_checkpointing = value + + def forward( + self, + states, + timestep: torch.LongTensor, + original_size: torch.Tensor, + target_size: torch.Tensor, + crop_coords: torch.Tensor, + ) -> Union[torch.Tensor, Transformer2DModelOutput]: + hidden_states = states[0] + encoder_hidden_states = states[1] + height, width = hidden_states.shape[-2:] + text_seq_length = encoder_hidden_states.shape[1] + + hidden_states = self.patch_embed( + hidden_states, encoder_hidden_states + ) # takes care of adding positional embeddings too. + emb = self.time_condition_embed(timestep, original_size, target_size, crop_coords, hidden_states.dtype) + + encoder_hidden_states = hidden_states[:, :text_seq_length] + hidden_states = hidden_states[:, text_seq_length:] + + hidden_states, encoder_hidden_states = self._forward_blocks(hidden_states, encoder_hidden_states, emb, states[2]) + + hidden_states = self.norm_out(hidden_states, emb) + hidden_states = self.proj_out(hidden_states) # (batch_size, height*width, patch_size*patch_size*out_channels) + + # unpatchify + patch_size = self.config.patch_size + height = height // patch_size + width = width // patch_size + + hidden_states = hidden_states.reshape( + shape=(hidden_states.shape[0], height, width, self.out_channels, patch_size, patch_size) + ) + hidden_states = torch.einsum("nhwcpq->nchpwq", hidden_states) + output = hidden_states.reshape( + shape=(hidden_states.shape[0], self.out_channels, height * patch_size, width * patch_size) + ) + + return Transformer2DModelOutput(sample=output) + + # forward blocks in range [start_idx, end_idx), then return input and output + def _forward_blocks_range(self, hidden_states, encoder_hidden_states, emb, start_idx, end_idx, **kwargs): + for _, block in enumerate(self.transformer_blocks[start_idx: end_idx]): + hidden_states, encoder_hidden_states = block( + hidden_states=hidden_states, + encoder_hidden_states=encoder_hidden_states, + emb=emb, + ) + + return hidden_states, encoder_hidden_states + + def _forward_blocks(self, hidden_states, encoder_hidden_states, emb, t_idx): + num_blocks = len(self.transformer_blocks) + + if not self.use_cache or (t_idx < self.cache_start_steps): + hidden_states, encoder_hidden_states = self._forward_blocks_range( + hidden_states, + encoder_hidden_states, + emb, + 0, + num_blocks + ) + else: + # infer [0, cache_start) + hidden_states, encoder_hidden_states = self._forward_blocks_range( + hidden_states, + encoder_hidden_states, + emb, + 0, + self.cache_start + ) + # infer [cache_start, cache_end) + cache_end = np.minimum(self.cache_start + self.num_cache_layer, num_blocks) + hidden_states_before_cache = hidden_states.clone() + encoder_hidden_states_before_cache = encoder_hidden_states.clone() + if t_idx % self.cache_interval == (self.cache_start_steps % self.cache_interval): + hidden_states, encoder_hidden_states = self._forward_blocks_range( + hidden_states, + encoder_hidden_states, + emb, + self.cache_start, + cache_end + ) + self.delta_cache = hidden_states - hidden_states_before_cache + self.delta_encoder_cache = encoder_hidden_states - encoder_hidden_states_before_cache + else: + hidden_states = hidden_states_before_cache + self.delta_cache + encoder_hidden_states = encoder_hidden_states_before_cache + self.delta_encoder_cache + # infer [cache_end, num_blocks) + hidden_states, encoder_hidden_states = self._forward_blocks_range( + hidden_states, + encoder_hidden_states, + emb, + cache_end, + num_blocks + ) + + return hidden_states, encoder_hidden_states + + def load_weights(self, state_dict, shard=False): + with torch.no_grad(): + if not shard: + self.load_state_dict(state_dict) + return {} + else: + weights = state_dict + + for i in range(self.num_layers): + if i != 26: + q_weight = weights.pop(f"transformer_blocks.{i}.attn1.to_q.weight", None) + q_bias = weights.pop(f"transformer_blocks.{i}.attn1.to_q.bias", None) + k_weight = weights.pop(f"transformer_blocks.{i}.attn1.to_k.weight", None) + k_bias = weights.pop(f"transformer_blocks.{i}.attn1.to_k.bias", None) + v_weight = weights.pop(f"transformer_blocks.{i}.attn1.to_v.weight", None) + v_bias = weights.pop(f"transformer_blocks.{i}.attn1.to_v.bias", None) + + # query, key, value的weight和bias权重存在同一个文件中，不会分开存储。 + if q_weight is not None and k_weight is not None and v_weight is not None: + qkv_weight = torch.cat([q_weight, k_weight, v_weight], dim=0).transpose(0, 1).contiguous() + qkv_bias = torch.cat([q_bias, k_bias, v_bias], dim=0).contiguous() + weights[f"transformer_blocks.{i}.attn1.to_qkv.weight"] = qkv_weight + weights[f"transformer_blocks.{i}.attn1.to_qkv.bias"] = qkv_bias + else: + if self.q_weight_cache is None: + self.q_weight_cache = weights.pop(f"transformer_blocks.{i}.attn1.to_q.weight", None) + if self.q_bias_cache is None: + self.q_bias_cache = weights.pop(f"transformer_blocks.{i}.attn1.to_q.bias", None) + if self.k_weight_cache is None: + self.k_weight_cache = weights.pop(f"transformer_blocks.{i}.attn1.to_k.weight", None) + if self.k_bias_cache is None: + self.k_bias_cache = weights.pop(f"transformer_blocks.{i}.attn1.to_k.bias", None) + if self.v_weight_cache is None: + self.v_weight_cache = weights.pop(f"transformer_blocks.{i}.attn1.to_v.weight", None) + if self.v_bias_cache is None: + self.v_bias_cache = weights.pop(f"transformer_blocks.{i}.attn1.to_v.bias", None) + + qk_weight_cache = self.q_weight_cache is not None and self.k_weight_cache is not None + if qk_weight_cache and self.v_weight_cache is not None: + qkv_weight = torch.cat( + [self.q_weight_cache, self.k_weight_cache, self.v_weight_cache], + dim=0 + ).transpose(0, 1).contiguous() + qkv_bias = torch.cat([self.q_bias_cache, self.k_bias_cache, self.v_bias_cache], dim=0).contiguous() + weights[f"transformer_blocks.26.attn1.to_qkv.weight"] = qkv_weight + weights[f"transformer_blocks.26.attn1.to_qkv.bias"] = qkv_bias + + self.load_state_dict(weights, strict=False, assign=True) + return weights.keys() diff --git a/MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/pipeline/__init__.py b/MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/pipeline/__init__.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/pipeline/__init__.py rename to MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/pipeline/__init__.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/pipeline/pipeline_cogview3plus.py b/MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/pipeline/pipeline_cogview3plus.py similarity index 97% rename from MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/pipeline/pipeline_cogview3plus.py rename to MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/pipeline/pipeline_cogview3plus.py index fe2bd5cfcd33a2a7e99cdf4f79b1130dc86e5cf5..4b07df76a687d5b1e50abd4779dd07a3d73011ab 100644 --- a/MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/pipeline/pipeline_cogview3plus.py +++ b/MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/pipeline/pipeline_cogview3plus.py @@ -1,339 +1,339 @@ -# Copyright 2024 The CogVideoX team, Tsinghua University & ZhipuAI and The HuggingFace Team. -# All rights reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# https://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -import inspect -from typing import List, Optional, Tuple, Union - -import torch -from transformers import T5EncoderModel, T5Tokenizer - -from diffusers.image_processor import VaeImageProcessor -from diffusers.pipelines.pipeline_utils import DiffusionPipeline -from diffusers.utils import logging -from diffusers.utils.torch_utils import randn_tensor -from diffusers import AutoencoderKL - -from ..models import CogView3PlusTransformer2DModel -from ..schedulers import CogVideoXDDIMScheduler -from .pipeline_output import CogView3PipelineOutput - -logger = logging.get_logger(__name__) # pylint: disable=invalid-name - - -def retrieve_timesteps( - scheduler, - num_inference_steps: Optional[int] = None, - device: Optional[Union[str, torch.device]] = None, - timesteps: Optional[List[int]] = None, - sigmas: Optional[List[float]] = None, - **kwargs, -): - if timesteps is not None and sigmas is not None: - raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") - if timesteps is not None: - accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) - if not accepts_timesteps: - raise ValueError( - f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" - f" timestep schedules. Please check whether you are using the correct scheduler." - ) - scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) - timesteps = scheduler.timesteps - num_inference_steps = len(timesteps) - elif sigmas is not None: - accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) - if not accept_sigmas: - raise ValueError( - f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" - f" sigmas schedules. Please check whether you are using the correct scheduler." - ) - scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) - timesteps = scheduler.timesteps - num_inference_steps = len(timesteps) - else: - scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) - timesteps = scheduler.timesteps - return timesteps, num_inference_steps - - -class CogView3PlusPipeline(DiffusionPipeline): - _optional_components = [] - model_cpu_offload_seq = "text_encoder->transformer->vae" - - _callback_tensor_inputs = [ - "latents", - "prompt_embeds", - "negative_prompt_embeds", - ] - - def __init__( - self, - tokenizer: T5Tokenizer, - text_encoder: T5EncoderModel, - vae: AutoencoderKL, - transformer: CogView3PlusTransformer2DModel, - scheduler: CogVideoXDDIMScheduler, - ): - super().__init__() - - self.register_modules( - tokenizer=tokenizer, text_encoder=text_encoder, vae=vae, transformer=transformer, scheduler=scheduler - ) - self.vae_scale_factor = ( - 2 ** (len(self.vae.config.block_out_channels) - 1) if hasattr(self, "vae") and self.vae is not None else 8 - ) - - self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) - - def _get_t5_prompt_embeds( - self, - prompt: Union[str, List[str]] = None, - num_images_per_prompt: int = 1, - max_sequence_length: int = 226, - device: Optional[torch.device] = None, - dtype: Optional[torch.dtype] = None, - ): - device = device or self._execution_device - dtype = dtype or self.text_encoder.dtype - - prompt = [prompt] if isinstance(prompt, str) else prompt - batch_size = len(prompt) - - text_inputs = self.tokenizer( - prompt, - padding="max_length", - max_length=max_sequence_length, - truncation=True, - add_special_tokens=True, - return_tensors="pt", - ) - text_input_ids = text_inputs.input_ids - untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids - - if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): - removed_text = self.tokenizer.batch_decode(untruncated_ids[:, max_sequence_length - 1 : -1]) - logger.warning( - "The following part of your input was truncated because `max_sequence_length` is set to " - f" {max_sequence_length} tokens: {removed_text}" - ) - - prompt_embeds = self.text_encoder(text_input_ids.to(device))[0] - prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) - - # duplicate text embeddings for each generation per prompt, using mps friendly method - _, seq_len, _ = prompt_embeds.shape - prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) - prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) - - return prompt_embeds - - def encode_prompt( - self, - prompt: Union[str, List[str]], - num_images_per_prompt: int = 1, - max_sequence_length: int = 224, - device: Optional[torch.device] = None, - dtype: Optional[torch.dtype] = None, - ): - device = device or self._execution_device - - prompt = [prompt] if isinstance(prompt, str) else prompt - prompt_embeds = self._get_t5_prompt_embeds( - prompt=prompt, - num_images_per_prompt=num_images_per_prompt, - max_sequence_length=max_sequence_length, - device=device, - dtype=dtype, - ) - negative_prompt_embeds = prompt_embeds.new_zeros(prompt_embeds.shape) - - return prompt_embeds, negative_prompt_embeds - - def prepare_latents(self, batch_size, num_channels_latents, image_size, dtype, device): - height = image_size[0] - width = image_size[1] - shape = ( - batch_size, - num_channels_latents, - int(height) // self.vae_scale_factor, - int(width) // self.vae_scale_factor, - ) - latents = randn_tensor(shape, device=device, dtype=dtype) - - # scale the initial noise by the standard deviation required by the scheduler - latents = latents * self.scheduler.init_noise_sigma - return latents - - def prepare_extra_step_kwargs(self, generator, eta): - accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) - extra_step_kwargs = {} - if accepts_eta: - extra_step_kwargs["eta"] = eta - - # check if the scheduler accepts generator - accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) - if accepts_generator: - extra_step_kwargs["generator"] = generator - return extra_step_kwargs - - def check_inputs( - self, - prompt, - height, - width, - ): - if height % 8 != 0 or width % 8 != 0: - raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") - - if prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): - raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") - - @property - def guidance_scale(self): - return self._guidance_scale - - @property - def do_classifier_free_guidance(self): - return self._guidance_scale > 1 - - @property - def num_timesteps(self): - return self._num_timesteps - - @property - def interrupt(self): - return self._interrupt - - @torch.no_grad() - def __call__( - self, - prompt: Optional[Union[str, List[str]]] = None, - image_size: Tuple[int, int] = None, - num_inference_steps: int = 50, - guidance_scale: float = 5.0, - num_images_per_prompt: int = 1, - ) -> Union[CogView3PipelineOutput, Tuple]: - if image_size is None: - height = self.transformer.config.sample_size * self.vae_scale_factor - width = self.transformer.config.sample_size * self.vae_scale_factor - else: - height = image_size[0] - width = image_size[1] - - original_size = (height, width) - target_size = (height, width) - - # 1. Check inputs. Raise error if not correct - self.check_inputs( - prompt, - height, - width, - ) - self._guidance_scale = guidance_scale - self._interrupt = False - - # 2. Default call parameters - if isinstance(prompt, str): - batch_size = 1 - else: - batch_size = len(prompt) - - device = self._execution_device - - # 3. Encode input prompt - prompt_embeds, negative_prompt_embeds = self.encode_prompt( - prompt, - num_images_per_prompt=num_images_per_prompt, - max_sequence_length=224, - device=device, - ) - if self.do_classifier_free_guidance: - prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) - - # 4. Prepare timesteps - timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device) - self._num_timesteps = len(timesteps) - - # 5. Prepare latents. - latent_channels = self.transformer.config.in_channels - latents = self.prepare_latents( - batch_size * num_images_per_prompt, - latent_channels, - (height, width), - prompt_embeds.dtype, - device, - ) - - extra_step_kwargs = self.prepare_extra_step_kwargs(None, 0.0) - - # 7. Prepare additional timestep conditions - original_size = torch.tensor([original_size], dtype=prompt_embeds.dtype) - target_size = torch.tensor([target_size], dtype=prompt_embeds.dtype) - crops_coords_top_left = torch.tensor([(0, 0)], dtype=prompt_embeds.dtype) - - if self.do_classifier_free_guidance: - original_size = torch.cat([original_size, original_size]) - target_size = torch.cat([target_size, target_size]) - crops_coords_top_left = torch.cat([crops_coords_top_left, crops_coords_top_left]) - - original_size = original_size.to(device).repeat(batch_size * num_images_per_prompt, 1) - target_size = target_size.to(device).repeat(batch_size * num_images_per_prompt, 1) - crops_coords_top_left = crops_coords_top_left.to(device).repeat(batch_size * num_images_per_prompt, 1) - - # 8. Denoising loop - num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) - - with self.progress_bar(total=num_inference_steps) as progress_bar: - # for DPM-solver++ - old_pred_original_sample = None - for i, t in enumerate(timesteps): - if self.interrupt: - continue - - latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents - latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) - - # broadcast to batch dimension in a way that's compatible with ONNX/Core ML - timestep = t.expand(latent_model_input.shape[0]) - - # predict noise model_output - noise_pred = self.transformer( - states=(latent_model_input, prompt_embeds, i), - timestep=timestep, - original_size=original_size, - target_size=target_size, - crop_coords=crops_coords_top_left, - )[0] - noise_pred = noise_pred.float() - - # perform guidance - if self.do_classifier_free_guidance: - noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) - noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) - - # compute the previous noisy sample x_t -> x_t-1 - latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] - latents = latents.to(prompt_embeds.dtype) - - if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): - progress_bar.update() - - image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, generator=None)[0] - image = self.image_processor.postprocess(image, output_type="pil") - - # Offload all models - self.maybe_free_model_hooks() - +# Copyright 2024 The CogVideoX team, Tsinghua University & ZhipuAI and The HuggingFace Team. +# All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# https://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import List, Optional, Tuple, Union + +import torch +from transformers import T5EncoderModel, T5Tokenizer + +from diffusers.image_processor import VaeImageProcessor +from diffusers.pipelines.pipeline_utils import DiffusionPipeline +from diffusers.utils import logging +from diffusers.utils.torch_utils import randn_tensor +from diffusers import AutoencoderKL + +from ..models import CogView3PlusTransformer2DModel +from ..schedulers import CogVideoXDDIMScheduler +from .pipeline_output import CogView3PipelineOutput + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +def retrieve_timesteps( + scheduler, + num_inference_steps: Optional[int] = None, + device: Optional[Union[str, torch.device]] = None, + timesteps: Optional[List[int]] = None, + sigmas: Optional[List[float]] = None, + **kwargs, +): + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class CogView3PlusPipeline(DiffusionPipeline): + _optional_components = [] + model_cpu_offload_seq = "text_encoder->transformer->vae" + + _callback_tensor_inputs = [ + "latents", + "prompt_embeds", + "negative_prompt_embeds", + ] + + def __init__( + self, + tokenizer: T5Tokenizer, + text_encoder: T5EncoderModel, + vae: AutoencoderKL, + transformer: CogView3PlusTransformer2DModel, + scheduler: CogVideoXDDIMScheduler, + ): + super().__init__() + + self.register_modules( + tokenizer=tokenizer, text_encoder=text_encoder, vae=vae, transformer=transformer, scheduler=scheduler + ) + self.vae_scale_factor = ( + 2 ** (len(self.vae.config.block_out_channels) - 1) if hasattr(self, "vae") and self.vae is not None else 8 + ) + + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + + def _get_t5_prompt_embeds( + self, + prompt: Union[str, List[str]] = None, + num_images_per_prompt: int = 1, + max_sequence_length: int = 226, + device: Optional[torch.device] = None, + dtype: Optional[torch.dtype] = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + add_special_tokens=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode(untruncated_ids[:, max_sequence_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because `max_sequence_length` is set to " + f" {max_sequence_length} tokens: {removed_text}" + ) + + prompt_embeds = self.text_encoder(text_input_ids.to(device))[0] + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + # duplicate text embeddings for each generation per prompt, using mps friendly method + _, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + return prompt_embeds + + def encode_prompt( + self, + prompt: Union[str, List[str]], + num_images_per_prompt: int = 1, + max_sequence_length: int = 224, + device: Optional[torch.device] = None, + dtype: Optional[torch.dtype] = None, + ): + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + prompt_embeds = self._get_t5_prompt_embeds( + prompt=prompt, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + negative_prompt_embeds = prompt_embeds.new_zeros(prompt_embeds.shape) + + return prompt_embeds, negative_prompt_embeds + + def prepare_latents(self, batch_size, num_channels_latents, image_size, dtype, device): + height = image_size[0] + width = image_size[1] + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + latents = randn_tensor(shape, device=device, dtype=dtype) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + def prepare_extra_step_kwargs(self, generator, eta): + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + height, + width, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + def __call__( + self, + prompt: Optional[Union[str, List[str]]] = None, + image_size: Tuple[int, int] = None, + num_inference_steps: int = 50, + guidance_scale: float = 5.0, + num_images_per_prompt: int = 1, + ) -> Union[CogView3PipelineOutput, Tuple]: + if image_size is None: + height = self.transformer.config.sample_size * self.vae_scale_factor + width = self.transformer.config.sample_size * self.vae_scale_factor + else: + height = image_size[0] + width = image_size[1] + + original_size = (height, width) + target_size = (height, width) + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + height, + width, + ) + self._guidance_scale = guidance_scale + self._interrupt = False + + # 2. Default call parameters + if isinstance(prompt, str): + batch_size = 1 + else: + batch_size = len(prompt) + + device = self._execution_device + + # 3. Encode input prompt + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=224, + device=device, + ) + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + + # 4. Prepare timesteps + timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device) + self._num_timesteps = len(timesteps) + + # 5. Prepare latents. + latent_channels = self.transformer.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + latent_channels, + (height, width), + prompt_embeds.dtype, + device, + ) + + extra_step_kwargs = self.prepare_extra_step_kwargs(None, 0.0) + + # 7. Prepare additional timestep conditions + original_size = torch.tensor([original_size], dtype=prompt_embeds.dtype) + target_size = torch.tensor([target_size], dtype=prompt_embeds.dtype) + crops_coords_top_left = torch.tensor([(0, 0)], dtype=prompt_embeds.dtype) + + if self.do_classifier_free_guidance: + original_size = torch.cat([original_size, original_size]) + target_size = torch.cat([target_size, target_size]) + crops_coords_top_left = torch.cat([crops_coords_top_left, crops_coords_top_left]) + + original_size = original_size.to(device).repeat(batch_size * num_images_per_prompt, 1) + target_size = target_size.to(device).repeat(batch_size * num_images_per_prompt, 1) + crops_coords_top_left = crops_coords_top_left.to(device).repeat(batch_size * num_images_per_prompt, 1) + + # 8. Denoising loop + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + + with self.progress_bar(total=num_inference_steps) as progress_bar: + # for DPM-solver++ + old_pred_original_sample = None + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latent_model_input.shape[0]) + + # predict noise model_output + noise_pred = self.transformer( + states=(latent_model_input, prompt_embeds, i), + timestep=timestep, + original_size=original_size, + target_size=target_size, + crop_coords=crops_coords_top_left, + )[0] + noise_pred = noise_pred.float() + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + latents = latents.to(prompt_embeds.dtype) + + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, generator=None)[0] + image = self.image_processor.postprocess(image, output_type="pil") + + # Offload all models + self.maybe_free_model_hooks() + return CogView3PipelineOutput(images=image) \ No newline at end of file diff --git a/MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/pipeline/pipeline_output.py b/MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/pipeline/pipeline_output.py similarity index 96% rename from MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/pipeline/pipeline_output.py rename to MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/pipeline/pipeline_output.py index 11f8976f0e5ab01333d91c8d452ae25f2e798eae..e56a4485d78c6c0b8d572fd0ef09be4ed71ec5e1 100644 --- a/MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/pipeline/pipeline_output.py +++ b/MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/pipeline/pipeline_output.py @@ -1,21 +1,21 @@ -from dataclasses import dataclass -from typing import List, Union - -import numpy as np -import PIL.Image - -from diffusers.utils import BaseOutput - - -@dataclass -class CogView3PipelineOutput(BaseOutput): - """ - Output class for CogView3 pipelines. - - Args: - images (`List[PIL.Image.Image]` or `np.ndarray`) - List of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width, - num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline. - """ - +from dataclasses import dataclass +from typing import List, Union + +import numpy as np +import PIL.Image + +from diffusers.utils import BaseOutput + + +@dataclass +class CogView3PipelineOutput(BaseOutput): + """ + Output class for CogView3 pipelines. + + Args: + images (`List[PIL.Image.Image]` or `np.ndarray`) + List of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width, + num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline. + """ + images: Union[List[PIL.Image.Image], np.ndarray] \ No newline at end of file diff --git a/MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/schedulers/__init__.py b/MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/schedulers/__init__.py similarity index 99% rename from MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/schedulers/__init__.py rename to MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/schedulers/__init__.py index 7a8f559a28aee98d7c3724b0b674f7f8e1230a63..f98b6e1dec52d8560138e4c62a07abcf96766d1c 100644 --- a/MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/schedulers/__init__.py +++ b/MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/schedulers/__init__.py @@ -1,2 +1,2 @@ -from .scheduling_ddim_cogvideox import CogVideoXDDIMScheduler +from .scheduling_ddim_cogvideox import CogVideoXDDIMScheduler from .scheduling_utils import SchedulerMixin \ No newline at end of file diff --git a/MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/schedulers/scheduling_ddim_cogvideox.py b/MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/schedulers/scheduling_ddim_cogvideox.py similarity index 97% rename from MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/schedulers/scheduling_ddim_cogvideox.py rename to MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/schedulers/scheduling_ddim_cogvideox.py index b3f6ce229b251658a2efaf4ee113942898735017..b4f81796e9806a58cf994b31a324ff2ecf9acd89 100644 --- a/MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/schedulers/scheduling_ddim_cogvideox.py +++ b/MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/schedulers/scheduling_ddim_cogvideox.py @@ -1,276 +1,276 @@ -# Copyright 2024 The CogVideoX team, Tsinghua University & ZhipuAI and The HuggingFace Team. -# All rights reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# https://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -import math -from dataclasses import dataclass -from typing import List, Optional, Tuple, Union - -import numpy as np -import torch - -from diffusers.configuration_utils import ConfigMixin, register_to_config -from diffusers.utils import BaseOutput -from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin - - -@dataclass -class DDIMSchedulerOutput(BaseOutput): - prev_sample: torch.Tensor - pred_original_sample: Optional[torch.Tensor] = None - - -def betas_for_alpha_bar( - num_diffusion_timesteps, - max_beta=0.999, - alpha_transform_type="cosine", -): - if alpha_transform_type == "cosine": - - def alpha_bar_fn(t): - return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 - - elif alpha_transform_type == "exp": - - def alpha_bar_fn(t): - return math.exp(t * -12.0) - - else: - raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") - - betas = [] - for i in range(num_diffusion_timesteps): - t1 = i / num_diffusion_timesteps - t2 = (i + 1) / num_diffusion_timesteps - betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) - return torch.tensor(betas, dtype=torch.float32) - - -def rescale_zero_terminal_snr(alphas_cumprod): - alphas_bar_sqrt = alphas_cumprod.sqrt() - - # Store old values. - alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone() - alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone() - - # Shift so the last timestep is zero. - alphas_bar_sqrt -= alphas_bar_sqrt_T - - # Scale so the first timestep is back to the old value. - alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T) - - # Convert alphas_bar_sqrt to betas - alphas_bar = alphas_bar_sqrt**2 # Revert sqrt - - return alphas_bar - - -class CogVideoXDDIMScheduler(SchedulerMixin, ConfigMixin): - _compatibles = [e.name for e in KarrasDiffusionSchedulers] - order = 1 - - @register_to_config - def __init__( - self, - num_train_timesteps: int = 1000, - beta_start: float = 0.00085, - beta_end: float = 0.0120, - beta_schedule: str = "scaled_linear", - trained_betas: Optional[Union[np.ndarray, List[float]]] = None, - set_alpha_to_one: bool = True, - rescale_betas_zero_snr: bool = False, - snr_shift_scale: float = 3.0, - ): - if trained_betas is not None: - self.betas = torch.tensor(trained_betas, dtype=torch.float32) - elif beta_schedule == "linear": - self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) - elif beta_schedule == "scaled_linear": - # this schedule is very specific to the latent diffusion model. - self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float64) ** 2 - elif beta_schedule == "squaredcos_cap_v2": - # Glide cosine schedule - self.betas = betas_for_alpha_bar(num_train_timesteps) - else: - raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") - - self.alphas = 1.0 - self.betas - self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) - - # Modify: SNR shift following SD3 - self.alphas_cumprod = self.alphas_cumprod / (snr_shift_scale + (1 - snr_shift_scale) * self.alphas_cumprod) - - # Rescale for zero SNR - if rescale_betas_zero_snr: - self.alphas_cumprod = rescale_zero_terminal_snr(self.alphas_cumprod) - - self.final_alpha_cumprod = torch.tensor(1.0) if set_alpha_to_one else self.alphas_cumprod[0] - - # standard deviation of the initial noise distribution - self.init_noise_sigma = 1.0 - - # setable values - self.num_inference_steps = None - self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[::-1].copy().astype(np.int64)) - - def _get_variance(self, timestep, prev_timestep): - alpha_prod_t = self.alphas_cumprod[timestep] - alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod - beta_prod_t = 1 - alpha_prod_t - beta_prod_t_prev = 1 - alpha_prod_t_prev - - variance = (beta_prod_t_prev / beta_prod_t) * (1 - alpha_prod_t / alpha_prod_t_prev) - - return variance - - def scale_model_input(self, sample: torch.Tensor, timestep: Optional[int] = None) -> torch.Tensor: - return sample - - def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None): - """ - Sets the discrete timesteps used for the diffusion chain (to be run before inference). - - Args: - num_inference_steps (`int`): - The number of diffusion steps used when generating samples with a pre-trained model. - """ - - if num_inference_steps > self.config.num_train_timesteps: - raise ValueError( - f"`num_inference_steps`: {num_inference_steps} cannot be larger than `self.config.train_timesteps`:" - f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle" - f" maximal {self.config.num_train_timesteps} timesteps." - ) - - self.num_inference_steps = num_inference_steps - - if self.config.timestep_spacing == "linspace": - timesteps = ( - np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps) - .round()[::-1] - .copy() - .astype(np.int64) - ) - elif self.config.timestep_spacing == "leading": - step_ratio = self.config.num_train_timesteps // self.num_inference_steps - # creates integer timesteps by multiplying by ratio - # casting to int to avoid issues when num_inference_step is power of 3 - timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.int64) - timesteps += self.config.steps_offset - elif self.config.timestep_spacing == "trailing": - step_ratio = self.config.num_train_timesteps / self.num_inference_steps - # creates integer timesteps by multiplying by ratio - # casting to int to avoid issues when num_inference_step is power of 3 - timesteps = np.round(np.arange(self.config.num_train_timesteps, 0, -step_ratio)).astype(np.int64) - timesteps -= 1 - else: - raise ValueError( - f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'leading' or 'trailing'." - ) - - self.timesteps = torch.from_numpy(timesteps).to(device) - - def step( - self, - model_output: torch.Tensor, - timestep: int, - sample: torch.Tensor, - return_dict: bool = True, - ) -> Union[DDIMSchedulerOutput, Tuple]: - if self.num_inference_steps is None: - raise ValueError( - "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" - ) - - prev_timestep = timestep - self.config.num_train_timesteps // self.num_inference_steps - - # 2. compute alphas, betas - alpha_prod_t = self.alphas_cumprod[timestep] - alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod - - beta_prod_t = 1 - alpha_prod_t - - # 3. compute predicted original sample from predicted noise also called - if self.config.prediction_type == "epsilon": - pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5) - elif self.config.prediction_type == "sample": - pred_original_sample = model_output - elif self.config.prediction_type == "v_prediction": - pred_original_sample = (alpha_prod_t**0.5) * sample - (beta_prod_t**0.5) * model_output - else: - raise ValueError( - f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or" - " `v_prediction`" - ) - - a_t = ((1 - alpha_prod_t_prev) / (1 - alpha_prod_t)) ** 0.5 - b_t = alpha_prod_t_prev**0.5 - alpha_prod_t**0.5 * a_t - - prev_sample = a_t * sample + b_t * pred_original_sample - - if not return_dict: - return ( - prev_sample, - pred_original_sample, - ) - - return DDIMSchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_original_sample) - - def add_noise( - self, - original_samples: torch.Tensor, - noise: torch.Tensor, - timesteps: torch.IntTensor, - ) -> torch.Tensor: - # Make sure alphas_cumprod and timestep have same device and dtype as original_samples - # Move the self.alphas_cumprod to device to avoid redundant CPU to GPU data movement - # for the subsequent add_noise calls - self.alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device) - alphas_cumprod = self.alphas_cumprod.to(dtype=original_samples.dtype) - timesteps = timesteps.to(original_samples.device) - - sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 - sqrt_alpha_prod = sqrt_alpha_prod.flatten() - while len(sqrt_alpha_prod.shape) < len(original_samples.shape): - sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) - - sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 - sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() - while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape): - sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) - - noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise - return noisy_samples - - def get_velocity(self, sample: torch.Tensor, noise: torch.Tensor, timesteps: torch.IntTensor) -> torch.Tensor: - # Make sure alphas_cumprod and timestep have same device and dtype as sample - self.alphas_cumprod = self.alphas_cumprod.to(device=sample.device) - alphas_cumprod = self.alphas_cumprod.to(dtype=sample.dtype) - timesteps = timesteps.to(sample.device) - - sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 - sqrt_alpha_prod = sqrt_alpha_prod.flatten() - while len(sqrt_alpha_prod.shape) < len(sample.shape): - sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) - - sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 - sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() - while len(sqrt_one_minus_alpha_prod.shape) < len(sample.shape): - sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) - - velocity = sqrt_alpha_prod * noise - sqrt_one_minus_alpha_prod * sample - return velocity - - def __len__(self): +# Copyright 2024 The CogVideoX team, Tsinghua University & ZhipuAI and The HuggingFace Team. +# All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# https://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import math +from dataclasses import dataclass +from typing import List, Optional, Tuple, Union + +import numpy as np +import torch + +from diffusers.configuration_utils import ConfigMixin, register_to_config +from diffusers.utils import BaseOutput +from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin + + +@dataclass +class DDIMSchedulerOutput(BaseOutput): + prev_sample: torch.Tensor + pred_original_sample: Optional[torch.Tensor] = None + + +def betas_for_alpha_bar( + num_diffusion_timesteps, + max_beta=0.999, + alpha_transform_type="cosine", +): + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +def rescale_zero_terminal_snr(alphas_cumprod): + alphas_bar_sqrt = alphas_cumprod.sqrt() + + # Store old values. + alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone() + alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone() + + # Shift so the last timestep is zero. + alphas_bar_sqrt -= alphas_bar_sqrt_T + + # Scale so the first timestep is back to the old value. + alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T) + + # Convert alphas_bar_sqrt to betas + alphas_bar = alphas_bar_sqrt**2 # Revert sqrt + + return alphas_bar + + +class CogVideoXDDIMScheduler(SchedulerMixin, ConfigMixin): + _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.00085, + beta_end: float = 0.0120, + beta_schedule: str = "scaled_linear", + trained_betas: Optional[Union[np.ndarray, List[float]]] = None, + set_alpha_to_one: bool = True, + rescale_betas_zero_snr: bool = False, + snr_shift_scale: float = 3.0, + ): + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float64) ** 2 + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + + # Modify: SNR shift following SD3 + self.alphas_cumprod = self.alphas_cumprod / (snr_shift_scale + (1 - snr_shift_scale) * self.alphas_cumprod) + + # Rescale for zero SNR + if rescale_betas_zero_snr: + self.alphas_cumprod = rescale_zero_terminal_snr(self.alphas_cumprod) + + self.final_alpha_cumprod = torch.tensor(1.0) if set_alpha_to_one else self.alphas_cumprod[0] + + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + # setable values + self.num_inference_steps = None + self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[::-1].copy().astype(np.int64)) + + def _get_variance(self, timestep, prev_timestep): + alpha_prod_t = self.alphas_cumprod[timestep] + alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod + beta_prod_t = 1 - alpha_prod_t + beta_prod_t_prev = 1 - alpha_prod_t_prev + + variance = (beta_prod_t_prev / beta_prod_t) * (1 - alpha_prod_t / alpha_prod_t_prev) + + return variance + + def scale_model_input(self, sample: torch.Tensor, timestep: Optional[int] = None) -> torch.Tensor: + return sample + + def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + """ + + if num_inference_steps > self.config.num_train_timesteps: + raise ValueError( + f"`num_inference_steps`: {num_inference_steps} cannot be larger than `self.config.train_timesteps`:" + f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle" + f" maximal {self.config.num_train_timesteps} timesteps." + ) + + self.num_inference_steps = num_inference_steps + + if self.config.timestep_spacing == "linspace": + timesteps = ( + np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps) + .round()[::-1] + .copy() + .astype(np.int64) + ) + elif self.config.timestep_spacing == "leading": + step_ratio = self.config.num_train_timesteps // self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.int64) + timesteps += self.config.steps_offset + elif self.config.timestep_spacing == "trailing": + step_ratio = self.config.num_train_timesteps / self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = np.round(np.arange(self.config.num_train_timesteps, 0, -step_ratio)).astype(np.int64) + timesteps -= 1 + else: + raise ValueError( + f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'leading' or 'trailing'." + ) + + self.timesteps = torch.from_numpy(timesteps).to(device) + + def step( + self, + model_output: torch.Tensor, + timestep: int, + sample: torch.Tensor, + return_dict: bool = True, + ) -> Union[DDIMSchedulerOutput, Tuple]: + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + prev_timestep = timestep - self.config.num_train_timesteps // self.num_inference_steps + + # 2. compute alphas, betas + alpha_prod_t = self.alphas_cumprod[timestep] + alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod + + beta_prod_t = 1 - alpha_prod_t + + # 3. compute predicted original sample from predicted noise also called + if self.config.prediction_type == "epsilon": + pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5) + elif self.config.prediction_type == "sample": + pred_original_sample = model_output + elif self.config.prediction_type == "v_prediction": + pred_original_sample = (alpha_prod_t**0.5) * sample - (beta_prod_t**0.5) * model_output + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or" + " `v_prediction`" + ) + + a_t = ((1 - alpha_prod_t_prev) / (1 - alpha_prod_t)) ** 0.5 + b_t = alpha_prod_t_prev**0.5 - alpha_prod_t**0.5 * a_t + + prev_sample = a_t * sample + b_t * pred_original_sample + + if not return_dict: + return ( + prev_sample, + pred_original_sample, + ) + + return DDIMSchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_original_sample) + + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.IntTensor, + ) -> torch.Tensor: + # Make sure alphas_cumprod and timestep have same device and dtype as original_samples + # Move the self.alphas_cumprod to device to avoid redundant CPU to GPU data movement + # for the subsequent add_noise calls + self.alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device) + alphas_cumprod = self.alphas_cumprod.to(dtype=original_samples.dtype) + timesteps = timesteps.to(original_samples.device) + + sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(original_samples.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise + return noisy_samples + + def get_velocity(self, sample: torch.Tensor, noise: torch.Tensor, timesteps: torch.IntTensor) -> torch.Tensor: + # Make sure alphas_cumprod and timestep have same device and dtype as sample + self.alphas_cumprod = self.alphas_cumprod.to(device=sample.device) + alphas_cumprod = self.alphas_cumprod.to(dtype=sample.dtype) + timesteps = timesteps.to(sample.device) + + sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(sample.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(sample.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + velocity = sqrt_alpha_prod * noise - sqrt_one_minus_alpha_prod * sample + return velocity + + def __len__(self): return self.config.num_train_timesteps \ No newline at end of file diff --git a/MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/schedulers/scheduling_utils.py b/MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/schedulers/scheduling_utils.py similarity index 97% rename from MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/schedulers/scheduling_utils.py rename to MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/schedulers/scheduling_utils.py index d854366c7791c96c5e890b1aae0d4800ac657e1f..eeb6e77dee08d3c5a8245394afd7b090847c8d28 100644 --- a/MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/schedulers/scheduling_utils.py +++ b/MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/schedulers/scheduling_utils.py @@ -1,113 +1,113 @@ -# Copyright 2024 The HuggingFace Team. All rights reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# https://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -import importlib -import os -from dataclasses import dataclass -from enum import Enum -from typing import Optional, Union - -import torch -from huggingface_hub.utils import validate_hf_hub_args - -from diffusers.utils import BaseOutput, PushToHubMixin - - -SCHEDULER_CONFIG_NAME = "scheduler_config.json" - - -class KarrasDiffusionSchedulers(Enum): - DDIMScheduler = 1 - DDPMScheduler = 2 - PNDMScheduler = 3 - LMSDiscreteScheduler = 4 - EulerDiscreteScheduler = 5 - HeunDiscreteScheduler = 6 - EulerAncestralDiscreteScheduler = 7 - DPMSolverMultistepScheduler = 8 - DPMSolverSinglestepScheduler = 9 - KDPM2DiscreteScheduler = 10 - KDPM2AncestralDiscreteScheduler = 11 - DEISMultistepScheduler = 12 - UniPCMultistepScheduler = 13 - DPMSolverSDEScheduler = 14 - EDMEulerScheduler = 15 - - -AysSchedules = { - "StableDiffusionTimesteps": [999, 850, 736, 645, 545, 455, 343, 233, 124, 24], - "StableDiffusionSigmas": [14.615, 6.475, 3.861, 2.697, 1.886, 1.396, 0.963, 0.652, 0.399, 0.152, 0.0], - "StableDiffusionXLTimesteps": [999, 845, 730, 587, 443, 310, 193, 116, 53, 13], - "StableDiffusionXLSigmas": [14.615, 6.315, 3.771, 2.181, 1.342, 0.862, 0.555, 0.380, 0.234, 0.113, 0.0], - "StableDiffusionVideoSigmas": [700.00, 54.5, 15.886, 7.977, 4.248, 1.789, 0.981, 0.403, 0.173, 0.034, 0.0], -} - - -@dataclass -class SchedulerOutput(BaseOutput): - """ - Base class for the output of a scheduler's `step` function. - - Args: - prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): - Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the - denoising loop. - """ - - prev_sample: torch.Tensor - - -class SchedulerMixin(PushToHubMixin): - - config_name = SCHEDULER_CONFIG_NAME - _compatibles = [] - has_compatibles = True - - @classmethod - @validate_hf_hub_args - def from_pretrained( - cls, - pretrained_model_name_or_path: Optional[Union[str, os.PathLike]] = None, - subfolder: Optional[str] = None, - return_unused_kwargs=False, - **kwargs, - ): - - config, kwargs, _ = cls.load_config( - pretrained_model_name_or_path=pretrained_model_name_or_path, - subfolder=subfolder, - return_unused_kwargs=True, - return_commit_hash=True, - **kwargs, - ) - return cls.from_config(config, return_unused_kwargs=return_unused_kwargs, **kwargs) - - @property - def compatibles(self): - """ - Returns all schedulers that are compatible with this scheduler - - Returns: - `List[SchedulerMixin]`: List of compatible schedulers - """ - return self._get_compatibles() - - @classmethod - def _get_compatibles(cls): - compatible_classes_str = list(set([cls.__name__] + cls._compatibles)) - diffusers_library = importlib.import_module(__name__.split(".")[0]) - compatible_classes = [ - getattr(diffusers_library, c) for c in compatible_classes_str if hasattr(diffusers_library, c) - ] +# Copyright 2024 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# https://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import importlib +import os +from dataclasses import dataclass +from enum import Enum +from typing import Optional, Union + +import torch +from huggingface_hub.utils import validate_hf_hub_args + +from diffusers.utils import BaseOutput, PushToHubMixin + + +SCHEDULER_CONFIG_NAME = "scheduler_config.json" + + +class KarrasDiffusionSchedulers(Enum): + DDIMScheduler = 1 + DDPMScheduler = 2 + PNDMScheduler = 3 + LMSDiscreteScheduler = 4 + EulerDiscreteScheduler = 5 + HeunDiscreteScheduler = 6 + EulerAncestralDiscreteScheduler = 7 + DPMSolverMultistepScheduler = 8 + DPMSolverSinglestepScheduler = 9 + KDPM2DiscreteScheduler = 10 + KDPM2AncestralDiscreteScheduler = 11 + DEISMultistepScheduler = 12 + UniPCMultistepScheduler = 13 + DPMSolverSDEScheduler = 14 + EDMEulerScheduler = 15 + + +AysSchedules = { + "StableDiffusionTimesteps": [999, 850, 736, 645, 545, 455, 343, 233, 124, 24], + "StableDiffusionSigmas": [14.615, 6.475, 3.861, 2.697, 1.886, 1.396, 0.963, 0.652, 0.399, 0.152, 0.0], + "StableDiffusionXLTimesteps": [999, 845, 730, 587, 443, 310, 193, 116, 53, 13], + "StableDiffusionXLSigmas": [14.615, 6.315, 3.771, 2.181, 1.342, 0.862, 0.555, 0.380, 0.234, 0.113, 0.0], + "StableDiffusionVideoSigmas": [700.00, 54.5, 15.886, 7.977, 4.248, 1.789, 0.981, 0.403, 0.173, 0.034, 0.0], +} + + +@dataclass +class SchedulerOutput(BaseOutput): + """ + Base class for the output of a scheduler's `step` function. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + """ + + prev_sample: torch.Tensor + + +class SchedulerMixin(PushToHubMixin): + + config_name = SCHEDULER_CONFIG_NAME + _compatibles = [] + has_compatibles = True + + @classmethod + @validate_hf_hub_args + def from_pretrained( + cls, + pretrained_model_name_or_path: Optional[Union[str, os.PathLike]] = None, + subfolder: Optional[str] = None, + return_unused_kwargs=False, + **kwargs, + ): + + config, kwargs, _ = cls.load_config( + pretrained_model_name_or_path=pretrained_model_name_or_path, + subfolder=subfolder, + return_unused_kwargs=True, + return_commit_hash=True, + **kwargs, + ) + return cls.from_config(config, return_unused_kwargs=return_unused_kwargs, **kwargs) + + @property + def compatibles(self): + """ + Returns all schedulers that are compatible with this scheduler + + Returns: + `List[SchedulerMixin]`: List of compatible schedulers + """ + return self._get_compatibles() + + @classmethod + def _get_compatibles(cls): + compatible_classes_str = list(set([cls.__name__] + cls._compatibles)) + diffusers_library = importlib.import_module(__name__.split(".")[0]) + compatible_classes = [ + getattr(diffusers_library, c) for c in compatible_classes_str if hasattr(diffusers_library, c) + ] return compatible_classes \ No newline at end of file diff --git a/MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/vae/__init__.py b/MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/vae/__init__.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/cogview3/cogview3plus/vae/__init__.py rename to MindIE/MultiModal/CogView3-Plus-3B/cogview3plus/vae/__init__.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/cogview3/inference_cogview3plus.py b/MindIE/MultiModal/CogView3-Plus-3B/inference_cogview3plus.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/cogview3/inference_cogview3plus.py rename to MindIE/MultiModal/CogView3-Plus-3B/inference_cogview3plus.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/cogview3/requirents.txt b/MindIE/MultiModal/CogView3-Plus-3B/requirents.txt similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/cogview3/requirents.txt rename to MindIE/MultiModal/CogView3-Plus-3B/requirents.txt diff --git a/MindIE/MindIE-Torch/built-in/foundation/Flux.1-DEV/FLUX1dev/__init__.py b/MindIE/MultiModal/Flux.1-DEV/FLUX1dev/__init__.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/Flux.1-DEV/FLUX1dev/__init__.py rename to MindIE/MultiModal/Flux.1-DEV/FLUX1dev/__init__.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/Flux.1-DEV/FLUX1dev/layers/__init__.py b/MindIE/MultiModal/Flux.1-DEV/FLUX1dev/layers/__init__.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/Flux.1-DEV/FLUX1dev/layers/__init__.py rename to MindIE/MultiModal/Flux.1-DEV/FLUX1dev/layers/__init__.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/Flux.1-DEV/FLUX1dev/layers/attention_processor.py b/MindIE/MultiModal/Flux.1-DEV/FLUX1dev/layers/attention_processor.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/Flux.1-DEV/FLUX1dev/layers/attention_processor.py rename to MindIE/MultiModal/Flux.1-DEV/FLUX1dev/layers/attention_processor.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/Flux.1-DEV/FLUX1dev/layers/embedding.py b/MindIE/MultiModal/Flux.1-DEV/FLUX1dev/layers/embedding.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/Flux.1-DEV/FLUX1dev/layers/embedding.py rename to MindIE/MultiModal/Flux.1-DEV/FLUX1dev/layers/embedding.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/Flux.1-DEV/FLUX1dev/models/__init__.py b/MindIE/MultiModal/Flux.1-DEV/FLUX1dev/models/__init__.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/Flux.1-DEV/FLUX1dev/models/__init__.py rename to MindIE/MultiModal/Flux.1-DEV/FLUX1dev/models/__init__.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/Flux.1-DEV/FLUX1dev/models/modeling_utils.py b/MindIE/MultiModal/Flux.1-DEV/FLUX1dev/models/modeling_utils.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/Flux.1-DEV/FLUX1dev/models/modeling_utils.py rename to MindIE/MultiModal/Flux.1-DEV/FLUX1dev/models/modeling_utils.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/Flux.1-DEV/FLUX1dev/models/transformer_flux.py b/MindIE/MultiModal/Flux.1-DEV/FLUX1dev/models/transformer_flux.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/Flux.1-DEV/FLUX1dev/models/transformer_flux.py rename to MindIE/MultiModal/Flux.1-DEV/FLUX1dev/models/transformer_flux.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/Flux.1-DEV/FLUX1dev/pipeline/__init__.py b/MindIE/MultiModal/Flux.1-DEV/FLUX1dev/pipeline/__init__.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/Flux.1-DEV/FLUX1dev/pipeline/__init__.py rename to MindIE/MultiModal/Flux.1-DEV/FLUX1dev/pipeline/__init__.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/Flux.1-DEV/FLUX1dev/pipeline/pipeline_flux.py b/MindIE/MultiModal/Flux.1-DEV/FLUX1dev/pipeline/pipeline_flux.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/Flux.1-DEV/FLUX1dev/pipeline/pipeline_flux.py rename to MindIE/MultiModal/Flux.1-DEV/FLUX1dev/pipeline/pipeline_flux.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/Flux.1-DEV/README.md b/MindIE/MultiModal/Flux.1-DEV/README.md similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/Flux.1-DEV/README.md rename to MindIE/MultiModal/Flux.1-DEV/README.md diff --git a/MindIE/MindIE-Torch/built-in/foundation/Flux.1-DEV/inference_flux.py b/MindIE/MultiModal/Flux.1-DEV/inference_flux.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/Flux.1-DEV/inference_flux.py rename to MindIE/MultiModal/Flux.1-DEV/inference_flux.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/Flux.1-DEV/prompts.txt b/MindIE/MultiModal/Flux.1-DEV/prompts.txt similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/Flux.1-DEV/prompts.txt rename to MindIE/MultiModal/Flux.1-DEV/prompts.txt diff --git a/MindIE/MindIE-Torch/built-in/foundation/Flux.1-DEV/requirements.txt b/MindIE/MultiModal/Flux.1-DEV/requirements.txt similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/Flux.1-DEV/requirements.txt rename to MindIE/MultiModal/Flux.1-DEV/requirements.txt diff --git a/MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/README.md b/MindIE/MultiModal/HunyuanDiT/README.md similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/README.md rename to MindIE/MultiModal/HunyuanDiT/README.md diff --git a/MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/hydit/__init__.py b/MindIE/MultiModal/HunyuanDiT/hydit/__init__.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/hydit/__init__.py rename to MindIE/MultiModal/HunyuanDiT/hydit/__init__.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/hydit/layers/__init__.py b/MindIE/MultiModal/HunyuanDiT/hydit/layers/__init__.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/hydit/layers/__init__.py rename to MindIE/MultiModal/HunyuanDiT/hydit/layers/__init__.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/hydit/layers/activation.py b/MindIE/MultiModal/HunyuanDiT/hydit/layers/activation.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/hydit/layers/activation.py rename to MindIE/MultiModal/HunyuanDiT/hydit/layers/activation.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/hydit/layers/attention.py b/MindIE/MultiModal/HunyuanDiT/hydit/layers/attention.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/hydit/layers/attention.py rename to MindIE/MultiModal/HunyuanDiT/hydit/layers/attention.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/hydit/layers/embedding.py b/MindIE/MultiModal/HunyuanDiT/hydit/layers/embedding.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/hydit/layers/embedding.py rename to MindIE/MultiModal/HunyuanDiT/hydit/layers/embedding.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/hydit/layers/mlp.py b/MindIE/MultiModal/HunyuanDiT/hydit/layers/mlp.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/hydit/layers/mlp.py rename to MindIE/MultiModal/HunyuanDiT/hydit/layers/mlp.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/hydit/layers/norm.py b/MindIE/MultiModal/HunyuanDiT/hydit/layers/norm.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/hydit/layers/norm.py rename to MindIE/MultiModal/HunyuanDiT/hydit/layers/norm.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/hydit/layers/poolers.py b/MindIE/MultiModal/HunyuanDiT/hydit/layers/poolers.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/hydit/layers/poolers.py rename to MindIE/MultiModal/HunyuanDiT/hydit/layers/poolers.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/hydit/models/__init__.py b/MindIE/MultiModal/HunyuanDiT/hydit/models/__init__.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/hydit/models/__init__.py rename to MindIE/MultiModal/HunyuanDiT/hydit/models/__init__.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/hydit/models/hydit.py b/MindIE/MultiModal/HunyuanDiT/hydit/models/hydit.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/hydit/models/hydit.py rename to MindIE/MultiModal/HunyuanDiT/hydit/models/hydit.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/hydit/models/model_load_utils.py b/MindIE/MultiModal/HunyuanDiT/hydit/models/model_load_utils.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/hydit/models/model_load_utils.py rename to MindIE/MultiModal/HunyuanDiT/hydit/models/model_load_utils.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/hydit/models/model_utils.py b/MindIE/MultiModal/HunyuanDiT/hydit/models/model_utils.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/hydit/models/model_utils.py rename to MindIE/MultiModal/HunyuanDiT/hydit/models/model_utils.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/hydit/pipeline/__init__.py b/MindIE/MultiModal/HunyuanDiT/hydit/pipeline/__init__.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/hydit/pipeline/__init__.py rename to MindIE/MultiModal/HunyuanDiT/hydit/pipeline/__init__.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/hydit/pipeline/hydit_pipeline.py b/MindIE/MultiModal/HunyuanDiT/hydit/pipeline/hydit_pipeline.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/hydit/pipeline/hydit_pipeline.py rename to MindIE/MultiModal/HunyuanDiT/hydit/pipeline/hydit_pipeline.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/hydit/schedulers/__init__.py b/MindIE/MultiModal/HunyuanDiT/hydit/schedulers/__init__.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/hydit/schedulers/__init__.py rename to MindIE/MultiModal/HunyuanDiT/hydit/schedulers/__init__.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/hydit/schedulers/ddpm.py b/MindIE/MultiModal/HunyuanDiT/hydit/schedulers/ddpm.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/hydit/schedulers/ddpm.py rename to MindIE/MultiModal/HunyuanDiT/hydit/schedulers/ddpm.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/hydit/utils/__init__.py b/MindIE/MultiModal/HunyuanDiT/hydit/utils/__init__.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/hydit/utils/__init__.py rename to MindIE/MultiModal/HunyuanDiT/hydit/utils/__init__.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/hydit/utils/file_utils.py b/MindIE/MultiModal/HunyuanDiT/hydit/utils/file_utils.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/hydit/utils/file_utils.py rename to MindIE/MultiModal/HunyuanDiT/hydit/utils/file_utils.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/hydit/utils/utils.py b/MindIE/MultiModal/HunyuanDiT/hydit/utils/utils.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/hydit/utils/utils.py rename to MindIE/MultiModal/HunyuanDiT/hydit/utils/utils.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/inference_hydit.py b/MindIE/MultiModal/HunyuanDiT/inference_hydit.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/inference_hydit.py rename to MindIE/MultiModal/HunyuanDiT/inference_hydit.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/lora/__init__.py b/MindIE/MultiModal/HunyuanDiT/lora/__init__.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/lora/__init__.py rename to MindIE/MultiModal/HunyuanDiT/lora/__init__.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/lora/hydit_lora.py b/MindIE/MultiModal/HunyuanDiT/lora/hydit_lora.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/lora/hydit_lora.py rename to MindIE/MultiModal/HunyuanDiT/lora/hydit_lora.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/prompts/example_prompts.txt b/MindIE/MultiModal/HunyuanDiT/prompts/example_prompts.txt similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/prompts/example_prompts.txt rename to MindIE/MultiModal/HunyuanDiT/prompts/example_prompts.txt diff --git a/MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/requirents.txt b/MindIE/MultiModal/HunyuanDiT/requirents.txt similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/hunyuan_dit/requirents.txt rename to MindIE/MultiModal/HunyuanDiT/requirents.txt diff --git a/MindIE/MindIE-Torch/built-in/foundation/opensora1.2/README.md b/MindIE/MultiModal/OpenSora-v1.2/README.md similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/opensora1.2/README.md rename to MindIE/MultiModal/OpenSora-v1.2/README.md diff --git a/MindIE/MindIE-Torch/built-in/foundation/opensora1.2/inference_opensora12.py b/MindIE/MultiModal/OpenSora-v1.2/inference_opensora12.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/opensora1.2/inference_opensora12.py rename to MindIE/MultiModal/OpenSora-v1.2/inference_opensora12.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/__init__.py b/MindIE/MultiModal/OpenSora-v1.2/opensora/__init__.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/__init__.py rename to MindIE/MultiModal/OpenSora-v1.2/opensora/__init__.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/layer/__init__.py b/MindIE/MultiModal/OpenSora-v1.2/opensora/layer/__init__.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/layer/__init__.py rename to MindIE/MultiModal/OpenSora-v1.2/opensora/layer/__init__.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/layer/activation.py b/MindIE/MultiModal/OpenSora-v1.2/opensora/layer/activation.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/layer/activation.py rename to MindIE/MultiModal/OpenSora-v1.2/opensora/layer/activation.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/layer/attention.py b/MindIE/MultiModal/OpenSora-v1.2/opensora/layer/attention.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/layer/attention.py rename to MindIE/MultiModal/OpenSora-v1.2/opensora/layer/attention.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/layer/comm.py b/MindIE/MultiModal/OpenSora-v1.2/opensora/layer/comm.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/layer/comm.py rename to MindIE/MultiModal/OpenSora-v1.2/opensora/layer/comm.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/layer/conv.py b/MindIE/MultiModal/OpenSora-v1.2/opensora/layer/conv.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/layer/conv.py rename to MindIE/MultiModal/OpenSora-v1.2/opensora/layer/conv.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/layer/embdding.py b/MindIE/MultiModal/OpenSora-v1.2/opensora/layer/embdding.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/layer/embdding.py rename to MindIE/MultiModal/OpenSora-v1.2/opensora/layer/embdding.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/layer/mlp.py b/MindIE/MultiModal/OpenSora-v1.2/opensora/layer/mlp.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/layer/mlp.py rename to MindIE/MultiModal/OpenSora-v1.2/opensora/layer/mlp.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/layer/norm.py b/MindIE/MultiModal/OpenSora-v1.2/opensora/layer/norm.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/layer/norm.py rename to MindIE/MultiModal/OpenSora-v1.2/opensora/layer/norm.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/layer/parallel_mgr.py b/MindIE/MultiModal/OpenSora-v1.2/opensora/layer/parallel_mgr.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/layer/parallel_mgr.py rename to MindIE/MultiModal/OpenSora-v1.2/opensora/layer/parallel_mgr.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/layer/utils.py b/MindIE/MultiModal/OpenSora-v1.2/opensora/layer/utils.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/layer/utils.py rename to MindIE/MultiModal/OpenSora-v1.2/opensora/layer/utils.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/pipeline/__init__.py b/MindIE/MultiModal/OpenSora-v1.2/opensora/pipeline/__init__.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/pipeline/__init__.py rename to MindIE/MultiModal/OpenSora-v1.2/opensora/pipeline/__init__.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/pipeline/compile_pipe.py b/MindIE/MultiModal/OpenSora-v1.2/opensora/pipeline/compile_pipe.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/pipeline/compile_pipe.py rename to MindIE/MultiModal/OpenSora-v1.2/opensora/pipeline/compile_pipe.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/pipeline/open_sora_pipeline.py b/MindIE/MultiModal/OpenSora-v1.2/opensora/pipeline/open_sora_pipeline.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/pipeline/open_sora_pipeline.py rename to MindIE/MultiModal/OpenSora-v1.2/opensora/pipeline/open_sora_pipeline.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/pipeline/pipeline_utils.py b/MindIE/MultiModal/OpenSora-v1.2/opensora/pipeline/pipeline_utils.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/pipeline/pipeline_utils.py rename to MindIE/MultiModal/OpenSora-v1.2/opensora/pipeline/pipeline_utils.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/schedulers/__init__.py b/MindIE/MultiModal/OpenSora-v1.2/opensora/schedulers/__init__.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/schedulers/__init__.py rename to MindIE/MultiModal/OpenSora-v1.2/opensora/schedulers/__init__.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/schedulers/rectified_flow.py b/MindIE/MultiModal/OpenSora-v1.2/opensora/schedulers/rectified_flow.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/schedulers/rectified_flow.py rename to MindIE/MultiModal/OpenSora-v1.2/opensora/schedulers/rectified_flow.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/stdit3/__init__.py b/MindIE/MultiModal/OpenSora-v1.2/opensora/stdit3/__init__.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/stdit3/__init__.py rename to MindIE/MultiModal/OpenSora-v1.2/opensora/stdit3/__init__.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/stdit3/stdit3.py b/MindIE/MultiModal/OpenSora-v1.2/opensora/stdit3/stdit3.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/stdit3/stdit3.py rename to MindIE/MultiModal/OpenSora-v1.2/opensora/stdit3/stdit3.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/utils/__init__.py b/MindIE/MultiModal/OpenSora-v1.2/opensora/utils/__init__.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/utils/__init__.py rename to MindIE/MultiModal/OpenSora-v1.2/opensora/utils/__init__.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/utils/patch_utils.py b/MindIE/MultiModal/OpenSora-v1.2/opensora/utils/patch_utils.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/utils/patch_utils.py rename to MindIE/MultiModal/OpenSora-v1.2/opensora/utils/patch_utils.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/utils/utils.py b/MindIE/MultiModal/OpenSora-v1.2/opensora/utils/utils.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/utils/utils.py rename to MindIE/MultiModal/OpenSora-v1.2/opensora/utils/utils.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/vae/VideoAutoencoder.py b/MindIE/MultiModal/OpenSora-v1.2/opensora/vae/VideoAutoencoder.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/vae/VideoAutoencoder.py rename to MindIE/MultiModal/OpenSora-v1.2/opensora/vae/VideoAutoencoder.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/vae/__init__.py b/MindIE/MultiModal/OpenSora-v1.2/opensora/vae/__init__.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/vae/__init__.py rename to MindIE/MultiModal/OpenSora-v1.2/opensora/vae/__init__.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/vae/vae_temporal.py b/MindIE/MultiModal/OpenSora-v1.2/opensora/vae/vae_temporal.py similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/opensora1.2/opensora/vae/vae_temporal.py rename to MindIE/MultiModal/OpenSora-v1.2/opensora/vae/vae_temporal.py diff --git a/MindIE/MindIE-Torch/built-in/foundation/opensora1.2/prompts/t2v_sora.txt b/MindIE/MultiModal/OpenSora-v1.2/prompts/t2v_sora.txt similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/opensora1.2/prompts/t2v_sora.txt rename to MindIE/MultiModal/OpenSora-v1.2/prompts/t2v_sora.txt diff --git a/MindIE/MindIE-Torch/built-in/foundation/opensora1.2/requirents.txt b/MindIE/MultiModal/OpenSora-v1.2/requirents.txt similarity index 100% rename from MindIE/MindIE-Torch/built-in/foundation/opensora1.2/requirents.txt rename to MindIE/MultiModal/OpenSora-v1.2/requirents.txt diff --git a/MindIE/MultiModal/StableAudioOpen-1.0 b/MindIE/MultiModal/StableAudioOpen-1.0 new file mode 160000 index 0000000000000000000000000000000000000000..795a72eb18d24edf0a2c2518c7fda2b567f9c61f --- /dev/null +++ b/MindIE/MultiModal/StableAudioOpen-1.0 @@ -0,0 +1 @@ +Subproject commit 795a72eb18d24edf0a2c2518c7fda2b567f9c61f diff --git a/MindIE/MultiModal/StableDiffusion-1.5/README.md b/MindIE/MultiModal/StableDiffusion-1.5/README.md new file mode 100644 index 0000000000000000000000000000000000000000..34893568d49320c576a224e3b93d7424df219a4b --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-1.5/README.md @@ -0,0 +1,130 @@ +--- +license: apache-2.0 +pipeline_tag: text-to-image +frameworks: + - PyTorch +library_name: openmind +hardwares: + - NPU +language: + - en +--- +## 一、准备运行环境 + + **表 1** 版本配套表 + + | 配套 | 版本 | 环境准备指导 | + | ----- | ----- |-----| + | Python | 3.10.2 | - | + | torch | 2.1.0 | - | + +### 1.1 获取CANN&MindIE安装包&环境准备 +- 设备支持 +Atlas 800I A2推理设备：支持的卡数为1或2 +Atlas 300I Duo推理卡：支持的卡数为1，可双芯并行 +- [Atlas 800I A2](https://www.hiascend.com/developer/download/community/result?module=pt+ie+cann&product=4&model=32) +- [Atlas 300I Duo](https://www.hiascend.com/developer/download/community/result?module=pt+ie+cann&product=2&model=17) +- [环境准备指导](https://www.hiascend.com/document/detail/zh/CANNCommunityEdition/80RC2alpha002/softwareinst/instg/instg_0001.html) + +### 1.2 CANN安装 +```shell +# 增加软件包可执行权限，{version}表示软件版本号，{arch}表示CPU架构，{soc}表示昇腾AI处理器的版本。 +chmod +x ./Ascend-cann-toolkit_{version}_linux-{arch}.run +chmod +x ./Ascend-cann-kernels-{soc}_{version}_linux.run +# 校验软件包安装文件的一致性和完整性 +./Ascend-cann-toolkit_{version}_linux-{arch}.run --check +./Ascend-cann-kernels-{soc}_{version}_linux.run --check +# 安装 +./Ascend-cann-toolkit_{version}_linux-{arch}.run --install +./Ascend-cann-kernels-{soc}_{version}_linux.run --install + +# 设置环境变量 +source /usr/local/Ascend/ascend-toolkit/set_env.sh +``` + + +### 1.3 MindIE安装 +```shell +# 增加软件包可执行权限，{version}表示软件版本号，{arch}表示CPU架构。 +chmod +x ./Ascend-mindie_${version}_linux-${arch}.run +./Ascend-mindie_${version}_linux-${arch}.run --check + +# 方式一：默认路径安装 +./Ascend-mindie_${version}_linux-${arch}.run --install +# 设置环境变量 +cd /usr/local/Ascend/mindie && source set_env.sh + +# 方式二：指定路径安装 +./Ascend-mindie_${version}_linux-${arch}.run --install-path=${AieInstallPath} +# 设置环境变量 +cd ${AieInstallPath}/mindie && source set_env.sh +``` + +### 1.4 Torch_npu安装 +安装pytorch框架 版本2.1.0 +[安装包下载](https://download.pytorch.org/whl/cpu/torch/) + +使用pip安装 +```shell +# {version}表示软件版本号，{arch}表示CPU架构。 +pip install torch-${version}-cp310-cp310-linux_${arch}.whl +``` +下载 pytorch_v{pytorchversion}_py{pythonversion}.tar.gz +```shell +tar -xzvf pytorch_v{pytorchversion}_py{pythonversion}.tar.gz +# 解压后，会有whl包 +pip install torch_npu-{pytorchversion}.xxxx.{arch}.whl +``` +## 二、下载本仓库 + +### 2.1 下载到本地 +```shell +git clone https://gitee.com/ascend/ModelZoo-PyTorch.git +``` + +### 2.2 环境依赖安装 +```bash +pip3 install -r requirements.txt +``` + +## 三、Stable-Diffusion-v1.5 使用 + +### 3.1 权重及配置文件说明 +下载权重和配置文件 +```shell +# stable-diffusion-v1.5: +git clone https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5 +``` + +### 3.2 修改配置文件 +将model_index.json中所有的`diffusers`字段修改为`stablediffusion` + +### 3.3 单卡功能测试 +设置权重路径 +```shell +model_base='./stable-diffusion-v1-5' +``` +执行命令： +```shell +python3 inference_stablediffusion.py \ + --model ${model_base} \ + --prompt_file ./prompts/prompts.txt \ + --steps 50 \ + --save_dir ./results \ + --device 0 +``` +参数说明： +- --model：模型权重路径。 +- --prompt_file：提示词文件。 +- --steps: 图片生成迭代次数。 +- --save_dir：生成图片的存放目录。 +- --device：推理设备ID。 + + +### 3.4 模型推理性能 + +性能参考下列数据。 + +| 硬件形态 | 迭代次数 | 平均耗时| +| :------: |:----:|:----:| +| Atlas 800I A2(8*32G) | 50 | 2.821s | \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-1.5/inference_stablediffusion.py b/MindIE/MultiModal/StableDiffusion-1.5/inference_stablediffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..6dbc957391cc02ce148d1bbf449ef4d071fca53b --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-1.5/inference_stablediffusion.py @@ -0,0 +1,93 @@ +import torch +import torch_npu +import time +import os +import argparse +from stablediffusion import StableDiffusionPipeline + + +def parse_arguments(): + parser = argparse.ArgumentParser() + parser.add_argument( + "--prompt_file", + type=str, + default="./prompts/prompts.txt", + help="The prompts file to guide images generation.", + ) + parser.add_argument( + "--negative_prompt", + type=str, + default="", + help="The prompt or prompts to guide what to not include in image generation.", + ) + parser.add_argument( + "--steps", + type=int, + default=50, + help="The number of denoising steps.", + ) + parser.add_argument( + "--model", + type=str, + default="./stable-diffusion-v1.5", + help="The path of stable-diffusion.", + ) + parser.add_argument( + "--device", + type=int, + default=0, + help="NPU device id.", + ) + parser.add_argument( + "--save_dir", + type=str, + default="./results", + help="Path to save result audio files.", + ) + return parser.parse_args() + + +def main(): + args = parse_arguments() + save_dir = args.save_dir + if not os.path.exists(save_dir): + os.makedirs(save_dir) + + torch_npu.npu.set_device(args.device) + torch.manual_seed(1) + npu_stream = torch_npu.npu.Stream() + + pipe = StableDiffusionPipeline.from_pretrained(args.model, torch_dtype=torch.float16) + pipe.to("npu") + + total_time = 0 + prompts_num = 0 + average_time = 0 + skip = 2 + with os.fdopen(os.open(args.prompt_file, os.O_RDONLY), "r") as f: + for i, prompt in enumerate(f): + with torch.no_grad(): + npu_stream.synchronize() + begin = time.time() + image = pipe( + prompt=prompt, + negative_prompt=args.negative_prompt, + num_inference_steps=args.steps, + ).images[0] + npu_stream.synchronize() + end = time.time() + if i > skip - 1: + total_time += end - begin + prompts_num = i+1 + image_save_path = os.path.join(save_dir, f"images_{i}.png") + image.save(image_save_path) + if prompts_num > skip: + average_time = total_time / (prompts_num-skip) + else: + raise ValueError("Infer average time skip first two prompts, ensure that prompts.txt \ + contains more than three prompts") + print(f"Infer average time: {average_time:.3f}s\n") + + +if __name__ == "__main__": + main() \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-1.5/prompts/prompts.txt b/MindIE/MultiModal/StableDiffusion-1.5/prompts/prompts.txt new file mode 100644 index 0000000000000000000000000000000000000000..a375a0bb63931d0d5da6c6d91df1e14f870f47d0 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-1.5/prompts/prompts.txt @@ -0,0 +1,16 @@ +Beautiful illustration of The ocean. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Islands in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Seaports in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of The waves. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Grassland. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Wheat. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Hut Tong. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of The boat. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Pine trees. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Bamboo. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of The temple. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Cloud in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Sun in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Spring. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Lotus. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Snow piles. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-1.5/requirements.txt b/MindIE/MultiModal/StableDiffusion-1.5/requirements.txt new file mode 100644 index 0000000000000000000000000000000000000000..6d39028a97778341c84978d9643b966f25895618 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-1.5/requirements.txt @@ -0,0 +1,5 @@ +setuptools==57.5.0 +torch==2.1.0 +diffusers==0.26.3 +transformers==4.46.0 +open_clip_torch==2.20.0 \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/__init__.py b/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..529145f57160863a45c35608f5ffcb6fbbc23e99 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/__init__.py @@ -0,0 +1,5 @@ +from .pipeline import StableDiffusionPipeline, DiffusionPipeline +from .models import UNet2DConditionModel +from .models.model_utils import ModelMixin +from .schedulers import PNDMScheduler, SchedulerMixin +from .vae import AutoencoderKL \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/layers/__init__.py b/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/layers/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/layers/activations.py b/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/layers/activations.py new file mode 100644 index 0000000000000000000000000000000000000000..c045238004962da70fad9a62b8b4a25f200a1e95 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/layers/activations.py @@ -0,0 +1,123 @@ +# coding=utf-8 +# Copyright 2023 HuggingFace Inc. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import torch +import torch.nn.functional as F +from torch import nn + +from diffusers.utils import USE_PEFT_BACKEND +from diffusers.models.lora import LoRACompatibleLinear + + +ACTIVATION_FUNCTIONS = { + "swish": nn.SiLU(), + "silu": nn.SiLU(), + "mish": nn.Mish(), + "gelu": nn.GELU(), + "relu": nn.ReLU(), +} + + +def get_activation(act_fn: str) -> nn.Module: + """Helper function to get activation function from string. + + Args: + act_fn (str): Name of activation function. + + Returns: + nn.Module: Activation function. + """ + + act_fn = act_fn.lower() + if act_fn in ACTIVATION_FUNCTIONS: + return ACTIVATION_FUNCTIONS[act_fn] + else: + raise ValueError(f"Unsupported activation function: {act_fn}") + + +class GELU(nn.Module): + r""" + GELU activation function with tanh approximation support with `approximate="tanh"`. + + Parameters: + dim_in (`int`): The number of channels in the input. + dim_out (`int`): The number of channels in the output. + approximate (`str`, *optional*, defaults to `"none"`): If `"tanh"`, use tanh approximation. + bias (`bool`, defaults to True): Whether to use a bias in the linear layer. + """ + + def __init__(self, dim_in: int, dim_out: int, approximate: str = "none", bias: bool = True): + super().__init__() + self.proj = nn.Linear(dim_in, dim_out, bias=bias) + self.approximate = approximate + + def gelu(self, gate: torch.Tensor) -> torch.Tensor: + if gate.device.type != "mps": + return F.gelu(gate, approximate=self.approximate) + # mps: gelu is not implemented for float16 + return F.gelu(gate.to(dtype=torch.float32), approximate=self.approximate).to(dtype=gate.dtype) + + def forward(self, hidden_states): + hidden_states = self.proj(hidden_states) + hidden_states = self.gelu(hidden_states) + return hidden_states + + +class GEGLU(nn.Module): + r""" + A [variant](https://arxiv.org/abs/2002.05202) of the gated linear unit activation function. + + Parameters: + dim_in (`int`): The number of channels in the input. + dim_out (`int`): The number of channels in the output. + bias (`bool`, defaults to True): Whether to use a bias in the linear layer. + """ + + def __init__(self, dim_in: int, dim_out: int, bias: bool = True): + super().__init__() + linear_cls = LoRACompatibleLinear if not USE_PEFT_BACKEND else nn.Linear + + self.proj = linear_cls(dim_in, dim_out * 2, bias=bias) + + def gelu(self, gate: torch.Tensor) -> torch.Tensor: + if gate.device.type != "mps": + return F.gelu(gate) + # mps: gelu is not implemented for float16 + return F.gelu(gate.to(dtype=torch.float32)).to(dtype=gate.dtype) + + def forward(self, hidden_states, scale: float = 1.0): + args = () if USE_PEFT_BACKEND else (scale,) + hidden_states, gate = self.proj(hidden_states, *args).chunk(2, dim=-1) + return hidden_states * self.gelu(gate) + + +class ApproximateGELU(nn.Module): + r""" + The approximate form of the Gaussian Error Linear Unit (GELU). For more details, see section 2 of this + [paper](https://arxiv.org/abs/1606.08415). + + Parameters: + dim_in (`int`): The number of channels in the input. + dim_out (`int`): The number of channels in the output. + bias (`bool`, defaults to True): Whether to use a bias in the linear layer. + """ + + def __init__(self, dim_in: int, dim_out: int, bias: bool = True): + super().__init__() + self.proj = nn.Linear(dim_in, dim_out, bias=bias) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + x = self.proj(x) + return x * torch.sigmoid(1.702 * x) diff --git a/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/layers/attention.py b/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/layers/attention.py new file mode 100644 index 0000000000000000000000000000000000000000..9b2d570658f9af3e08ca74ffe65f9add1065f1a3 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/layers/attention.py @@ -0,0 +1,676 @@ +# Copyright 2023 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from typing import Any, Dict, Optional + +import torch +import torch.nn.functional as F +from torch import nn + +from diffusers.utils import USE_PEFT_BACKEND +from diffusers.utils.torch_utils import maybe_allow_in_graph +from diffusers.models.lora import LoRACompatibleLinear + +from .activations import GEGLU, GELU, ApproximateGELU +from .attention_processor import Attention +from .embeddings import SinusoidalPositionalEmbedding +from .normalization import AdaLayerNorm, AdaLayerNormContinuous, AdaLayerNormZero, RMSNorm + + +def _chunked_feed_forward( + ff: nn.Module, hidden_states: torch.Tensor, chunk_dim: int, chunk_size: int, lora_scale: Optional[float] = None +): + # "feed_forward_chunk_size" can be used to save memory + if hidden_states.shape[chunk_dim] % chunk_size != 0: + raise ValueError( + f"`hidden_states` dimension to be chunked: {hidden_states.shape[chunk_dim]} has to be divisible by chunk size: {chunk_size}. Make sure to set an appropriate `chunk_size` when calling `unet.enable_forward_chunking`." + ) + + num_chunks = hidden_states.shape[chunk_dim] // chunk_size + if lora_scale is None: + ff_output = torch.cat( + [ff(hid_slice) for hid_slice in hidden_states.chunk(num_chunks, dim=chunk_dim)], + dim=chunk_dim, + ) + else: + # TOOD(Patrick): LoRA scale can be removed once PEFT refactor is complete + ff_output = torch.cat( + [ff(hid_slice, scale=lora_scale) for hid_slice in hidden_states.chunk(num_chunks, dim=chunk_dim)], + dim=chunk_dim, + ) + + return ff_output + + +@maybe_allow_in_graph +class GatedSelfAttentionDense(nn.Module): + r""" + A gated self-attention dense layer that combines visual features and object features. + + Parameters: + query_dim (`int`): The number of channels in the query. + context_dim (`int`): The number of channels in the context. + n_heads (`int`): The number of heads to use for attention. + d_head (`int`): The number of channels in each head. + """ + + def __init__(self, query_dim: int, context_dim: int, n_heads: int, d_head: int): + super().__init__() + + # we need a linear projection since we need cat visual feature and obj feature + self.linear = nn.Linear(context_dim, query_dim) + + self.attn = Attention(query_dim=query_dim, heads=n_heads, dim_head=d_head) + self.ff = FeedForward(query_dim, activation_fn="geglu") + + self.norm1 = nn.LayerNorm(query_dim) + self.norm2 = nn.LayerNorm(query_dim) + + self.register_parameter("alpha_attn", nn.Parameter(torch.tensor(0.0))) + self.register_parameter("alpha_dense", nn.Parameter(torch.tensor(0.0))) + + self.enabled = True + + def forward(self, x: torch.Tensor, objs: torch.Tensor) -> torch.Tensor: + if not self.enabled: + return x + + n_visual = x.shape[1] + objs = self.linear(objs) + + x = x + self.alpha_attn.tanh() * self.attn(self.norm1(torch.cat([x, objs], dim=1)))[:, :n_visual, :] + x = x + self.alpha_dense.tanh() * self.ff(self.norm2(x)) + + return x + + +@maybe_allow_in_graph +class BasicTransformerBlock(nn.Module): + r""" + A basic Transformer block. + + Parameters: + dim (`int`): The number of channels in the input and output. + num_attention_heads (`int`): The number of heads to use for multi-head attention. + attention_head_dim (`int`): The number of channels in each head. + dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. + cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention. + activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward. + num_embeds_ada_norm (: + obj: `int`, *optional*): The number of diffusion steps used during training. See `Transformer2DModel`. + attention_bias (: + obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter. + only_cross_attention (`bool`, *optional*): + Whether to use only cross-attention layers. In this case two cross attention layers are used. + double_self_attention (`bool`, *optional*): + Whether to use two self-attention layers. In this case no cross attention layers are used. + upcast_attention (`bool`, *optional*): + Whether to upcast the attention computation to float32. This is useful for mixed precision training. + norm_elementwise_affine (`bool`, *optional*, defaults to `True`): + Whether to use learnable elementwise affine parameters for normalization. + norm_type (`str`, *optional*, defaults to `"layer_norm"`): + The normalization layer to use. Can be `"layer_norm"`, `"ada_norm"` or `"ada_norm_zero"`. + final_dropout (`bool` *optional*, defaults to False): + Whether to apply a final dropout after the last feed-forward layer. + attention_type (`str`, *optional*, defaults to `"default"`): + The type of attention to use. Can be `"default"` or `"gated"` or `"gated-text-image"`. + positional_embeddings (`str`, *optional*, defaults to `None`): + The type of positional embeddings to apply to. + num_positional_embeddings (`int`, *optional*, defaults to `None`): + The maximum number of positional embeddings to apply. + """ + + def __init__( + self, + dim: int, + num_attention_heads: int, + attention_head_dim: int, + dropout=0.0, + cross_attention_dim: Optional[int] = None, + activation_fn: str = "geglu", + num_embeds_ada_norm: Optional[int] = None, + attention_bias: bool = False, + only_cross_attention: bool = False, + double_self_attention: bool = False, + upcast_attention: bool = False, + norm_elementwise_affine: bool = True, + norm_type: str = "layer_norm", # 'layer_norm', 'ada_norm', 'ada_norm_zero', 'ada_norm_single', 'layer_norm_i2vgen' + norm_eps: float = 1e-5, + final_dropout: bool = False, + attention_type: str = "default", + positional_embeddings: Optional[str] = None, + num_positional_embeddings: Optional[int] = None, + ada_norm_continous_conditioning_embedding_dim: Optional[int] = None, + ada_norm_bias: Optional[int] = None, + ff_inner_dim: Optional[int] = None, + ff_bias: bool = True, + attention_out_bias: bool = True, + ): + super().__init__() + self.only_cross_attention = only_cross_attention + + self.use_ada_layer_norm_zero = (num_embeds_ada_norm is not None) and norm_type == "ada_norm_zero" + self.use_ada_layer_norm = (num_embeds_ada_norm is not None) and norm_type == "ada_norm" + self.use_ada_layer_norm_single = norm_type == "ada_norm_single" + self.use_layer_norm = norm_type == "layer_norm" + self.use_ada_layer_norm_continuous = norm_type == "ada_norm_continuous" + + if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None: + raise ValueError( + f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to" + f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}." + ) + + self.norm_type = norm_type + self.num_embeds_ada_norm = num_embeds_ada_norm + + if positional_embeddings and (num_positional_embeddings is None): + raise ValueError( + "If `positional_embedding` type is defined, `num_positition_embeddings` must also be defined." + ) + + if positional_embeddings == "sinusoidal": + self.pos_embed = SinusoidalPositionalEmbedding(dim, max_seq_length=num_positional_embeddings) + else: + self.pos_embed = None + + # Define 3 blocks. Each block has its own normalization layer. + # 1. Self-Attn + if norm_type == "ada_norm": + self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm) + elif norm_type == "ada_norm_zero": + self.norm1 = AdaLayerNormZero(dim, num_embeds_ada_norm) + elif norm_type == "ada_norm_continuous": + self.norm1 = AdaLayerNormContinuous( + dim, + ada_norm_continous_conditioning_embedding_dim, + norm_elementwise_affine, + norm_eps, + ada_norm_bias, + "rms_norm", + ) + else: + self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps) + + self.attn1 = Attention( + query_dim=dim, + heads=num_attention_heads, + dim_head=attention_head_dim, + dropout=dropout, + bias=attention_bias, + cross_attention_dim=cross_attention_dim if only_cross_attention else None, + upcast_attention=upcast_attention, + out_bias=attention_out_bias, + ) + + # 2. Cross-Attn + if cross_attention_dim is not None or double_self_attention: + # We currently only use AdaLayerNormZero for self attention where there will only be one attention block. + # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during + # the second cross attention block. + if norm_type == "ada_norm": + self.norm2 = AdaLayerNorm(dim, num_embeds_ada_norm) + elif norm_type == "ada_norm_continuous": + self.norm2 = AdaLayerNormContinuous( + dim, + ada_norm_continous_conditioning_embedding_dim, + norm_elementwise_affine, + norm_eps, + ada_norm_bias, + "rms_norm", + ) + else: + self.norm2 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine) + + self.attn2 = Attention( + query_dim=dim, + cross_attention_dim=cross_attention_dim if not double_self_attention else None, + heads=num_attention_heads, + dim_head=attention_head_dim, + dropout=dropout, + bias=attention_bias, + upcast_attention=upcast_attention, + out_bias=attention_out_bias, + ) # is self-attn if encoder_hidden_states is none + else: + self.norm2 = None + self.attn2 = None + + # 3. Feed-forward + if norm_type == "ada_norm_continuous": + self.norm3 = AdaLayerNormContinuous( + dim, + ada_norm_continous_conditioning_embedding_dim, + norm_elementwise_affine, + norm_eps, + ada_norm_bias, + "layer_norm", + ) + + elif norm_type in ["ada_norm_zero", "ada_norm", "layer_norm", "ada_norm_continuous"]: + self.norm3 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine) + elif norm_type == "layer_norm_i2vgen": + self.norm3 = None + + self.ff = FeedForward( + dim, + dropout=dropout, + activation_fn=activation_fn, + final_dropout=final_dropout, + inner_dim=ff_inner_dim, + bias=ff_bias, + ) + + # 4. Fuser + if attention_type == "gated" or attention_type == "gated-text-image": + self.fuser = GatedSelfAttentionDense(dim, cross_attention_dim, num_attention_heads, attention_head_dim) + + # 5. Scale-shift for PixArt-Alpha. + if norm_type == "ada_norm_single": + self.scale_shift_table = nn.Parameter(torch.randn(6, dim) / dim**0.5) + + # let chunk size default to None + self._chunk_size = None + self._chunk_dim = 0 + + def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int = 0): + # Sets chunk feed-forward + self._chunk_size = chunk_size + self._chunk_dim = dim + + def forward( + self, + hidden_states: torch.FloatTensor, + attention_mask: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + timestep: Optional[torch.LongTensor] = None, + cross_attention_kwargs: Dict[str, Any] = None, + class_labels: Optional[torch.LongTensor] = None, + added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None, + ) -> torch.FloatTensor: + # Notice that normalization is always applied before the real computation in the following blocks. + # 0. Self-Attention + batch_size = hidden_states.shape[0] + + if self.norm_type == "ada_norm": + norm_hidden_states = self.norm1(hidden_states, timestep) + elif self.norm_type == "ada_norm_zero": + norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1( + hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype + ) + elif self.norm_type in ["layer_norm", "layer_norm_i2vgen"]: + norm_hidden_states = self.norm1(hidden_states) + elif self.norm_type == "ada_norm_continuous": + norm_hidden_states = self.norm1(hidden_states, added_cond_kwargs["pooled_text_emb"]) + elif self.norm_type == "ada_norm_single": + shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = ( + self.scale_shift_table[None] + timestep.reshape(batch_size, 6, -1) + ).chunk(6, dim=1) + norm_hidden_states = self.norm1(hidden_states) + norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa + norm_hidden_states = norm_hidden_states.squeeze(1) + else: + raise ValueError("Incorrect norm used") + + if self.pos_embed is not None: + norm_hidden_states = self.pos_embed(norm_hidden_states) + + # 1. Retrieve lora scale. + lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0 + + # 2. Prepare GLIGEN inputs + cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {} + gligen_kwargs = cross_attention_kwargs.pop("gligen", None) + + attn_output = self.attn1( + norm_hidden_states, + encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None, + attention_mask=attention_mask, + **cross_attention_kwargs, + ) + if self.norm_type == "ada_norm_zero": + attn_output = gate_msa.unsqueeze(1) * attn_output + elif self.norm_type == "ada_norm_single": + attn_output = gate_msa * attn_output + + hidden_states = attn_output + hidden_states + if hidden_states.ndim == 4: + hidden_states = hidden_states.squeeze(1) + + # 2.5 GLIGEN Control + if gligen_kwargs is not None: + hidden_states = self.fuser(hidden_states, gligen_kwargs["objs"]) + + # 3. Cross-Attention + if self.attn2 is not None: + if self.norm_type == "ada_norm": + norm_hidden_states = self.norm2(hidden_states, timestep) + elif self.norm_type in ["ada_norm_zero", "layer_norm", "layer_norm_i2vgen"]: + norm_hidden_states = self.norm2(hidden_states) + elif self.norm_type == "ada_norm_single": + # For PixArt norm2 isn't applied here: + # https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L70C1-L76C103 + norm_hidden_states = hidden_states + elif self.norm_type == "ada_norm_continuous": + norm_hidden_states = self.norm2(hidden_states, added_cond_kwargs["pooled_text_emb"]) + else: + raise ValueError("Incorrect norm") + + if self.pos_embed is not None and self.norm_type != "ada_norm_single": + norm_hidden_states = self.pos_embed(norm_hidden_states) + + attn_output = self.attn2( + norm_hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=encoder_attention_mask, + **cross_attention_kwargs, + ) + hidden_states = attn_output + hidden_states + + # 4. Feed-forward + # i2vgen doesn't have this norm 🤷‍♂️ + if self.norm_type == "ada_norm_continuous": + norm_hidden_states = self.norm3(hidden_states, added_cond_kwargs["pooled_text_emb"]) + elif not self.norm_type == "ada_norm_single": + norm_hidden_states = self.norm3(hidden_states) + + if self.norm_type == "ada_norm_zero": + norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None] + + if self.norm_type == "ada_norm_single": + norm_hidden_states = self.norm2(hidden_states) + norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp + + if self._chunk_size is not None: + # "feed_forward_chunk_size" can be used to save memory + ff_output = _chunked_feed_forward( + self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size, lora_scale=lora_scale + ) + else: + ff_output = self.ff(norm_hidden_states, scale=lora_scale) + + if self.norm_type == "ada_norm_zero": + ff_output = gate_mlp.unsqueeze(1) * ff_output + elif self.norm_type == "ada_norm_single": + ff_output = gate_mlp * ff_output + + hidden_states = ff_output + hidden_states + if hidden_states.ndim == 4: + hidden_states = hidden_states.squeeze(1) + + return hidden_states + + +@maybe_allow_in_graph +class TemporalBasicTransformerBlock(nn.Module): + r""" + A basic Transformer block for video like data. + + Parameters: + dim (`int`): The number of channels in the input and output. + time_mix_inner_dim (`int`): The number of channels for temporal attention. + num_attention_heads (`int`): The number of heads to use for multi-head attention. + attention_head_dim (`int`): The number of channels in each head. + cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention. + """ + + def __init__( + self, + dim: int, + time_mix_inner_dim: int, + num_attention_heads: int, + attention_head_dim: int, + cross_attention_dim: Optional[int] = None, + ): + super().__init__() + self.is_res = dim == time_mix_inner_dim + + self.norm_in = nn.LayerNorm(dim) + + # Define 3 blocks. Each block has its own normalization layer. + # 1. Self-Attn + self.norm_in = nn.LayerNorm(dim) + self.ff_in = FeedForward( + dim, + dim_out=time_mix_inner_dim, + activation_fn="geglu", + ) + + self.norm1 = nn.LayerNorm(time_mix_inner_dim) + self.attn1 = Attention( + query_dim=time_mix_inner_dim, + heads=num_attention_heads, + dim_head=attention_head_dim, + cross_attention_dim=None, + ) + + # 2. Cross-Attn + if cross_attention_dim is not None: + # We currently only use AdaLayerNormZero for self attention where there will only be one attention block. + # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during + # the second cross attention block. + self.norm2 = nn.LayerNorm(time_mix_inner_dim) + self.attn2 = Attention( + query_dim=time_mix_inner_dim, + cross_attention_dim=cross_attention_dim, + heads=num_attention_heads, + dim_head=attention_head_dim, + ) # is self-attn if encoder_hidden_states is none + else: + self.norm2 = None + self.attn2 = None + + # 3. Feed-forward + self.norm3 = nn.LayerNorm(time_mix_inner_dim) + self.ff = FeedForward(time_mix_inner_dim, activation_fn="geglu") + + # let chunk size default to None + self._chunk_size = None + self._chunk_dim = None + + def set_chunk_feed_forward(self, chunk_size: Optional[int], **kwargs): + # Sets chunk feed-forward + self._chunk_size = chunk_size + # chunk dim should be hardcoded to 1 to have better speed vs. memory trade-off + self._chunk_dim = 1 + + def forward( + self, + hidden_states: torch.FloatTensor, + num_frames: int, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + ) -> torch.FloatTensor: + # Notice that normalization is always applied before the real computation in the following blocks. + # 0. Self-Attention + batch_size = hidden_states.shape[0] + + batch_frames, seq_length, channels = hidden_states.shape + batch_size = batch_frames // num_frames + + hidden_states = hidden_states[None, :].reshape(batch_size, num_frames, seq_length, channels) + hidden_states = hidden_states.permute(0, 2, 1, 3) + hidden_states = hidden_states.reshape(batch_size * seq_length, num_frames, channels) + + residual = hidden_states + hidden_states = self.norm_in(hidden_states) + + if self._chunk_size is not None: + hidden_states = _chunked_feed_forward(self.ff_in, hidden_states, self._chunk_dim, self._chunk_size) + else: + hidden_states = self.ff_in(hidden_states) + + if self.is_res: + hidden_states = hidden_states + residual + + norm_hidden_states = self.norm1(hidden_states) + attn_output = self.attn1(norm_hidden_states, encoder_hidden_states=None) + hidden_states = attn_output + hidden_states + + # 3. Cross-Attention + if self.attn2 is not None: + norm_hidden_states = self.norm2(hidden_states) + attn_output = self.attn2(norm_hidden_states, encoder_hidden_states=encoder_hidden_states) + hidden_states = attn_output + hidden_states + + # 4. Feed-forward + norm_hidden_states = self.norm3(hidden_states) + + if self._chunk_size is not None: + ff_output = _chunked_feed_forward(self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size) + else: + ff_output = self.ff(norm_hidden_states) + + if self.is_res: + hidden_states = ff_output + hidden_states + else: + hidden_states = ff_output + + hidden_states = hidden_states[None, :].reshape(batch_size, seq_length, num_frames, channels) + hidden_states = hidden_states.permute(0, 2, 1, 3) + hidden_states = hidden_states.reshape(batch_size * num_frames, seq_length, channels) + + return hidden_states + + +class SkipFFTransformerBlock(nn.Module): + def __init__( + self, + dim: int, + num_attention_heads: int, + attention_head_dim: int, + kv_input_dim: int, + kv_input_dim_proj_use_bias: bool, + dropout=0.0, + cross_attention_dim: Optional[int] = None, + attention_bias: bool = False, + attention_out_bias: bool = True, + ): + super().__init__() + if kv_input_dim != dim: + self.kv_mapper = nn.Linear(kv_input_dim, dim, kv_input_dim_proj_use_bias) + else: + self.kv_mapper = None + + self.norm1 = RMSNorm(dim, 1e-06) + + self.attn1 = Attention( + query_dim=dim, + heads=num_attention_heads, + dim_head=attention_head_dim, + dropout=dropout, + bias=attention_bias, + cross_attention_dim=cross_attention_dim, + out_bias=attention_out_bias, + ) + + self.norm2 = RMSNorm(dim, 1e-06) + + self.attn2 = Attention( + query_dim=dim, + cross_attention_dim=cross_attention_dim, + heads=num_attention_heads, + dim_head=attention_head_dim, + dropout=dropout, + bias=attention_bias, + out_bias=attention_out_bias, + ) + + def forward(self, hidden_states, encoder_hidden_states, cross_attention_kwargs): + cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {} + + if self.kv_mapper is not None: + encoder_hidden_states = self.kv_mapper(F.silu(encoder_hidden_states)) + + norm_hidden_states = self.norm1(hidden_states) + + attn_output = self.attn1( + norm_hidden_states, + encoder_hidden_states=encoder_hidden_states, + **cross_attention_kwargs, + ) + + hidden_states = attn_output + hidden_states + + norm_hidden_states = self.norm2(hidden_states) + + attn_output = self.attn2( + norm_hidden_states, + encoder_hidden_states=encoder_hidden_states, + **cross_attention_kwargs, + ) + + hidden_states = attn_output + hidden_states + + return hidden_states + + +class FeedForward(nn.Module): + r""" + A feed-forward layer. + + Parameters: + dim (`int`): The number of channels in the input. + dim_out (`int`, *optional*): The number of channels in the output. If not given, defaults to `dim`. + mult (`int`, *optional*, defaults to 4): The multiplier to use for the hidden dimension. + dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. + activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward. + final_dropout (`bool` *optional*, defaults to False): Apply a final dropout. + bias (`bool`, defaults to True): Whether to use a bias in the linear layer. + """ + + def __init__( + self, + dim: int, + dim_out: Optional[int] = None, + mult: int = 4, + dropout: float = 0.0, + activation_fn: str = "geglu", + final_dropout: bool = False, + inner_dim=None, + bias: bool = True, + ): + super().__init__() + if inner_dim is None: + inner_dim = int(dim * mult) + dim_out = dim_out if dim_out is not None else dim + linear_cls = LoRACompatibleLinear if not USE_PEFT_BACKEND else nn.Linear + + if activation_fn == "gelu": + act_fn = GELU(dim, inner_dim, bias=bias) + if activation_fn == "gelu-approximate": + act_fn = GELU(dim, inner_dim, approximate="tanh", bias=bias) + elif activation_fn == "geglu": + act_fn = GEGLU(dim, inner_dim, bias=bias) + elif activation_fn == "geglu-approximate": + act_fn = ApproximateGELU(dim, inner_dim, bias=bias) + + self.net = nn.ModuleList([]) + # project in + self.net.append(act_fn) + # project dropout + self.net.append(nn.Dropout(dropout)) + # project out + self.net.append(linear_cls(inner_dim, dim_out, bias=bias)) + # FF as used in Vision Transformer, MLP-Mixer, etc. have a final dropout + if final_dropout: + self.net.append(nn.Dropout(dropout)) + + def forward(self, hidden_states: torch.Tensor, scale: float = 1.0) -> torch.Tensor: + compatible_cls = (GEGLU,) if USE_PEFT_BACKEND else (GEGLU, LoRACompatibleLinear) + for module in self.net: + if isinstance(module, compatible_cls): + hidden_states = module(hidden_states, scale) + else: + hidden_states = module(hidden_states) + return hidden_states diff --git a/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/layers/attention_processor.py b/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/layers/attention_processor.py new file mode 100644 index 0000000000000000000000000000000000000000..ffc4fbfb232b83b85ae946ce2544eab02efb94b2 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/layers/attention_processor.py @@ -0,0 +1,1019 @@ +# Copyright 2023 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from importlib import import_module +from typing import Callable, Optional, Union + +import torch +import torch_npu +import torch.nn.functional as F +from torch import nn + +from diffusers.utils import USE_PEFT_BACKEND, deprecate, logging +from diffusers.utils.import_utils import is_xformers_available +from diffusers.models.attention_processor import ( + CustomDiffusionAttnProcessor, FusedAttnProcessor2_0, CustomDiffusionXFormersAttnProcessor, CustomDiffusionAttnProcessor2_0, + SpatialNorm, SlicedAttnProcessor, SlicedAttnAddedKVProcessor, LoRAAttnProcessor, LoRAAttnProcessor2_0, + LoRAXFormersAttnProcessor, LoRAAttnAddedKVProcessor, IPAdapterAttnProcessor, IPAdapterAttnProcessor2_0, + CustomDiffusionAttnProcessor, XFormersAttnAddedKVProcessor, XFormersAttnProcessor, +) +from diffusers.models.lora import LoRACompatibleLinear, LoRALinearLayer + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +class Attention(nn.Module): + r""" + A cross attention layer. + + Parameters: + query_dim (`int`): + The number of channels in the query. + cross_attention_dim (`int`, *optional*): + The number of channels in the encoder_hidden_states. If not given, defaults to `query_dim`. + heads (`int`, *optional*, defaults to 8): + The number of heads to use for multi-head attention. + dim_head (`int`, *optional*, defaults to 64): + The number of channels in each head. + dropout (`float`, *optional*, defaults to 0.0): + The dropout probability to use. + bias (`bool`, *optional*, defaults to False): + Set to `True` for the query, key, and value linear layers to contain a bias parameter. + upcast_attention (`bool`, *optional*, defaults to False): + Set to `True` to upcast the attention computation to `float32`. + upcast_softmax (`bool`, *optional*, defaults to False): + Set to `True` to upcast the softmax computation to `float32`. + cross_attention_norm (`str`, *optional*, defaults to `None`): + The type of normalization to use for the cross attention. Can be `None`, `layer_norm`, or `group_norm`. + cross_attention_norm_num_groups (`int`, *optional*, defaults to 32): + The number of groups to use for the group norm in the cross attention. + added_kv_proj_dim (`int`, *optional*, defaults to `None`): + The number of channels to use for the added key and value projections. If `None`, no projection is used. + norm_num_groups (`int`, *optional*, defaults to `None`): + The number of groups to use for the group norm in the attention. + spatial_norm_dim (`int`, *optional*, defaults to `None`): + The number of channels to use for the spatial normalization. + out_bias (`bool`, *optional*, defaults to `True`): + Set to `True` to use a bias in the output linear layer. + scale_qk (`bool`, *optional*, defaults to `True`): + Set to `True` to scale the query and key by `1 / sqrt(dim_head)`. + only_cross_attention (`bool`, *optional*, defaults to `False`): + Set to `True` to only use cross attention and not added_kv_proj_dim. Can only be set to `True` if + `added_kv_proj_dim` is not `None`. + eps (`float`, *optional*, defaults to 1e-5): + An additional value added to the denominator in group normalization that is used for numerical stability. + rescale_output_factor (`float`, *optional*, defaults to 1.0): + A factor to rescale the output by dividing it with this value. + residual_connection (`bool`, *optional*, defaults to `False`): + Set to `True` to add the residual connection to the output. + _from_deprecated_attn_block (`bool`, *optional*, defaults to `False`): + Set to `True` if the attention block is loaded from a deprecated state dict. + processor (`AttnProcessor`, *optional*, defaults to `None`): + The attention processor to use. If `None`, defaults to `AttnProcessor2_0` if `torch 2.x` is used and + `AttnProcessor` otherwise. + """ + + def __init__( + self, + query_dim: int, + cross_attention_dim: Optional[int] = None, + heads: int = 8, + dim_head: int = 64, + dropout: float = 0.0, + bias: bool = False, + upcast_attention: bool = False, + upcast_softmax: bool = False, + cross_attention_norm: Optional[str] = None, + cross_attention_norm_num_groups: int = 32, + added_kv_proj_dim: Optional[int] = None, + norm_num_groups: Optional[int] = None, + spatial_norm_dim: Optional[int] = None, + out_bias: bool = True, + scale_qk: bool = True, + only_cross_attention: bool = False, + eps: float = 1e-5, + rescale_output_factor: float = 1.0, + residual_connection: bool = False, + _from_deprecated_attn_block: bool = False, + processor: Optional["AttnProcessor"] = None, + out_dim: int = None, + ): + super().__init__() + self.inner_dim = out_dim if out_dim is not None else dim_head * heads + self.query_dim = query_dim + self.cross_attention_dim = cross_attention_dim if cross_attention_dim is not None else query_dim + self.upcast_attention = upcast_attention + self.upcast_softmax = upcast_softmax + self.rescale_output_factor = rescale_output_factor + self.residual_connection = residual_connection + self.dropout = dropout + self.fused_projections = False + self.out_dim = out_dim if out_dim is not None else query_dim + + # we make use of this private variable to know whether this class is loaded + # with an deprecated state dict so that we can convert it on the fly + self._from_deprecated_attn_block = _from_deprecated_attn_block + + self.scale_qk = scale_qk + self.scale = dim_head**-0.5 if self.scale_qk else 1.0 + + self.heads = out_dim // dim_head if out_dim is not None else heads + # for slice_size > 0 the attention score computation + # is split across the batch axis to save memory + # You can set slice_size with `set_attention_slice` + self.sliceable_head_dim = heads + + self.added_kv_proj_dim = added_kv_proj_dim + self.only_cross_attention = only_cross_attention + + if self.added_kv_proj_dim is None and self.only_cross_attention: + raise ValueError( + "`only_cross_attention` can only be set to True if `added_kv_proj_dim` is not None. Make sure to set either `only_cross_attention=False` or define `added_kv_proj_dim`." + ) + + if norm_num_groups is not None: + self.group_norm = nn.GroupNorm(num_channels=query_dim, num_groups=norm_num_groups, eps=eps, affine=True) + else: + self.group_norm = None + + if spatial_norm_dim is not None: + self.spatial_norm = SpatialNorm(f_channels=query_dim, zq_channels=spatial_norm_dim) + else: + self.spatial_norm = None + + if cross_attention_norm is None: + self.norm_cross = None + elif cross_attention_norm == "layer_norm": + self.norm_cross = nn.LayerNorm(self.cross_attention_dim) + elif cross_attention_norm == "group_norm": + if self.added_kv_proj_dim is not None: + # The given `encoder_hidden_states` are initially of shape + # (batch_size, seq_len, added_kv_proj_dim) before being projected + # to (batch_size, seq_len, cross_attention_dim). The norm is applied + # before the projection, so we need to use `added_kv_proj_dim` as + # the number of channels for the group norm. + norm_cross_num_channels = added_kv_proj_dim + else: + norm_cross_num_channels = self.cross_attention_dim + + self.norm_cross = nn.GroupNorm( + num_channels=norm_cross_num_channels, num_groups=cross_attention_norm_num_groups, eps=1e-5, affine=True + ) + else: + raise ValueError( + f"unknown cross_attention_norm: {cross_attention_norm}. Should be None, 'layer_norm' or 'group_norm'" + ) + + if USE_PEFT_BACKEND: + linear_cls = nn.Linear + else: + linear_cls = LoRACompatibleLinear + + self.linear_cls = linear_cls + self.to_q = linear_cls(query_dim, self.inner_dim, bias=bias) + + if not self.only_cross_attention: + # only relevant for the `AddedKVProcessor` classes + self.to_k = linear_cls(self.cross_attention_dim, self.inner_dim, bias=bias) + self.to_v = linear_cls(self.cross_attention_dim, self.inner_dim, bias=bias) + else: + self.to_k = None + self.to_v = None + + if self.added_kv_proj_dim is not None: + self.add_k_proj = linear_cls(added_kv_proj_dim, self.inner_dim) + self.add_v_proj = linear_cls(added_kv_proj_dim, self.inner_dim) + + self.to_out = nn.ModuleList([]) + self.to_out.append(linear_cls(self.inner_dim, self.out_dim, bias=out_bias)) + self.to_out.append(nn.Dropout(dropout)) + + # set attention processor + # We use the AttnProcessor2_0 by default when torch 2.x is used which uses + # torch.nn.functional.scaled_dot_product_attention for native Flash/memory_efficient_attention + # but only if it has the default `scale` argument. TODO remove scale_qk check when we move to torch 2.1 + if processor is None: + processor = ( + AttnProcessor2_0() if hasattr(F, "scaled_dot_product_attention") and self.scale_qk else AttnProcessor() + ) + self.set_processor(processor) + + def set_use_memory_efficient_attention_xformers( + self, use_memory_efficient_attention_xformers: bool, attention_op: Optional[Callable] = None + ) -> None: + r""" + Set whether to use memory efficient attention from `xformers` or not. + + Args: + use_memory_efficient_attention_xformers (`bool`): + Whether to use memory efficient attention from `xformers` or not. + attention_op (`Callable`, *optional*): + The attention operation to use. Defaults to `None` which uses the default attention operation from + `xformers`. + """ + is_lora = hasattr(self, "processor") and isinstance( + self.processor, + LORA_ATTENTION_PROCESSORS, + ) + is_custom_diffusion = hasattr(self, "processor") and isinstance( + self.processor, + (CustomDiffusionAttnProcessor, CustomDiffusionXFormersAttnProcessor, CustomDiffusionAttnProcessor2_0), + ) + is_added_kv_processor = hasattr(self, "processor") and isinstance( + self.processor, + ( + AttnAddedKVProcessor, + AttnAddedKVProcessor2_0, + SlicedAttnAddedKVProcessor, + XFormersAttnAddedKVProcessor, + LoRAAttnAddedKVProcessor, + ), + ) + + if use_memory_efficient_attention_xformers: + if is_added_kv_processor and (is_lora or is_custom_diffusion): + raise NotImplementedError( + f"Memory efficient attention is currently not supported for LoRA or custom diffusion for attention processor type {self.processor}" + ) + if not is_xformers_available(): + raise ModuleNotFoundError( + ( + "Refer to https://github.com/facebookresearch/xformers for more information on how to install" + " xformers" + ), + name="xformers", + ) + elif not torch.cuda.is_available(): + raise ValueError( + "torch.cuda.is_available() should be True but is False. xformers' memory efficient attention is" + " only available for GPU " + ) + else: + try: + # Make sure we can run the memory efficient attention + _ = xformers.ops.memory_efficient_attention( + torch.randn((1, 2, 40), device="cuda"), + torch.randn((1, 2, 40), device="cuda"), + torch.randn((1, 2, 40), device="cuda"), + ) + except Exception as e: + raise e + + if is_lora: + # TODO (sayakpaul): should we throw a warning if someone wants to use the xformers + # variant when using PT 2.0 now that we have LoRAAttnProcessor2_0? + processor = LoRAXFormersAttnProcessor( + hidden_size=self.processor.hidden_size, + cross_attention_dim=self.processor.cross_attention_dim, + rank=self.processor.rank, + attention_op=attention_op, + ) + processor.load_state_dict(self.processor.state_dict()) + processor.to(self.processor.to_q_lora.up.weight.device) + elif is_custom_diffusion: + processor = CustomDiffusionXFormersAttnProcessor( + train_kv=self.processor.train_kv, + train_q_out=self.processor.train_q_out, + hidden_size=self.processor.hidden_size, + cross_attention_dim=self.processor.cross_attention_dim, + attention_op=attention_op, + ) + processor.load_state_dict(self.processor.state_dict()) + if hasattr(self.processor, "to_k_custom_diffusion"): + processor.to(self.processor.to_k_custom_diffusion.weight.device) + elif is_added_kv_processor: + # TODO(Patrick, Suraj, William) - currently xformers doesn't work for UnCLIP + # which uses this type of cross attention ONLY because the attention mask of format + # [0, ..., -10.000, ..., 0, ...,] is not supported + # throw warning + logger.info( + "Memory efficient attention with `xformers` might currently not work correctly if an attention mask is required for the attention operation." + ) + processor = XFormersAttnAddedKVProcessor(attention_op=attention_op) + else: + processor = XFormersAttnProcessor(attention_op=attention_op) + else: + if is_lora: + attn_processor_class = ( + LoRAAttnProcessor2_0 if hasattr(F, "scaled_dot_product_attention") else LoRAAttnProcessor + ) + processor = attn_processor_class( + hidden_size=self.processor.hidden_size, + cross_attention_dim=self.processor.cross_attention_dim, + rank=self.processor.rank, + ) + processor.load_state_dict(self.processor.state_dict()) + processor.to(self.processor.to_q_lora.up.weight.device) + elif is_custom_diffusion: + attn_processor_class = ( + CustomDiffusionAttnProcessor2_0 + if hasattr(F, "scaled_dot_product_attention") + else CustomDiffusionAttnProcessor + ) + processor = attn_processor_class( + train_kv=self.processor.train_kv, + train_q_out=self.processor.train_q_out, + hidden_size=self.processor.hidden_size, + cross_attention_dim=self.processor.cross_attention_dim, + ) + processor.load_state_dict(self.processor.state_dict()) + if hasattr(self.processor, "to_k_custom_diffusion"): + processor.to(self.processor.to_k_custom_diffusion.weight.device) + else: + # set attention processor + # We use the AttnProcessor2_0 by default when torch 2.x is used which uses + # torch.nn.functional.scaled_dot_product_attention for native Flash/memory_efficient_attention + # but only if it has the default `scale` argument. TODO remove scale_qk check when we move to torch 2.1 + processor = ( + AttnProcessor2_0() + if hasattr(F, "scaled_dot_product_attention") and self.scale_qk + else AttnProcessor() + ) + + self.set_processor(processor) + + def set_attention_slice(self, slice_size: int) -> None: + r""" + Set the slice size for attention computation. + + Args: + slice_size (`int`): + The slice size for attention computation. + """ + if slice_size is not None and slice_size > self.sliceable_head_dim: + raise ValueError(f"slice_size {slice_size} has to be smaller or equal to {self.sliceable_head_dim}.") + + if slice_size is not None and self.added_kv_proj_dim is not None: + processor = SlicedAttnAddedKVProcessor(slice_size) + elif slice_size is not None: + processor = SlicedAttnProcessor(slice_size) + elif self.added_kv_proj_dim is not None: + processor = AttnAddedKVProcessor() + else: + # set attention processor + # We use the AttnProcessor2_0 by default when torch 2.x is used which uses + # torch.nn.functional.scaled_dot_product_attention for native Flash/memory_efficient_attention + # but only if it has the default `scale` argument. TODO remove scale_qk check when we move to torch 2.1 + processor = ( + AttnProcessor2_0() if hasattr(F, "scaled_dot_product_attention") and self.scale_qk else AttnProcessor() + ) + + self.set_processor(processor) + + def set_processor(self, processor: "AttnProcessor") -> None: + r""" + Set the attention processor to use. + + Args: + processor (`AttnProcessor`): + The attention processor to use. + """ + # if current processor is in `self._modules` and if passed `processor` is not, we need to + # pop `processor` from `self._modules` + if ( + hasattr(self, "processor") + and isinstance(self.processor, torch.nn.Module) + and not isinstance(processor, torch.nn.Module) + ): + logger.info(f"You are removing possibly trained weights of {self.processor} with {processor}") + self._modules.pop("processor") + + self.processor = processor + + def get_processor(self, return_deprecated_lora: bool = False) -> "AttentionProcessor": + r""" + Get the attention processor in use. + + Args: + return_deprecated_lora (`bool`, *optional*, defaults to `False`): + Set to `True` to return the deprecated LoRA attention processor. + + Returns: + "AttentionProcessor": The attention processor in use. + """ + if not return_deprecated_lora: + return self.processor + + # TODO(Sayak, Patrick). The rest of the function is needed to ensure backwards compatible + # serialization format for LoRA Attention Processors. It should be deleted once the integration + # with PEFT is completed. + is_lora_activated = { + name: module.lora_layer is not None + for name, module in self.named_modules() + if hasattr(module, "lora_layer") + } + + # 1. if no layer has a LoRA activated we can return the processor as usual + if not any(is_lora_activated.values()): + return self.processor + + # If doesn't apply LoRA do `add_k_proj` or `add_v_proj` + is_lora_activated.pop("add_k_proj", None) + is_lora_activated.pop("add_v_proj", None) + # 2. else it is not posssible that only some layers have LoRA activated + if not all(is_lora_activated.values()): + raise ValueError( + f"Make sure that either all layers or no layers have LoRA activated, but have {is_lora_activated}" + ) + + # 3. And we need to merge the current LoRA layers into the corresponding LoRA attention processor + non_lora_processor_cls_name = self.processor.__class__.__name__ + lora_processor_cls = getattr(import_module(__name__), "LoRA" + non_lora_processor_cls_name) + + hidden_size = self.inner_dim + + # now create a LoRA attention processor from the LoRA layers + if lora_processor_cls in [LoRAAttnProcessor, LoRAAttnProcessor2_0, LoRAXFormersAttnProcessor]: + kwargs = { + "cross_attention_dim": self.cross_attention_dim, + "rank": self.to_q.lora_layer.rank, + "network_alpha": self.to_q.lora_layer.network_alpha, + "q_rank": self.to_q.lora_layer.rank, + "q_hidden_size": self.to_q.lora_layer.out_features, + "k_rank": self.to_k.lora_layer.rank, + "k_hidden_size": self.to_k.lora_layer.out_features, + "v_rank": self.to_v.lora_layer.rank, + "v_hidden_size": self.to_v.lora_layer.out_features, + "out_rank": self.to_out[0].lora_layer.rank, + "out_hidden_size": self.to_out[0].lora_layer.out_features, + } + + if hasattr(self.processor, "attention_op"): + kwargs["attention_op"] = self.processor.attention_op + + lora_processor = lora_processor_cls(hidden_size, **kwargs) + lora_processor.to_q_lora.load_state_dict(self.to_q.lora_layer.state_dict()) + lora_processor.to_k_lora.load_state_dict(self.to_k.lora_layer.state_dict()) + lora_processor.to_v_lora.load_state_dict(self.to_v.lora_layer.state_dict()) + lora_processor.to_out_lora.load_state_dict(self.to_out[0].lora_layer.state_dict()) + elif lora_processor_cls == LoRAAttnAddedKVProcessor: + lora_processor = lora_processor_cls( + hidden_size, + cross_attention_dim=self.add_k_proj.weight.shape[0], + rank=self.to_q.lora_layer.rank, + network_alpha=self.to_q.lora_layer.network_alpha, + ) + lora_processor.to_q_lora.load_state_dict(self.to_q.lora_layer.state_dict()) + lora_processor.to_k_lora.load_state_dict(self.to_k.lora_layer.state_dict()) + lora_processor.to_v_lora.load_state_dict(self.to_v.lora_layer.state_dict()) + lora_processor.to_out_lora.load_state_dict(self.to_out[0].lora_layer.state_dict()) + + # only save if used + if self.add_k_proj.lora_layer is not None: + lora_processor.add_k_proj_lora.load_state_dict(self.add_k_proj.lora_layer.state_dict()) + lora_processor.add_v_proj_lora.load_state_dict(self.add_v_proj.lora_layer.state_dict()) + else: + lora_processor.add_k_proj_lora = None + lora_processor.add_v_proj_lora = None + else: + raise ValueError(f"{lora_processor_cls} does not exist.") + + return lora_processor + + def forward( + self, + hidden_states: torch.FloatTensor, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + **cross_attention_kwargs, + ) -> torch.Tensor: + r""" + The forward method of the `Attention` class. + + Args: + hidden_states (`torch.Tensor`): + The hidden states of the query. + encoder_hidden_states (`torch.Tensor`, *optional*): + The hidden states of the encoder. + attention_mask (`torch.Tensor`, *optional*): + The attention mask to use. If `None`, no mask is applied. + **cross_attention_kwargs: + Additional keyword arguments to pass along to the cross attention. + + Returns: + `torch.Tensor`: The output of the attention layer. + """ + # The `Attention` class can call different attention processors / attention functions + # here we simply pass along all tensors to the selected processor class + # For standard processors that are defined here, `**cross_attention_kwargs` is empty + return self.processor( + self, + hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=attention_mask, + **cross_attention_kwargs, + ) + + def batch_to_head_dim(self, tensor: torch.Tensor) -> torch.Tensor: + r""" + Reshape the tensor from `[batch_size, seq_len, dim]` to `[batch_size // heads, seq_len, dim * heads]`. `heads` + is the number of heads initialized while constructing the `Attention` class. + + Args: + tensor (`torch.Tensor`): The tensor to reshape. + + Returns: + `torch.Tensor`: The reshaped tensor. + """ + head_size = self.heads + batch_size, seq_len, dim = tensor.shape + tensor = tensor.reshape(batch_size // head_size, head_size, seq_len, dim) + tensor = tensor.permute(0, 2, 1, 3).reshape(batch_size // head_size, seq_len, dim * head_size) + return tensor + + def head_to_batch_dim(self, tensor: torch.Tensor, out_dim: int = 3) -> torch.Tensor: + r""" + Reshape the tensor from `[batch_size, seq_len, dim]` to `[batch_size, seq_len, heads, dim // heads]` `heads` is + the number of heads initialized while constructing the `Attention` class. + + Args: + tensor (`torch.Tensor`): The tensor to reshape. + out_dim (`int`, *optional*, defaults to `3`): The output dimension of the tensor. If `3`, the tensor is + reshaped to `[batch_size * heads, seq_len, dim // heads]`. + + Returns: + `torch.Tensor`: The reshaped tensor. + """ + head_size = self.heads + batch_size, seq_len, dim = tensor.shape + tensor = tensor.reshape(batch_size, seq_len, head_size, dim // head_size) + tensor = tensor.permute(0, 2, 1, 3) + + if out_dim == 3: + tensor = tensor.reshape(batch_size * head_size, seq_len, dim // head_size) + + return tensor + + def get_attention_scores( + self, query: torch.Tensor, key: torch.Tensor, attention_mask: torch.Tensor = None + ) -> torch.Tensor: + r""" + Compute the attention scores. + + Args: + query (`torch.Tensor`): The query tensor. + key (`torch.Tensor`): The key tensor. + attention_mask (`torch.Tensor`, *optional*): The attention mask to use. If `None`, no mask is applied. + + Returns: + `torch.Tensor`: The attention probabilities/scores. + """ + dtype = query.dtype + if self.upcast_attention: + query = query.float() + key = key.float() + + if attention_mask is None: + baddbmm_input = torch.empty( + query.shape[0], query.shape[1], key.shape[1], dtype=query.dtype, device=query.device + ) + beta = 0 + else: + baddbmm_input = attention_mask + beta = 1 + + attention_scores = torch.baddbmm( + baddbmm_input, + query, + key.transpose(-1, -2), + beta=beta, + alpha=self.scale, + ) + del baddbmm_input + + if self.upcast_softmax: + attention_scores = attention_scores.float() + + attention_probs = attention_scores.softmax(dim=-1) + del attention_scores + + attention_probs = attention_probs.to(dtype) + + return attention_probs + + def prepare_attention_mask( + self, attention_mask: torch.Tensor, target_length: int, batch_size: int, out_dim: int = 3 + ) -> torch.Tensor: + r""" + Prepare the attention mask for the attention computation. + + Args: + attention_mask (`torch.Tensor`): + The attention mask to prepare. + target_length (`int`): + The target length of the attention mask. This is the length of the attention mask after padding. + batch_size (`int`): + The batch size, which is used to repeat the attention mask. + out_dim (`int`, *optional*, defaults to `3`): + The output dimension of the attention mask. Can be either `3` or `4`. + + Returns: + `torch.Tensor`: The prepared attention mask. + """ + head_size = self.heads + if attention_mask is None: + return attention_mask + + current_length: int = attention_mask.shape[-1] + if current_length != target_length: + if attention_mask.device.type == "mps": + # HACK: MPS: Does not support padding by greater than dimension of input tensor. + # Instead, we can manually construct the padding tensor. + padding_shape = (attention_mask.shape[0], attention_mask.shape[1], target_length) + padding = torch.zeros(padding_shape, dtype=attention_mask.dtype, device=attention_mask.device) + attention_mask = torch.cat([attention_mask, padding], dim=2) + else: + # TODO: for pipelines such as stable-diffusion, padding cross-attn mask: + # we want to instead pad by (0, remaining_length), where remaining_length is: + # remaining_length: int = target_length - current_length + # TODO: re-enable tests/models/test_models_unet_2d_condition.py#test_model_xattn_padding + attention_mask = F.pad(attention_mask, (0, target_length), value=0.0) + + if out_dim == 3: + if attention_mask.shape[0] < batch_size * head_size: + attention_mask = attention_mask.repeat_interleave(head_size, dim=0) + elif out_dim == 4: + attention_mask = attention_mask.unsqueeze(1) + attention_mask = attention_mask.repeat_interleave(head_size, dim=1) + + return attention_mask + + def norm_encoder_hidden_states(self, encoder_hidden_states: torch.Tensor) -> torch.Tensor: + r""" + Normalize the encoder hidden states. Requires `self.norm_cross` to be specified when constructing the + `Attention` class. + + Args: + encoder_hidden_states (`torch.Tensor`): Hidden states of the encoder. + + Returns: + `torch.Tensor`: The normalized encoder hidden states. + """ + assert self.norm_cross is not None, "self.norm_cross must be defined to call self.norm_encoder_hidden_states" + + if isinstance(self.norm_cross, nn.LayerNorm): + encoder_hidden_states = self.norm_cross(encoder_hidden_states) + elif isinstance(self.norm_cross, nn.GroupNorm): + # Group norm norms along the channels dimension and expects + # input to be in the shape of (N, C, *). In this case, we want + # to norm along the hidden dimension, so we need to move + # (batch_size, sequence_length, hidden_size) -> + # (batch_size, hidden_size, sequence_length) + encoder_hidden_states = encoder_hidden_states.transpose(1, 2) + encoder_hidden_states = self.norm_cross(encoder_hidden_states) + encoder_hidden_states = encoder_hidden_states.transpose(1, 2) + else: + assert False + + return encoder_hidden_states + + @torch.no_grad() + def fuse_projections(self, fuse=True): + is_cross_attention = self.cross_attention_dim != self.query_dim + device = self.to_q.weight.data.device + dtype = self.to_q.weight.data.dtype + + if not is_cross_attention: + # fetch weight matrices. + concatenated_weights = torch.cat([self.to_q.weight.data, self.to_k.weight.data, self.to_v.weight.data]) + in_features = concatenated_weights.shape[1] + out_features = concatenated_weights.shape[0] + + # create a new single projection layer and copy over the weights. + self.to_qkv = self.linear_cls(in_features, out_features, bias=False, device=device, dtype=dtype) + self.to_qkv.weight.copy_(concatenated_weights) + + else: + concatenated_weights = torch.cat([self.to_k.weight.data, self.to_v.weight.data]) + in_features = concatenated_weights.shape[1] + out_features = concatenated_weights.shape[0] + + self.to_kv = self.linear_cls(in_features, out_features, bias=False, device=device, dtype=dtype) + self.to_kv.weight.copy_(concatenated_weights) + + self.fused_projections = fuse + + +class AttnProcessor: + r""" + Default processor for performing attention-related computations. + """ + + def __call__( + self, + attn: Attention, + hidden_states: torch.FloatTensor, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + temb: Optional[torch.FloatTensor] = None, + scale: float = 1.0, + ) -> torch.Tensor: + residual = hidden_states + + args = () if USE_PEFT_BACKEND else (scale,) + + if attn.spatial_norm is not None: + hidden_states = attn.spatial_norm(hidden_states, temb) + + input_ndim = hidden_states.ndim + + if input_ndim == 4: + batch_size, channel, height, width = hidden_states.shape + hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2) + + batch_size, sequence_length, _ = ( + hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape + ) + attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) + + if attn.group_norm is not None: + hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2) + + query = attn.to_q(hidden_states, *args) + + if encoder_hidden_states is None: + encoder_hidden_states = hidden_states + elif attn.norm_cross: + encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states) + + key = attn.to_k(encoder_hidden_states, *args) + value = attn.to_v(encoder_hidden_states, *args) + + query = attn.head_to_batch_dim(query) + key = attn.head_to_batch_dim(key) + value = attn.head_to_batch_dim(value) + + attention_probs = attn.get_attention_scores(query, key, attention_mask) + hidden_states = torch.bmm(attention_probs, value) + hidden_states = attn.batch_to_head_dim(hidden_states) + + # linear proj + hidden_states = attn.to_out[0](hidden_states, *args) + # dropout + hidden_states = attn.to_out[1](hidden_states) + + if input_ndim == 4: + hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width) + + if attn.residual_connection: + hidden_states = hidden_states + residual + + hidden_states = hidden_states / attn.rescale_output_factor + + return hidden_states + + +class AttnAddedKVProcessor: + r""" + Processor for performing attention-related computations with extra learnable key and value matrices for the text + encoder. + """ + + def __call__( + self, + attn: Attention, + hidden_states: torch.FloatTensor, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + scale: float = 1.0, + ) -> torch.Tensor: + residual = hidden_states + + args = () if USE_PEFT_BACKEND else (scale,) + + hidden_states = hidden_states.view(hidden_states.shape[0], hidden_states.shape[1], -1).transpose(1, 2) + batch_size, sequence_length, _ = hidden_states.shape + + attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) + + if encoder_hidden_states is None: + encoder_hidden_states = hidden_states + elif attn.norm_cross: + encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states) + + hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2) + + query = attn.to_q(hidden_states, *args) + query = attn.head_to_batch_dim(query) + + encoder_hidden_states_key_proj = attn.add_k_proj(encoder_hidden_states, *args) + encoder_hidden_states_value_proj = attn.add_v_proj(encoder_hidden_states, *args) + encoder_hidden_states_key_proj = attn.head_to_batch_dim(encoder_hidden_states_key_proj) + encoder_hidden_states_value_proj = attn.head_to_batch_dim(encoder_hidden_states_value_proj) + + if not attn.only_cross_attention: + key = attn.to_k(hidden_states, *args) + value = attn.to_v(hidden_states, *args) + key = attn.head_to_batch_dim(key) + value = attn.head_to_batch_dim(value) + key = torch.cat([encoder_hidden_states_key_proj, key], dim=1) + value = torch.cat([encoder_hidden_states_value_proj, value], dim=1) + else: + key = encoder_hidden_states_key_proj + value = encoder_hidden_states_value_proj + + attention_probs = attn.get_attention_scores(query, key, attention_mask) + hidden_states = torch.bmm(attention_probs, value) + hidden_states = attn.batch_to_head_dim(hidden_states) + + # linear proj + hidden_states = attn.to_out[0](hidden_states, *args) + # dropout + hidden_states = attn.to_out[1](hidden_states) + + hidden_states = hidden_states.transpose(-1, -2).reshape(residual.shape) + hidden_states = hidden_states + residual + + return hidden_states + + +class AttnAddedKVProcessor2_0: + r""" + Processor for performing scaled dot-product attention (enabled by default if you're using PyTorch 2.0), with extra + learnable key and value matrices for the text encoder. + """ + + def __init__(self): + if not hasattr(F, "scaled_dot_product_attention"): + raise ImportError( + "AttnAddedKVProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0." + ) + + def __call__( + self, + attn: Attention, + hidden_states: torch.FloatTensor, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + scale: float = 1.0, + ) -> torch.Tensor: + residual = hidden_states + + args = () if USE_PEFT_BACKEND else (scale,) + + hidden_states = hidden_states.view(hidden_states.shape[0], hidden_states.shape[1], -1).transpose(1, 2) + batch_size, sequence_length, _ = hidden_states.shape + + attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size, out_dim=4) + + if encoder_hidden_states is None: + encoder_hidden_states = hidden_states + elif attn.norm_cross: + encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states) + + hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2) + + query = attn.to_q(hidden_states, *args) + query = attn.head_to_batch_dim(query, out_dim=4) + + encoder_hidden_states_key_proj = attn.add_k_proj(encoder_hidden_states) + encoder_hidden_states_value_proj = attn.add_v_proj(encoder_hidden_states) + encoder_hidden_states_key_proj = attn.head_to_batch_dim(encoder_hidden_states_key_proj, out_dim=4) + encoder_hidden_states_value_proj = attn.head_to_batch_dim(encoder_hidden_states_value_proj, out_dim=4) + + if not attn.only_cross_attention: + key = attn.to_k(hidden_states, *args) + value = attn.to_v(hidden_states, *args) + key = attn.head_to_batch_dim(key, out_dim=4) + value = attn.head_to_batch_dim(value, out_dim=4) + key = torch.cat([encoder_hidden_states_key_proj, key], dim=2) + value = torch.cat([encoder_hidden_states_value_proj, value], dim=2) + else: + key = encoder_hidden_states_key_proj + value = encoder_hidden_states_value_proj + + # the output of sdp = (batch, num_heads, seq_len, head_dim) + # TODO: add support for attn.scale when we move to Torch 2.1 + hidden_states = F.scaled_dot_product_attention( + query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False + ) + hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, residual.shape[1]) + + # linear proj + hidden_states = attn.to_out[0](hidden_states, *args) + # dropout + hidden_states = attn.to_out[1](hidden_states) + + hidden_states = hidden_states.transpose(-1, -2).reshape(residual.shape) + hidden_states = hidden_states + residual + + return hidden_states + + +class AttnProcessor2_0: + r""" + Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0). + """ + + def __init__(self): + if not hasattr(F, "scaled_dot_product_attention"): + raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.") + + def __call__( + self, + attn: Attention, + hidden_states: torch.FloatTensor, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + temb: Optional[torch.FloatTensor] = None, + scale: float = 1.0, + ) -> torch.FloatTensor: + residual = hidden_states + if attn.spatial_norm is not None: + hidden_states = attn.spatial_norm(hidden_states, temb) + + input_ndim = hidden_states.ndim + + if input_ndim == 4: + batch_size, channel, height, width = hidden_states.shape + hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2) + + batch_size, sequence_length, _ = ( + hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape + ) + + if attention_mask is not None: + attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) + # scaled_dot_product_attention expects attention_mask shape to be + # (batch, heads, source_length, target_length) + attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1]) + + if attn.group_norm is not None: + hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2) + + args = () if USE_PEFT_BACKEND else (scale,) + query = attn.to_q(hidden_states, *args) + + if encoder_hidden_states is None: + encoder_hidden_states = hidden_states + elif attn.norm_cross: + encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states) + + key = attn.to_k(encoder_hidden_states, *args) + value = attn.to_v(encoder_hidden_states, *args) + + inner_dim = key.shape[-1] + head_dim = inner_dim // attn.heads + + query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) + + key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) + value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) + + hidden_states = torch_npu.npu_fusion_attention( + query, key, value, + atten_mask=attention_mask, + input_layout='BNSD', + scale=head_dim**-0.5, + head_num=attn.heads, + )[0] + # hidden_states = F.scaled_dot_product_attention( + # query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False + # ) + + hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim) + hidden_states = hidden_states.to(query.dtype) + + # linear proj + hidden_states = attn.to_out[0](hidden_states, *args) + # dropout + hidden_states = attn.to_out[1](hidden_states) + + if input_ndim == 4: + hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width) + + if attn.residual_connection: + hidden_states = hidden_states + residual + + hidden_states = hidden_states / attn.rescale_output_factor + + return hidden_states + + + +ADDED_KV_ATTENTION_PROCESSORS = ( + AttnAddedKVProcessor, + AttnAddedKVProcessor2_0, +) + +CROSS_ATTENTION_PROCESSORS = ( + AttnProcessor, + AttnProcessor2_0, +) + +AttentionProcessor = Union[ + AttnProcessor, + AttnProcessor2_0, + AttnAddedKVProcessor, + AttnAddedKVProcessor2_0, +] diff --git a/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/layers/embeddings.py b/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/layers/embeddings.py new file mode 100644 index 0000000000000000000000000000000000000000..9f46aa5f3b03048bfb3affeafb20894982775c5d --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/layers/embeddings.py @@ -0,0 +1,747 @@ +# Copyright 2023 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +import math +from typing import List, Optional, Tuple, Union + +import numpy as np +import torch +from torch import nn + +from diffusers.utils import USE_PEFT_BACKEND, deprecate +from diffusers.models.lora import LoRACompatibleLinear + +from .activations import get_activation +from diffusers.models.embeddings import PatchEmbed, ImagePositionalEmbeddings, PixArtAlphaCombinedTimestepSizeEmbeddings + +def get_timestep_embedding( + timesteps: torch.Tensor, + embedding_dim: int, + flip_sin_to_cos: bool = False, + downscale_freq_shift: float = 1, + scale: float = 1, + max_period: int = 10000, +): + """ + This matches the implementation in Denoising Diffusion Probabilistic Models: Create sinusoidal timestep embeddings. + + :param timesteps: a 1-D Tensor of N indices, one per batch element. + These may be fractional. + :param embedding_dim: the dimension of the output. :param max_period: controls the minimum frequency of the + embeddings. :return: an [N x dim] Tensor of positional embeddings. + """ + assert len(timesteps.shape) == 1, "Timesteps should be a 1d-array" + + half_dim = embedding_dim // 2 + exponent = -math.log(max_period) * torch.arange( + start=0, end=half_dim, dtype=torch.float32, device=timesteps.device + ) + exponent = exponent / (half_dim - downscale_freq_shift) + + emb = torch.exp(exponent) + emb = timesteps[:, None].float() * emb[None, :] + + # scale embeddings + emb = scale * emb + + # concat sine and cosine embeddings + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=-1) + + # flip sine and cosine embeddings + if flip_sin_to_cos: + emb = torch.cat([emb[:, half_dim:], emb[:, :half_dim]], dim=-1) + + # zero pad + if embedding_dim % 2 == 1: + emb = torch.nn.functional.pad(emb, (0, 1, 0, 0)) + return emb + + +def get_2d_sincos_pos_embed( + embed_dim, grid_size, cls_token=False, extra_tokens=0, interpolation_scale=1.0, base_size=16 +): + """ + grid_size: int of the grid height and width return: pos_embed: [grid_size*grid_size, embed_dim] or + [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token) + """ + if isinstance(grid_size, int): + grid_size = (grid_size, grid_size) + + grid_h = np.arange(grid_size[0], dtype=np.float32) / (grid_size[0] / base_size) / interpolation_scale + grid_w = np.arange(grid_size[1], dtype=np.float32) / (grid_size[1] / base_size) / interpolation_scale + grid = np.meshgrid(grid_w, grid_h) # here w goes first + grid = np.stack(grid, axis=0) + + grid = grid.reshape([2, 1, grid_size[1], grid_size[0]]) + pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid) + if cls_token and extra_tokens > 0: + pos_embed = np.concatenate([np.zeros([extra_tokens, embed_dim]), pos_embed], axis=0) + return pos_embed + + +def get_2d_sincos_pos_embed_from_grid(embed_dim, grid): + if embed_dim % 2 != 0: + raise ValueError("embed_dim must be divisible by 2") + + # use half of dimensions to encode grid_h + emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0]) # (H*W, D/2) + emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1]) # (H*W, D/2) + + emb = np.concatenate([emb_h, emb_w], axis=1) # (H*W, D) + return emb + + +def get_1d_sincos_pos_embed_from_grid(embed_dim, pos): + """ + embed_dim: output dimension for each position pos: a list of positions to be encoded: size (M,) out: (M, D) + """ + if embed_dim % 2 != 0: + raise ValueError("embed_dim must be divisible by 2") + + omega = np.arange(embed_dim // 2, dtype=np.float64) + omega /= embed_dim / 2.0 + omega = 1.0 / 10000**omega # (D/2,) + + pos = pos.reshape(-1) # (M,) + out = np.einsum("m,d->md", pos, omega) # (M, D/2), outer product + + emb_sin = np.sin(out) # (M, D/2) + emb_cos = np.cos(out) # (M, D/2) + + emb = np.concatenate([emb_sin, emb_cos], axis=1) # (M, D) + return emb + + +class TimestepEmbedding(nn.Module): + def __init__( + self, + in_channels: int, + time_embed_dim: int, + act_fn: str = "silu", + out_dim: int = None, + post_act_fn: Optional[str] = None, + cond_proj_dim=None, + sample_proj_bias=True, + ): + super().__init__() + linear_cls = nn.Linear if USE_PEFT_BACKEND else LoRACompatibleLinear + + self.linear_1 = linear_cls(in_channels, time_embed_dim, sample_proj_bias) + + if cond_proj_dim is not None: + self.cond_proj = nn.Linear(cond_proj_dim, in_channels, bias=False) + else: + self.cond_proj = None + + self.act = get_activation(act_fn) + + if out_dim is not None: + time_embed_dim_out = out_dim + else: + time_embed_dim_out = time_embed_dim + self.linear_2 = linear_cls(time_embed_dim, time_embed_dim_out, sample_proj_bias) + + if post_act_fn is None: + self.post_act = None + else: + self.post_act = get_activation(post_act_fn) + + def forward(self, sample, condition=None): + if condition is not None: + sample = sample + self.cond_proj(condition) + sample = self.linear_1(sample) + + if self.act is not None: + sample = self.act(sample) + + sample = self.linear_2(sample) + + if self.post_act is not None: + sample = self.post_act(sample) + return sample + + +class Timesteps(nn.Module): + def __init__(self, num_channels: int, flip_sin_to_cos: bool, downscale_freq_shift: float): + super().__init__() + self.num_channels = num_channels + self.flip_sin_to_cos = flip_sin_to_cos + self.downscale_freq_shift = downscale_freq_shift + + def forward(self, timesteps): + t_emb = get_timestep_embedding( + timesteps, + self.num_channels, + flip_sin_to_cos=self.flip_sin_to_cos, + downscale_freq_shift=self.downscale_freq_shift, + ) + return t_emb + + +class GaussianFourierProjection(nn.Module): + """Gaussian Fourier embeddings for noise levels.""" + + def __init__( + self, embedding_size: int = 256, scale: float = 1.0, set_W_to_weight=True, log=True, flip_sin_to_cos=False + ): + super().__init__() + self.weight = nn.Parameter(torch.randn(embedding_size) * scale, requires_grad=False) + self.log = log + self.flip_sin_to_cos = flip_sin_to_cos + + if set_W_to_weight: + # to delete later + self.W = nn.Parameter(torch.randn(embedding_size) * scale, requires_grad=False) + + self.weight = self.W + + def forward(self, x): + if self.log: + x = torch.log(x) + + x_proj = x[:, None] * self.weight[None, :] * 2 * np.pi + + if self.flip_sin_to_cos: + out = torch.cat([torch.cos(x_proj), torch.sin(x_proj)], dim=-1) + else: + out = torch.cat([torch.sin(x_proj), torch.cos(x_proj)], dim=-1) + return out + + +class SinusoidalPositionalEmbedding(nn.Module): + """Apply positional information to a sequence of embeddings. + + Takes in a sequence of embeddings with shape (batch_size, seq_length, embed_dim) and adds positional embeddings to + them + + Args: + embed_dim: (int): Dimension of the positional embedding. + max_seq_length: Maximum sequence length to apply positional embeddings + + """ + + def __init__(self, embed_dim: int, max_seq_length: int = 32): + super().__init__() + position = torch.arange(max_seq_length).unsqueeze(1) + div_term = torch.exp(torch.arange(0, embed_dim, 2) * (-math.log(10000.0) / embed_dim)) + pe = torch.zeros(1, max_seq_length, embed_dim) + pe[0, :, 0::2] = torch.sin(position * div_term) + pe[0, :, 1::2] = torch.cos(position * div_term) + self.register_buffer("pe", pe) + + def forward(self, x): + _, seq_length, _ = x.shape + x = x + self.pe[:, :seq_length] + return x + + +class LabelEmbedding(nn.Module): + """ + Embeds class labels into vector representations. Also handles label dropout for classifier-free guidance. + + Args: + num_classes (`int`): The number of classes. + hidden_size (`int`): The size of the vector embeddings. + dropout_prob (`float`): The probability of dropping a label. + """ + + def __init__(self, num_classes, hidden_size, dropout_prob): + super().__init__() + use_cfg_embedding = dropout_prob > 0 + self.embedding_table = nn.Embedding(num_classes + use_cfg_embedding, hidden_size) + self.num_classes = num_classes + self.dropout_prob = dropout_prob + + def token_drop(self, labels, force_drop_ids=None): + """ + Drops labels to enable classifier-free guidance. + """ + if force_drop_ids is None: + drop_ids = torch.rand(labels.shape[0], device=labels.device) < self.dropout_prob + else: + drop_ids = torch.tensor(force_drop_ids == 1) + labels = torch.where(drop_ids, self.num_classes, labels) + return labels + + def forward(self, labels: torch.LongTensor, force_drop_ids=None): + use_dropout = self.dropout_prob > 0 + if (self.training and use_dropout) or (force_drop_ids is not None): + labels = self.token_drop(labels, force_drop_ids) + embeddings = self.embedding_table(labels) + return embeddings + + +class TextImageProjection(nn.Module): + def __init__( + self, + text_embed_dim: int = 1024, + image_embed_dim: int = 768, + cross_attention_dim: int = 768, + num_image_text_embeds: int = 10, + ): + super().__init__() + + self.num_image_text_embeds = num_image_text_embeds + self.image_embeds = nn.Linear(image_embed_dim, self.num_image_text_embeds * cross_attention_dim) + self.text_proj = nn.Linear(text_embed_dim, cross_attention_dim) + + def forward(self, text_embeds: torch.FloatTensor, image_embeds: torch.FloatTensor): + batch_size = text_embeds.shape[0] + + # image + image_text_embeds = self.image_embeds(image_embeds) + image_text_embeds = image_text_embeds.reshape(batch_size, self.num_image_text_embeds, -1) + + # text + text_embeds = self.text_proj(text_embeds) + + return torch.cat([image_text_embeds, text_embeds], dim=1) + + +class ImageProjection(nn.Module): + def __init__( + self, + image_embed_dim: int = 768, + cross_attention_dim: int = 768, + num_image_text_embeds: int = 32, + ): + super().__init__() + + self.num_image_text_embeds = num_image_text_embeds + self.image_embeds = nn.Linear(image_embed_dim, self.num_image_text_embeds * cross_attention_dim) + self.norm = nn.LayerNorm(cross_attention_dim) + + def forward(self, image_embeds: torch.FloatTensor): + batch_size = image_embeds.shape[0] + + # image + image_embeds = self.image_embeds(image_embeds) + image_embeds = image_embeds.reshape(batch_size, self.num_image_text_embeds, -1) + image_embeds = self.norm(image_embeds) + return image_embeds + + +class IPAdapterFullImageProjection(nn.Module): + def __init__(self, image_embed_dim=1024, cross_attention_dim=1024): + super().__init__() + from .attention import FeedForward + + self.ff = FeedForward(image_embed_dim, cross_attention_dim, mult=1, activation_fn="gelu") + self.norm = nn.LayerNorm(cross_attention_dim) + + def forward(self, image_embeds: torch.FloatTensor): + return self.norm(self.ff(image_embeds)) + + +class CombinedTimestepLabelEmbeddings(nn.Module): + def __init__(self, num_classes, embedding_dim, class_dropout_prob=0.1): + super().__init__() + + self.time_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=1) + self.timestep_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=embedding_dim) + self.class_embedder = LabelEmbedding(num_classes, embedding_dim, class_dropout_prob) + + def forward(self, timestep, class_labels, hidden_dtype=None): + timesteps_proj = self.time_proj(timestep) + timesteps_emb = self.timestep_embedder(timesteps_proj.to(dtype=hidden_dtype)) # (N, D) + + class_labels = self.class_embedder(class_labels) # (N, D) + + conditioning = timesteps_emb + class_labels # (N, D) + + return conditioning + + +class TextTimeEmbedding(nn.Module): + def __init__(self, encoder_dim: int, time_embed_dim: int, num_heads: int = 64): + super().__init__() + self.norm1 = nn.LayerNorm(encoder_dim) + self.pool = AttentionPooling(num_heads, encoder_dim) + self.proj = nn.Linear(encoder_dim, time_embed_dim) + self.norm2 = nn.LayerNorm(time_embed_dim) + + def forward(self, hidden_states): + hidden_states = self.norm1(hidden_states) + hidden_states = self.pool(hidden_states) + hidden_states = self.proj(hidden_states) + hidden_states = self.norm2(hidden_states) + return hidden_states + + +class TextImageTimeEmbedding(nn.Module): + def __init__(self, text_embed_dim: int = 768, image_embed_dim: int = 768, time_embed_dim: int = 1536): + super().__init__() + self.text_proj = nn.Linear(text_embed_dim, time_embed_dim) + self.text_norm = nn.LayerNorm(time_embed_dim) + self.image_proj = nn.Linear(image_embed_dim, time_embed_dim) + + def forward(self, text_embeds: torch.FloatTensor, image_embeds: torch.FloatTensor): + # text + time_text_embeds = self.text_proj(text_embeds) + time_text_embeds = self.text_norm(time_text_embeds) + + # image + time_image_embeds = self.image_proj(image_embeds) + + return time_image_embeds + time_text_embeds + + +class ImageTimeEmbedding(nn.Module): + def __init__(self, image_embed_dim: int = 768, time_embed_dim: int = 1536): + super().__init__() + self.image_proj = nn.Linear(image_embed_dim, time_embed_dim) + self.image_norm = nn.LayerNorm(time_embed_dim) + + def forward(self, image_embeds: torch.FloatTensor): + # image + time_image_embeds = self.image_proj(image_embeds) + time_image_embeds = self.image_norm(time_image_embeds) + return time_image_embeds + + +class ImageHintTimeEmbedding(nn.Module): + def __init__(self, image_embed_dim: int = 768, time_embed_dim: int = 1536): + super().__init__() + self.image_proj = nn.Linear(image_embed_dim, time_embed_dim) + self.image_norm = nn.LayerNorm(time_embed_dim) + self.input_hint_block = nn.Sequential( + nn.Conv2d(3, 16, 3, padding=1), + nn.SiLU(), + nn.Conv2d(16, 16, 3, padding=1), + nn.SiLU(), + nn.Conv2d(16, 32, 3, padding=1, stride=2), + nn.SiLU(), + nn.Conv2d(32, 32, 3, padding=1), + nn.SiLU(), + nn.Conv2d(32, 96, 3, padding=1, stride=2), + nn.SiLU(), + nn.Conv2d(96, 96, 3, padding=1), + nn.SiLU(), + nn.Conv2d(96, 256, 3, padding=1, stride=2), + nn.SiLU(), + nn.Conv2d(256, 4, 3, padding=1), + ) + + def forward(self, image_embeds: torch.FloatTensor, hint: torch.FloatTensor): + # image + time_image_embeds = self.image_proj(image_embeds) + time_image_embeds = self.image_norm(time_image_embeds) + hint = self.input_hint_block(hint) + return time_image_embeds, hint + + +class AttentionPooling(nn.Module): + # Copied from https://github.com/deep-floyd/IF/blob/2f91391f27dd3c468bf174be5805b4cc92980c0b/deepfloyd_if/model/nn.py#L54 + + def __init__(self, num_heads, embed_dim, dtype=None): + super().__init__() + self.dtype = dtype + self.positional_embedding = nn.Parameter(torch.randn(1, embed_dim) / embed_dim**0.5) + self.k_proj = nn.Linear(embed_dim, embed_dim, dtype=self.dtype) + self.q_proj = nn.Linear(embed_dim, embed_dim, dtype=self.dtype) + self.v_proj = nn.Linear(embed_dim, embed_dim, dtype=self.dtype) + self.num_heads = num_heads + self.dim_per_head = embed_dim // self.num_heads + + def forward(self, x): + bs, length, width = x.size() + + def shape(x): + # (bs, length, width) --> (bs, length, n_heads, dim_per_head) + x = x.view(bs, -1, self.num_heads, self.dim_per_head) + # (bs, length, n_heads, dim_per_head) --> (bs, n_heads, length, dim_per_head) + x = x.transpose(1, 2) + # (bs, n_heads, length, dim_per_head) --> (bs*n_heads, length, dim_per_head) + x = x.reshape(bs * self.num_heads, -1, self.dim_per_head) + # (bs*n_heads, length, dim_per_head) --> (bs*n_heads, dim_per_head, length) + x = x.transpose(1, 2) + return x + + class_token = x.mean(dim=1, keepdim=True) + self.positional_embedding.to(x.dtype) + x = torch.cat([class_token, x], dim=1) # (bs, length+1, width) + + # (bs*n_heads, class_token_length, dim_per_head) + q = shape(self.q_proj(class_token)) + # (bs*n_heads, length+class_token_length, dim_per_head) + k = shape(self.k_proj(x)) + v = shape(self.v_proj(x)) + + # (bs*n_heads, class_token_length, length+class_token_length): + scale = 1 / math.sqrt(math.sqrt(self.dim_per_head)) + weight = torch.einsum("bct,bcs->bts", q * scale, k * scale) # More stable with f16 than dividing afterwards + weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype) + + # (bs*n_heads, dim_per_head, class_token_length) + a = torch.einsum("bts,bcs->bct", weight, v) + + # (bs, length+1, width) + a = a.reshape(bs, -1, 1).transpose(1, 2) + + return a[:, 0, :] # cls_token + + +def get_fourier_embeds_from_boundingbox(embed_dim, box): + """ + Args: + embed_dim: int + box: a 3-D tensor [B x N x 4] representing the bounding boxes for GLIGEN pipeline + Returns: + [B x N x embed_dim] tensor of positional embeddings + """ + + batch_size, num_boxes = box.shape[:2] + + emb = 100 ** (torch.arange(embed_dim) / embed_dim) + emb = emb[None, None, None].to(device=box.device, dtype=box.dtype) + emb = emb * box.unsqueeze(-1) + + emb = torch.stack((emb.sin(), emb.cos()), dim=-1) + emb = emb.permute(0, 1, 3, 4, 2).reshape(batch_size, num_boxes, embed_dim * 2 * 4) + + return emb + + +class GLIGENTextBoundingboxProjection(nn.Module): + def __init__(self, positive_len, out_dim, feature_type="text-only", fourier_freqs=8): + super().__init__() + self.positive_len = positive_len + self.out_dim = out_dim + + self.fourier_embedder_dim = fourier_freqs + self.position_dim = fourier_freqs * 2 * 4 # 2: sin/cos, 4: xyxy + + if isinstance(out_dim, tuple): + out_dim = out_dim[0] + + if feature_type == "text-only": + self.linears = nn.Sequential( + nn.Linear(self.positive_len + self.position_dim, 512), + nn.SiLU(), + nn.Linear(512, 512), + nn.SiLU(), + nn.Linear(512, out_dim), + ) + self.null_positive_feature = torch.nn.Parameter(torch.zeros([self.positive_len])) + + elif feature_type == "text-image": + self.linears_text = nn.Sequential( + nn.Linear(self.positive_len + self.position_dim, 512), + nn.SiLU(), + nn.Linear(512, 512), + nn.SiLU(), + nn.Linear(512, out_dim), + ) + self.linears_image = nn.Sequential( + nn.Linear(self.positive_len + self.position_dim, 512), + nn.SiLU(), + nn.Linear(512, 512), + nn.SiLU(), + nn.Linear(512, out_dim), + ) + self.null_text_feature = torch.nn.Parameter(torch.zeros([self.positive_len])) + self.null_image_feature = torch.nn.Parameter(torch.zeros([self.positive_len])) + + self.null_position_feature = torch.nn.Parameter(torch.zeros([self.position_dim])) + + def forward( + self, + boxes, + masks, + positive_embeddings=None, + phrases_masks=None, + image_masks=None, + phrases_embeddings=None, + image_embeddings=None, + ): + masks = masks.unsqueeze(-1) + + # embedding position (it may includes padding as placeholder) + xyxy_embedding = get_fourier_embeds_from_boundingbox(self.fourier_embedder_dim, boxes) # B*N*4 -> B*N*C + + # learnable null embedding + xyxy_null = self.null_position_feature.view(1, 1, -1) + + # replace padding with learnable null embedding + xyxy_embedding = xyxy_embedding * masks + (1 - masks) * xyxy_null + + # positionet with text only information + if positive_embeddings is not None: + # learnable null embedding + positive_null = self.null_positive_feature.view(1, 1, -1) + + # replace padding with learnable null embedding + positive_embeddings = positive_embeddings * masks + (1 - masks) * positive_null + + objs = self.linears(torch.cat([positive_embeddings, xyxy_embedding], dim=-1)) + + # positionet with text and image infomation + else: + phrases_masks = phrases_masks.unsqueeze(-1) + image_masks = image_masks.unsqueeze(-1) + + # learnable null embedding + text_null = self.null_text_feature.view(1, 1, -1) + image_null = self.null_image_feature.view(1, 1, -1) + + # replace padding with learnable null embedding + phrases_embeddings = phrases_embeddings * phrases_masks + (1 - phrases_masks) * text_null + image_embeddings = image_embeddings * image_masks + (1 - image_masks) * image_null + + objs_text = self.linears_text(torch.cat([phrases_embeddings, xyxy_embedding], dim=-1)) + objs_image = self.linears_image(torch.cat([image_embeddings, xyxy_embedding], dim=-1)) + objs = torch.cat([objs_text, objs_image], dim=1) + + return objs + + +class PixArtAlphaTextProjection(nn.Module): + """ + Projects caption embeddings. Also handles dropout for classifier-free guidance. + + Adapted from https://github.com/PixArt-alpha/PixArt-alpha/blob/master/diffusion/model/nets/PixArt_blocks.py + """ + + def __init__(self, in_features, hidden_size, num_tokens=120): + super().__init__() + self.linear_1 = nn.Linear(in_features=in_features, out_features=hidden_size, bias=True) + self.act_1 = nn.GELU(approximate="tanh") + self.linear_2 = nn.Linear(in_features=hidden_size, out_features=hidden_size, bias=True) + + def forward(self, caption): + hidden_states = self.linear_1(caption) + hidden_states = self.act_1(hidden_states) + hidden_states = self.linear_2(hidden_states) + return hidden_states + + +class IPAdapterPlusImageProjection(nn.Module): + """Resampler of IP-Adapter Plus. + + Args: + ---- + embed_dims (int): The feature dimension. Defaults to 768. + output_dims (int): The number of output channels, that is the same + number of the channels in the + `unet.config.cross_attention_dim`. Defaults to 1024. + hidden_dims (int): The number of hidden channels. Defaults to 1280. + depth (int): The number of blocks. Defaults to 8. + dim_head (int): The number of head channels. Defaults to 64. + heads (int): Parallel attention heads. Defaults to 16. + num_queries (int): The number of queries. Defaults to 8. + ffn_ratio (float): The expansion ratio of feedforward network hidden + layer channels. Defaults to 4. + """ + + def __init__( + self, + embed_dims: int = 768, + output_dims: int = 1024, + hidden_dims: int = 1280, + depth: int = 4, + dim_head: int = 64, + heads: int = 16, + num_queries: int = 8, + ffn_ratio: float = 4, + ) -> None: + super().__init__() + from .attention import FeedForward # Lazy import to avoid circular import + + self.latents = nn.Parameter(torch.randn(1, num_queries, hidden_dims) / hidden_dims**0.5) + + self.proj_in = nn.Linear(embed_dims, hidden_dims) + + self.proj_out = nn.Linear(hidden_dims, output_dims) + self.norm_out = nn.LayerNorm(output_dims) + + self.layers = nn.ModuleList([]) + for _ in range(depth): + self.layers.append( + nn.ModuleList( + [ + nn.LayerNorm(hidden_dims), + nn.LayerNorm(hidden_dims), + Attention( + query_dim=hidden_dims, + dim_head=dim_head, + heads=heads, + out_bias=False, + ), + nn.Sequential( + nn.LayerNorm(hidden_dims), + FeedForward(hidden_dims, hidden_dims, activation_fn="gelu", mult=ffn_ratio, bias=False), + ), + ] + ) + ) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + """Forward pass. + + Args: + ---- + x (torch.Tensor): Input Tensor. + + Returns: + ------- + torch.Tensor: Output Tensor. + """ + latents = self.latents.repeat(x.size(0), 1, 1) + + x = self.proj_in(x) + + for ln0, ln1, attn, ff in self.layers: + residual = latents + + encoder_hidden_states = ln0(x) + latents = ln1(latents) + encoder_hidden_states = torch.cat([encoder_hidden_states, latents], dim=-2) + latents = attn(latents, encoder_hidden_states) + residual + latents = ff(latents) + latents + + latents = self.proj_out(latents) + return self.norm_out(latents) + + +class MultiIPAdapterImageProjection(nn.Module): + def __init__(self, IPAdapterImageProjectionLayers: Union[List[nn.Module], Tuple[nn.Module]]): + super().__init__() + self.image_projection_layers = nn.ModuleList(IPAdapterImageProjectionLayers) + + def forward(self, image_embeds: List[torch.FloatTensor]): + projected_image_embeds = [] + + # currently, we accept `image_embeds` as + # 1. a tensor (deprecated) with shape [batch_size, embed_dim] or [batch_size, sequence_length, embed_dim] + # 2. list of `n` tensors where `n` is number of ip-adapters, each tensor can hae shape [batch_size, num_images, embed_dim] or [batch_size, num_images, sequence_length, embed_dim] + if not isinstance(image_embeds, list): + deprecation_message = ( + "You have passed a tensor as `image_embeds`.This is deprecated and will be removed in a future release." + " Please make sure to update your script to pass `image_embeds` as a list of tensors to supress this warning." + ) + deprecate("image_embeds not a list", "1.0.0", deprecation_message, standard_warn=False) + image_embeds = [image_embeds.unsqueeze(1)] + + if len(image_embeds) != len(self.image_projection_layers): + raise ValueError( + f"image_embeds must have the same length as image_projection_layers, got {len(image_embeds)} and {len(self.image_projection_layers)}" + ) + + for image_embed, image_projection_layer in zip(image_embeds, self.image_projection_layers): + batch_size, num_images = image_embed.shape[0], image_embed.shape[1] + image_embed = image_embed.reshape((batch_size * num_images,) + image_embed.shape[2:]) + image_embed = image_projection_layer(image_embed) + image_embed = image_embed.reshape((batch_size, num_images) + image_embed.shape[1:]) + + projected_image_embeds.append(image_embed) + + return projected_image_embeds diff --git a/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/layers/normalization.py b/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/layers/normalization.py new file mode 100644 index 0000000000000000000000000000000000000000..5acced9ea6a6c10c800e2560c6390e3134bc82a5 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/layers/normalization.py @@ -0,0 +1,254 @@ +# coding=utf-8 +# Copyright 2023 HuggingFace Inc. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import numbers +from typing import Dict, Optional, Tuple + +import torch +import torch.nn as nn +import torch.nn.functional as F + +from diffusers.utils import is_torch_version +from .activations import get_activation +from .embeddings import CombinedTimestepLabelEmbeddings, PixArtAlphaCombinedTimestepSizeEmbeddings + + +class AdaLayerNorm(nn.Module): + r""" + Norm layer modified to incorporate timestep embeddings. + + Parameters: + embedding_dim (`int`): The size of each embedding vector. + num_embeddings (`int`): The size of the embeddings dictionary. + """ + + def __init__(self, embedding_dim: int, num_embeddings: int): + super().__init__() + self.emb = nn.Embedding(num_embeddings, embedding_dim) + self.silu = nn.SiLU() + self.linear = nn.Linear(embedding_dim, embedding_dim * 2) + self.norm = nn.LayerNorm(embedding_dim, elementwise_affine=False) + + def forward(self, x: torch.Tensor, timestep: torch.Tensor) -> torch.Tensor: + emb = self.linear(self.silu(self.emb(timestep))) + scale, shift = torch.chunk(emb, 2) + x = self.norm(x) * (1 + scale) + shift + return x + + +class AdaLayerNormZero(nn.Module): + r""" + Norm layer adaptive layer norm zero (adaLN-Zero). + + Parameters: + embedding_dim (`int`): The size of each embedding vector. + num_embeddings (`int`): The size of the embeddings dictionary. + """ + + def __init__(self, embedding_dim: int, num_embeddings: int): + super().__init__() + + self.emb = CombinedTimestepLabelEmbeddings(num_embeddings, embedding_dim) + + self.silu = nn.SiLU() + self.linear = nn.Linear(embedding_dim, 6 * embedding_dim, bias=True) + self.norm = nn.LayerNorm(embedding_dim, elementwise_affine=False, eps=1e-6) + + def forward( + self, + x: torch.Tensor, + timestep: torch.Tensor, + class_labels: torch.LongTensor, + hidden_dtype: Optional[torch.dtype] = None, + ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]: + emb = self.linear(self.silu(self.emb(timestep, class_labels, hidden_dtype=hidden_dtype))) + shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = emb.chunk(6, dim=1) + x = self.norm(x) * (1 + scale_msa[:, None]) + shift_msa[:, None] + return x, gate_msa, shift_mlp, scale_mlp, gate_mlp + + +class AdaLayerNormSingle(nn.Module): + r""" + Norm layer adaptive layer norm single (adaLN-single). + + As proposed in PixArt-Alpha (see: https://arxiv.org/abs/2310.00426; Section 2.3). + + Parameters: + embedding_dim (`int`): The size of each embedding vector. + use_additional_conditions (`bool`): To use additional conditions for normalization or not. + """ + + def __init__(self, embedding_dim: int, use_additional_conditions: bool = False): + super().__init__() + + self.emb = PixArtAlphaCombinedTimestepSizeEmbeddings( + embedding_dim, size_emb_dim=embedding_dim // 3, use_additional_conditions=use_additional_conditions + ) + + self.silu = nn.SiLU() + self.linear = nn.Linear(embedding_dim, 6 * embedding_dim, bias=True) + + def forward( + self, + timestep: torch.Tensor, + added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None, + batch_size: Optional[int] = None, + hidden_dtype: Optional[torch.dtype] = None, + ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]: + # No modulation happening here. + embedded_timestep = self.emb(timestep, **added_cond_kwargs, batch_size=batch_size, hidden_dtype=hidden_dtype) + return self.linear(self.silu(embedded_timestep)), embedded_timestep + + +class AdaGroupNorm(nn.Module): + r""" + GroupNorm layer modified to incorporate timestep embeddings. + + Parameters: + embedding_dim (`int`): The size of each embedding vector. + num_embeddings (`int`): The size of the embeddings dictionary. + num_groups (`int`): The number of groups to separate the channels into. + act_fn (`str`, *optional*, defaults to `None`): The activation function to use. + eps (`float`, *optional*, defaults to `1e-5`): The epsilon value to use for numerical stability. + """ + + def __init__( + self, embedding_dim: int, out_dim: int, num_groups: int, act_fn: Optional[str] = None, eps: float = 1e-5 + ): + super().__init__() + self.num_groups = num_groups + self.eps = eps + + if act_fn is None: + self.act = None + else: + self.act = get_activation(act_fn) + + self.linear = nn.Linear(embedding_dim, out_dim * 2) + + def forward(self, x: torch.Tensor, emb: torch.Tensor) -> torch.Tensor: + if self.act: + emb = self.act(emb) + emb = self.linear(emb) + emb = emb[:, :, None, None] + scale, shift = emb.chunk(2, dim=1) + + x = F.group_norm(x, self.num_groups, eps=self.eps) + x = x * (1 + scale) + shift + return x + + +class AdaLayerNormContinuous(nn.Module): + def __init__( + self, + embedding_dim: int, + conditioning_embedding_dim: int, + # NOTE: It is a bit weird that the norm layer can be configured to have scale and shift parameters + # because the output is immediately scaled and shifted by the projected conditioning embeddings. + # Note that AdaLayerNorm does not let the norm layer have scale and shift parameters. + # However, this is how it was implemented in the original code, and it's rather likely you should + # set `elementwise_affine` to False. + elementwise_affine=True, + eps=1e-5, + bias=True, + norm_type="layer_norm", + ): + super().__init__() + self.silu = nn.SiLU() + self.linear = nn.Linear(conditioning_embedding_dim, embedding_dim * 2, bias=bias) + if norm_type == "layer_norm": + self.norm = LayerNorm(embedding_dim, eps, elementwise_affine, bias) + elif norm_type == "rms_norm": + self.norm = RMSNorm(embedding_dim, eps, elementwise_affine) + else: + raise ValueError(f"unknown norm_type {norm_type}") + + def forward(self, x: torch.Tensor, conditioning_embedding: torch.Tensor) -> torch.Tensor: + emb = self.linear(self.silu(conditioning_embedding)) + scale, shift = torch.chunk(emb, 2, dim=1) + x = self.norm(x) * (1 + scale)[:, None, :] + shift[:, None, :] + return x + + +if is_torch_version(">=", "2.1.0"): + LayerNorm = nn.LayerNorm +else: + # Has optional bias parameter compared to torch layer norm + # TODO: replace with torch layernorm once min required torch version >= 2.1 + class LayerNorm(nn.Module): + def __init__(self, dim, eps: float = 1e-5, elementwise_affine: bool = True, bias: bool = True): + super().__init__() + + self.eps = eps + + if isinstance(dim, numbers.Integral): + dim = (dim,) + + self.dim = torch.Size(dim) + + if elementwise_affine: + self.weight = nn.Parameter(torch.ones(dim)) + self.bias = nn.Parameter(torch.zeros(dim)) if bias else None + else: + self.weight = None + self.bias = None + + def forward(self, input): + return F.layer_norm(input, self.dim, self.weight, self.bias, self.eps) + + +class RMSNorm(nn.Module): + def __init__(self, dim, eps: float, elementwise_affine: bool = True): + super().__init__() + + self.eps = eps + + if isinstance(dim, numbers.Integral): + dim = (dim,) + + self.dim = torch.Size(dim) + + if elementwise_affine: + self.weight = nn.Parameter(torch.ones(dim)) + else: + self.weight = None + + def forward(self, hidden_states): + input_dtype = hidden_states.dtype + variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True) + hidden_states = hidden_states * torch.rsqrt(variance + self.eps) + + if self.weight is not None: + # convert into half-precision if necessary + if self.weight.dtype in [torch.float16, torch.bfloat16]: + hidden_states = hidden_states.to(self.weight.dtype) + hidden_states = hidden_states * self.weight + else: + hidden_states = hidden_states.to(input_dtype) + + return hidden_states + + +class GlobalResponseNorm(nn.Module): + # Taken from https://github.com/facebookresearch/ConvNeXt-V2/blob/3608f67cc1dae164790c5d0aead7bf2d73d9719b/models/utils.py#L105 + def __init__(self, dim): + super().__init__() + self.gamma = nn.Parameter(torch.zeros(1, 1, 1, dim)) + self.beta = nn.Parameter(torch.zeros(1, 1, 1, dim)) + + def forward(self, x): + gx = torch.norm(x, p=2, dim=(1, 2), keepdim=True) + nx = gx / (gx.mean(dim=-1, keepdim=True) + 1e-6) + return self.gamma * (x * nx) + self.beta + x diff --git a/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/models/__init__.py b/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/models/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..5af450275b73656a52deb912765cab1461ffd090 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/models/__init__.py @@ -0,0 +1 @@ +from .unets.unet_2d_condition import UNet2DConditionModel \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/models/model_utils.py b/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/models/model_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..0c9d557c2e7ca0446054d8ec36a71ee1f9de7d80 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/models/model_utils.py @@ -0,0 +1 @@ +from diffusers.models.modeling_utils import ModelMixin \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/models/transformers/__init__.py b/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/models/transformers/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..f5429a94ed6f940af278e7a7b7050fadf912f627 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/models/transformers/__init__.py @@ -0,0 +1,2 @@ +from .dual_transformer_2d import DualTransformer2DModel +from .transformer_2d import Transformer2DModel \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/models/transformers/dual_transformer_2d.py b/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/models/transformers/dual_transformer_2d.py new file mode 100644 index 0000000000000000000000000000000000000000..21b135c2eb86a6cc1ec2799c61c094e814893267 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/models/transformers/dual_transformer_2d.py @@ -0,0 +1,155 @@ +# Copyright 2023 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from typing import Optional + +from torch import nn + +from .transformer_2d import Transformer2DModel, Transformer2DModelOutput + + +class DualTransformer2DModel(nn.Module): + """ + Dual transformer wrapper that combines two `Transformer2DModel`s for mixed inference. + + Parameters: + num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention. + attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head. + in_channels (`int`, *optional*): + Pass if the input is continuous. The number of channels in the input and output. + num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use. + dropout (`float`, *optional*, defaults to 0.1): The dropout probability to use. + cross_attention_dim (`int`, *optional*): The number of encoder_hidden_states dimensions to use. + sample_size (`int`, *optional*): Pass if the input is discrete. The width of the latent images. + Note that this is fixed at training time as it is used for learning a number of position embeddings. See + `ImagePositionalEmbeddings`. + num_vector_embeds (`int`, *optional*): + Pass if the input is discrete. The number of classes of the vector embeddings of the latent pixels. + Includes the class for the masked latent pixel. + activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward. + num_embeds_ada_norm ( `int`, *optional*): Pass if at least one of the norm_layers is `AdaLayerNorm`. + The number of diffusion steps used during training. Note that this is fixed at training time as it is used + to learn a number of embeddings that are added to the hidden states. During inference, you can denoise for + up to but not more than steps than `num_embeds_ada_norm`. + attention_bias (`bool`, *optional*): + Configure if the TransformerBlocks' attention should contain a bias parameter. + """ + + def __init__( + self, + num_attention_heads: int = 16, + attention_head_dim: int = 88, + in_channels: Optional[int] = None, + num_layers: int = 1, + dropout: float = 0.0, + norm_num_groups: int = 32, + cross_attention_dim: Optional[int] = None, + attention_bias: bool = False, + sample_size: Optional[int] = None, + num_vector_embeds: Optional[int] = None, + activation_fn: str = "geglu", + num_embeds_ada_norm: Optional[int] = None, + ): + super().__init__() + self.transformers = nn.ModuleList( + [ + Transformer2DModel( + num_attention_heads=num_attention_heads, + attention_head_dim=attention_head_dim, + in_channels=in_channels, + num_layers=num_layers, + dropout=dropout, + norm_num_groups=norm_num_groups, + cross_attention_dim=cross_attention_dim, + attention_bias=attention_bias, + sample_size=sample_size, + num_vector_embeds=num_vector_embeds, + activation_fn=activation_fn, + num_embeds_ada_norm=num_embeds_ada_norm, + ) + for _ in range(2) + ] + ) + + # Variables that can be set by a pipeline: + + # The ratio of transformer1 to transformer2's output states to be combined during inference + self.mix_ratio = 0.5 + + # The shape of `encoder_hidden_states` is expected to be + # `(batch_size, condition_lengths[0]+condition_lengths[1], num_features)` + self.condition_lengths = [77, 257] + + # Which transformer to use to encode which condition. + # E.g. `(1, 0)` means that we'll use `transformers[1](conditions[0])` and `transformers[0](conditions[1])` + self.transformer_index_for_condition = [1, 0] + + def forward( + self, + hidden_states, + encoder_hidden_states, + timestep=None, + attention_mask=None, + cross_attention_kwargs=None, + return_dict: bool = True, + ): + """ + Args: + hidden_states ( When discrete, `torch.LongTensor` of shape `(batch size, num latent pixels)`. + When continuous, `torch.FloatTensor` of shape `(batch size, channel, height, width)`): Input + hidden_states. + encoder_hidden_states ( `torch.LongTensor` of shape `(batch size, encoder_hidden_states dim)`, *optional*): + Conditional embeddings for cross attention layer. If not given, cross-attention defaults to + self-attention. + timestep ( `torch.long`, *optional*): + Optional timestep to be applied as an embedding in AdaLayerNorm's. Used to indicate denoising step. + attention_mask (`torch.FloatTensor`, *optional*): + Optional attention mask to be applied in Attention. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`models.unets.unet_2d_condition.UNet2DConditionOutput`] instead of a plain tuple. + + Returns: + [`~models.transformer_2d.Transformer2DModelOutput`] or `tuple`: + [`~models.transformer_2d.Transformer2DModelOutput`] if `return_dict` is True, otherwise a `tuple`. When + returning a tuple, the first element is the sample tensor. + """ + input_states = hidden_states + + encoded_states = [] + tokens_start = 0 + # attention_mask is not used yet + for i in range(2): + # for each of the two transformers, pass the corresponding condition tokens + condition_state = encoder_hidden_states[:, tokens_start : tokens_start + self.condition_lengths[i]] + transformer_index = self.transformer_index_for_condition[i] + encoded_state = self.transformers[transformer_index]( + input_states, + encoder_hidden_states=condition_state, + timestep=timestep, + cross_attention_kwargs=cross_attention_kwargs, + return_dict=False, + )[0] + encoded_states.append(encoded_state - input_states) + tokens_start += self.condition_lengths[i] + + output_states = encoded_states[0] * self.mix_ratio + encoded_states[1] * (1 - self.mix_ratio) + output_states = output_states + input_states + + if not return_dict: + return (output_states,) + + return Transformer2DModelOutput(sample=output_states) diff --git a/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/models/transformers/transformer_2d.py b/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/models/transformers/transformer_2d.py new file mode 100644 index 0000000000000000000000000000000000000000..e33621f2497a737e92cd1e6e5d7386ea1f83b76c --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/models/transformers/transformer_2d.py @@ -0,0 +1,459 @@ +# Copyright 2023 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from dataclasses import dataclass +from typing import Any, Dict, Optional + +import torch +import torch.nn.functional as F +from torch import nn + +from diffusers.configuration_utils import ConfigMixin, register_to_config +from diffusers.utils import USE_PEFT_BACKEND, BaseOutput, deprecate, is_torch_version +from diffusers.models.lora import LoRACompatibleConv, LoRACompatibleLinear + +from stablediffusion.layers.attention import BasicTransformerBlock +from stablediffusion.layers.embeddings import ImagePositionalEmbeddings, PatchEmbed, PixArtAlphaTextProjection +from stablediffusion.layers.normalization import AdaLayerNormSingle +from stablediffusion.models.model_utils import ModelMixin + + +@dataclass +class Transformer2DModelOutput(BaseOutput): + """ + The output of [`Transformer2DModel`]. + + Args: + sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` or `(batch size, num_vector_embeds - 1, num_latent_pixels)` if [`Transformer2DModel`] is discrete): + The hidden states output conditioned on the `encoder_hidden_states` input. If discrete, returns probability + distributions for the unnoised latent pixels. + """ + + sample: torch.FloatTensor + + +class Transformer2DModel(ModelMixin, ConfigMixin): + """ + A 2D Transformer model for image-like data. + + Parameters: + num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention. + attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head. + in_channels (`int`, *optional*): + The number of channels in the input and output (specify if the input is **continuous**). + num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use. + dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. + cross_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use. + sample_size (`int`, *optional*): The width of the latent images (specify if the input is **discrete**). + This is fixed during training since it is used to learn a number of position embeddings. + num_vector_embeds (`int`, *optional*): + The number of classes of the vector embeddings of the latent pixels (specify if the input is **discrete**). + Includes the class for the masked latent pixel. + activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to use in feed-forward. + num_embeds_ada_norm ( `int`, *optional*): + The number of diffusion steps used during training. Pass if at least one of the norm_layers is + `AdaLayerNorm`. This is fixed during training since it is used to learn a number of embeddings that are + added to the hidden states. + + During inference, you can denoise for up to but not more steps than `num_embeds_ada_norm`. + attention_bias (`bool`, *optional*): + Configure if the `TransformerBlocks` attention should contain a bias parameter. + """ + + _supports_gradient_checkpointing = True + + @register_to_config + def __init__( + self, + num_attention_heads: int = 16, + attention_head_dim: int = 88, + in_channels: Optional[int] = None, + out_channels: Optional[int] = None, + num_layers: int = 1, + dropout: float = 0.0, + norm_num_groups: int = 32, + cross_attention_dim: Optional[int] = None, + attention_bias: bool = False, + sample_size: Optional[int] = None, + num_vector_embeds: Optional[int] = None, + patch_size: Optional[int] = None, + activation_fn: str = "geglu", + num_embeds_ada_norm: Optional[int] = None, + use_linear_projection: bool = False, + only_cross_attention: bool = False, + double_self_attention: bool = False, + upcast_attention: bool = False, + norm_type: str = "layer_norm", + norm_elementwise_affine: bool = True, + norm_eps: float = 1e-5, + attention_type: str = "default", + caption_channels: int = None, + ): + super().__init__() + self.use_linear_projection = use_linear_projection + self.num_attention_heads = num_attention_heads + self.attention_head_dim = attention_head_dim + inner_dim = num_attention_heads * attention_head_dim + + conv_cls = nn.Conv2d if USE_PEFT_BACKEND else LoRACompatibleConv + linear_cls = nn.Linear if USE_PEFT_BACKEND else LoRACompatibleLinear + + # 1. Transformer2DModel can process both standard continuous images of shape `(batch_size, num_channels, width, height)` as well as quantized image embeddings of shape `(batch_size, num_image_vectors)` + # Define whether input is continuous or discrete depending on configuration + self.is_input_continuous = (in_channels is not None) and (patch_size is None) + self.is_input_vectorized = num_vector_embeds is not None + self.is_input_patches = in_channels is not None and patch_size is not None + + if norm_type == "layer_norm" and num_embeds_ada_norm is not None: + deprecation_message = ( + f"The configuration file of this model: {self.__class__} is outdated. `norm_type` is either not set or" + " incorrectly set to `'layer_norm'`.Make sure to set `norm_type` to `'ada_norm'` in the config." + " Please make sure to update the config accordingly as leaving `norm_type` might led to incorrect" + " results in future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it" + " would be very nice if you could open a Pull request for the `transformer/config.json` file" + ) + deprecate("norm_type!=num_embeds_ada_norm", "1.0.0", deprecation_message, standard_warn=False) + norm_type = "ada_norm" + + if self.is_input_continuous and self.is_input_vectorized: + raise ValueError( + f"Cannot define both `in_channels`: {in_channels} and `num_vector_embeds`: {num_vector_embeds}. Make" + " sure that either `in_channels` or `num_vector_embeds` is None." + ) + elif self.is_input_vectorized and self.is_input_patches: + raise ValueError( + f"Cannot define both `num_vector_embeds`: {num_vector_embeds} and `patch_size`: {patch_size}. Make" + " sure that either `num_vector_embeds` or `num_patches` is None." + ) + elif not self.is_input_continuous and not self.is_input_vectorized and not self.is_input_patches: + raise ValueError( + f"Has to define `in_channels`: {in_channels}, `num_vector_embeds`: {num_vector_embeds}, or patch_size:" + f" {patch_size}. Make sure that `in_channels`, `num_vector_embeds` or `num_patches` is not None." + ) + + # 2. Define input layers + if self.is_input_continuous: + self.in_channels = in_channels + + self.norm = torch.nn.GroupNorm(num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True) + if use_linear_projection: + self.proj_in = linear_cls(in_channels, inner_dim) + else: + self.proj_in = conv_cls(in_channels, inner_dim, kernel_size=1, stride=1, padding=0) + elif self.is_input_vectorized: + assert sample_size is not None, "Transformer2DModel over discrete input must provide sample_size" + assert num_vector_embeds is not None, "Transformer2DModel over discrete input must provide num_embed" + + self.height = sample_size + self.width = sample_size + self.num_vector_embeds = num_vector_embeds + self.num_latent_pixels = self.height * self.width + + self.latent_image_embedding = ImagePositionalEmbeddings( + num_embed=num_vector_embeds, embed_dim=inner_dim, height=self.height, width=self.width + ) + elif self.is_input_patches: + assert sample_size is not None, "Transformer2DModel over patched input must provide sample_size" + + self.height = sample_size + self.width = sample_size + + self.patch_size = patch_size + interpolation_scale = self.config.sample_size // 64 # => 64 (= 512 pixart) has interpolation scale 1 + interpolation_scale = max(interpolation_scale, 1) + self.pos_embed = PatchEmbed( + height=sample_size, + width=sample_size, + patch_size=patch_size, + in_channels=in_channels, + embed_dim=inner_dim, + interpolation_scale=interpolation_scale, + ) + + # 3. Define transformers blocks + self.transformer_blocks = nn.ModuleList( + [ + BasicTransformerBlock( + inner_dim, + num_attention_heads, + attention_head_dim, + dropout=dropout, + cross_attention_dim=cross_attention_dim, + activation_fn=activation_fn, + num_embeds_ada_norm=num_embeds_ada_norm, + attention_bias=attention_bias, + only_cross_attention=only_cross_attention, + double_self_attention=double_self_attention, + upcast_attention=upcast_attention, + norm_type=norm_type, + norm_elementwise_affine=norm_elementwise_affine, + norm_eps=norm_eps, + attention_type=attention_type, + ) + for d in range(num_layers) + ] + ) + + # 4. Define output layers + self.out_channels = in_channels if out_channels is None else out_channels + if self.is_input_continuous: + # TODO: should use out_channels for continuous projections + if use_linear_projection: + self.proj_out = linear_cls(inner_dim, in_channels) + else: + self.proj_out = conv_cls(inner_dim, in_channels, kernel_size=1, stride=1, padding=0) + elif self.is_input_vectorized: + self.norm_out = nn.LayerNorm(inner_dim) + self.out = nn.Linear(inner_dim, self.num_vector_embeds - 1) + elif self.is_input_patches and norm_type != "ada_norm_single": + self.norm_out = nn.LayerNorm(inner_dim, elementwise_affine=False, eps=1e-6) + self.proj_out_1 = nn.Linear(inner_dim, 2 * inner_dim) + self.proj_out_2 = nn.Linear(inner_dim, patch_size * patch_size * self.out_channels) + elif self.is_input_patches and norm_type == "ada_norm_single": + self.norm_out = nn.LayerNorm(inner_dim, elementwise_affine=False, eps=1e-6) + self.scale_shift_table = nn.Parameter(torch.randn(2, inner_dim) / inner_dim**0.5) + self.proj_out = nn.Linear(inner_dim, patch_size * patch_size * self.out_channels) + + # 5. PixArt-Alpha blocks. + self.adaln_single = None + self.use_additional_conditions = False + if norm_type == "ada_norm_single": + self.use_additional_conditions = self.config.sample_size == 128 + # TODO(Sayak, PVP) clean this, for now we use sample size to determine whether to use + # additional conditions until we find better name + self.adaln_single = AdaLayerNormSingle(inner_dim, use_additional_conditions=self.use_additional_conditions) + + self.caption_projection = None + if caption_channels is not None: + self.caption_projection = PixArtAlphaTextProjection(in_features=caption_channels, hidden_size=inner_dim) + + self.gradient_checkpointing = False + + def _set_gradient_checkpointing(self, module, value=False): + if hasattr(module, "gradient_checkpointing"): + module.gradient_checkpointing = value + + def forward( + self, + hidden_states: torch.Tensor, + encoder_hidden_states: Optional[torch.Tensor] = None, + timestep: Optional[torch.LongTensor] = None, + added_cond_kwargs: Dict[str, torch.Tensor] = None, + class_labels: Optional[torch.LongTensor] = None, + cross_attention_kwargs: Dict[str, Any] = None, + attention_mask: Optional[torch.Tensor] = None, + encoder_attention_mask: Optional[torch.Tensor] = None, + return_dict: bool = True, + ): + """ + The [`Transformer2DModel`] forward method. + + Args: + hidden_states (`torch.LongTensor` of shape `(batch size, num latent pixels)` if discrete, `torch.FloatTensor` of shape `(batch size, channel, height, width)` if continuous): + Input `hidden_states`. + encoder_hidden_states ( `torch.FloatTensor` of shape `(batch size, sequence len, embed dims)`, *optional*): + Conditional embeddings for cross attention layer. If not given, cross-attention defaults to + self-attention. + timestep ( `torch.LongTensor`, *optional*): + Used to indicate denoising step. Optional timestep to be applied as an embedding in `AdaLayerNorm`. + class_labels ( `torch.LongTensor` of shape `(batch size, num classes)`, *optional*): + Used to indicate class labels conditioning. Optional class labels to be applied as an embedding in + `AdaLayerZeroNorm`. + cross_attention_kwargs ( `Dict[str, Any]`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + attention_mask ( `torch.Tensor`, *optional*): + An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask + is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large + negative values to the attention scores corresponding to "discard" tokens. + encoder_attention_mask ( `torch.Tensor`, *optional*): + Cross-attention mask applied to `encoder_hidden_states`. Two formats supported: + + * Mask `(batch, sequence_length)` True = keep, False = discard. + * Bias `(batch, 1, sequence_length)` 0 = keep, -10000 = discard. + + If `ndim == 2`: will be interpreted as a mask, then converted into a bias consistent with the format + above. This bias will be added to the cross-attention scores. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] instead of a plain + tuple. + + Returns: + If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a + `tuple` where the first element is the sample tensor. + """ + # ensure attention_mask is a bias, and give it a singleton query_tokens dimension. + # we may have done this conversion already, e.g. if we came here via UNet2DConditionModel#forward. + # we can tell by counting dims; if ndim == 2: it's a mask rather than a bias. + # expects mask of shape: + # [batch, key_tokens] + # adds singleton query_tokens dimension: + # [batch, 1, key_tokens] + # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes: + # [batch, heads, query_tokens, key_tokens] (e.g. torch sdp attn) + # [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn) + if attention_mask is not None and attention_mask.ndim == 2: + # assume that mask is expressed as: + # (1 = keep, 0 = discard) + # convert mask into a bias that can be added to attention scores: + # (keep = +0, discard = -10000.0) + attention_mask = (1 - attention_mask.to(hidden_states.dtype)) * -10000.0 + attention_mask = attention_mask.unsqueeze(1) + + # convert encoder_attention_mask to a bias the same way we do for attention_mask + if encoder_attention_mask is not None and encoder_attention_mask.ndim == 2: + encoder_attention_mask = (1 - encoder_attention_mask.to(hidden_states.dtype)) * -10000.0 + encoder_attention_mask = encoder_attention_mask.unsqueeze(1) + + # Retrieve lora scale. + lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0 + + # 1. Input + if self.is_input_continuous: + batch, _, height, width = hidden_states.shape + residual = hidden_states + + hidden_states = self.norm(hidden_states) + if not self.use_linear_projection: + hidden_states = ( + self.proj_in(hidden_states, scale=lora_scale) + if not USE_PEFT_BACKEND + else self.proj_in(hidden_states) + ) + inner_dim = hidden_states.shape[1] + hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * width, inner_dim) + else: + inner_dim = hidden_states.shape[1] + hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * width, inner_dim) + hidden_states = ( + self.proj_in(hidden_states, scale=lora_scale) + if not USE_PEFT_BACKEND + else self.proj_in(hidden_states) + ) + + elif self.is_input_vectorized: + hidden_states = self.latent_image_embedding(hidden_states) + elif self.is_input_patches: + height, width = hidden_states.shape[-2] // self.patch_size, hidden_states.shape[-1] // self.patch_size + hidden_states = self.pos_embed(hidden_states) + + if self.adaln_single is not None: + if self.use_additional_conditions and added_cond_kwargs is None: + raise ValueError( + "`added_cond_kwargs` cannot be None when using additional conditions for `adaln_single`." + ) + batch_size = hidden_states.shape[0] + timestep, embedded_timestep = self.adaln_single( + timestep, added_cond_kwargs, batch_size=batch_size, hidden_dtype=hidden_states.dtype + ) + + # 2. Blocks + if self.caption_projection is not None: + batch_size = hidden_states.shape[0] + encoder_hidden_states = self.caption_projection(encoder_hidden_states) + encoder_hidden_states = encoder_hidden_states.view(batch_size, -1, hidden_states.shape[-1]) + + for block in self.transformer_blocks: + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {} + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(block), + hidden_states, + attention_mask, + encoder_hidden_states, + encoder_attention_mask, + timestep, + cross_attention_kwargs, + class_labels, + **ckpt_kwargs, + ) + else: + hidden_states = block( + hidden_states, + attention_mask=attention_mask, + encoder_hidden_states=encoder_hidden_states, + encoder_attention_mask=encoder_attention_mask, + timestep=timestep, + cross_attention_kwargs=cross_attention_kwargs, + class_labels=class_labels, + ) + + # 3. Output + if self.is_input_continuous: + if not self.use_linear_projection: + hidden_states = hidden_states.reshape(batch, height, width, inner_dim).permute(0, 3, 1, 2).contiguous() + hidden_states = ( + self.proj_out(hidden_states, scale=lora_scale) + if not USE_PEFT_BACKEND + else self.proj_out(hidden_states) + ) + else: + hidden_states = ( + self.proj_out(hidden_states, scale=lora_scale) + if not USE_PEFT_BACKEND + else self.proj_out(hidden_states) + ) + hidden_states = hidden_states.reshape(batch, height, width, inner_dim).permute(0, 3, 1, 2).contiguous() + + output = hidden_states + residual + elif self.is_input_vectorized: + hidden_states = self.norm_out(hidden_states) + logits = self.out(hidden_states) + # (batch, self.num_vector_embeds - 1, self.num_latent_pixels) + logits = logits.permute(0, 2, 1) + + # log(p(x_0)) + output = F.log_softmax(logits.double(), dim=1).float() + + if self.is_input_patches: + if self.config.norm_type != "ada_norm_single": + conditioning = self.transformer_blocks[0].norm1.emb( + timestep, class_labels, hidden_dtype=hidden_states.dtype + ) + shift, scale = self.proj_out_1(F.silu(conditioning)).chunk(2, dim=1) + hidden_states = self.norm_out(hidden_states) * (1 + scale[:, None]) + shift[:, None] + hidden_states = self.proj_out_2(hidden_states) + elif self.config.norm_type == "ada_norm_single": + shift, scale = (self.scale_shift_table[None] + embedded_timestep[:, None]).chunk(2, dim=1) + hidden_states = self.norm_out(hidden_states) + # Modulation + hidden_states = hidden_states * (1 + scale) + shift + hidden_states = self.proj_out(hidden_states) + hidden_states = hidden_states.squeeze(1) + + # unpatchify + if self.adaln_single is None: + height = width = int(hidden_states.shape[1] ** 0.5) + hidden_states = hidden_states.reshape( + shape=(-1, height, width, self.patch_size, self.patch_size, self.out_channels) + ) + hidden_states = torch.einsum("nhwpqc->nchpwq", hidden_states) + output = hidden_states.reshape( + shape=(-1, self.out_channels, height * self.patch_size, width * self.patch_size) + ) + + if not return_dict: + return (output,) + + return Transformer2DModelOutput(sample=output) diff --git a/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/models/unets/__init__.py b/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/models/unets/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..1d3437072dac140d915d4e7922e1edbf31ed34de --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/models/unets/__init__.py @@ -0,0 +1 @@ +from .unet_2d_condition import UNet2DConditionModel \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/models/unets/unet_2d_blocks.py b/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/models/unets/unet_2d_blocks.py new file mode 100644 index 0000000000000000000000000000000000000000..b7d60b8bf0ce78a88f3f6cf29093e54812cc6cb4 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/models/unets/unet_2d_blocks.py @@ -0,0 +1,3592 @@ +# Copyright 2023 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from typing import Any, Dict, Optional, Tuple, Union + +import numpy as np +import torch +import torch.nn.functional as F +from torch import nn + +from diffusers.utils import is_torch_version, logging +from diffusers.utils.torch_utils import apply_freeu +from diffusers.models.resnet import ( + Downsample2D, + FirDownsample2D, + FirUpsample2D, + KDownsample2D, + KUpsample2D, + ResnetBlock2D, + ResnetBlockCondNorm2D, + Upsample2D, +) + +from stablediffusion.layers.activations import get_activation +from stablediffusion.layers.attention_processor import Attention, AttnAddedKVProcessor, AttnAddedKVProcessor2_0 +from stablediffusion.layers.normalization import AdaGroupNorm +from ..transformers.dual_transformer_2d import DualTransformer2DModel +from ..transformers.transformer_2d import Transformer2DModel + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +def get_down_block( + down_block_type: str, + num_layers: int, + in_channels: int, + out_channels: int, + temb_channels: int, + add_downsample: bool, + resnet_eps: float, + resnet_act_fn: str, + transformer_layers_per_block: int = 1, + num_attention_heads: Optional[int] = None, + resnet_groups: Optional[int] = None, + cross_attention_dim: Optional[int] = None, + downsample_padding: Optional[int] = None, + dual_cross_attention: bool = False, + use_linear_projection: bool = False, + only_cross_attention: bool = False, + upcast_attention: bool = False, + resnet_time_scale_shift: str = "default", + attention_type: str = "default", + resnet_skip_time_act: bool = False, + resnet_out_scale_factor: float = 1.0, + cross_attention_norm: Optional[str] = None, + attention_head_dim: Optional[int] = None, + downsample_type: Optional[str] = None, + dropout: float = 0.0, +): + # If attn head dim is not defined, we default it to the number of heads + if attention_head_dim is None: + logger.warn( + f"It is recommended to provide `attention_head_dim` when calling `get_down_block`. Defaulting `attention_head_dim` to {num_attention_heads}." + ) + attention_head_dim = num_attention_heads + + down_block_type = down_block_type[7:] if down_block_type.startswith("UNetRes") else down_block_type + if down_block_type == "DownBlock2D": + return DownBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + downsample_padding=downsample_padding, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + elif down_block_type == "ResnetDownsampleBlock2D": + return ResnetDownsampleBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + resnet_time_scale_shift=resnet_time_scale_shift, + skip_time_act=resnet_skip_time_act, + output_scale_factor=resnet_out_scale_factor, + ) + elif down_block_type == "AttnDownBlock2D": + if add_downsample is False: + downsample_type = None + else: + downsample_type = downsample_type or "conv" # default to 'conv' + return AttnDownBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + downsample_padding=downsample_padding, + attention_head_dim=attention_head_dim, + resnet_time_scale_shift=resnet_time_scale_shift, + downsample_type=downsample_type, + ) + elif down_block_type == "CrossAttnDownBlock2D": + if cross_attention_dim is None: + raise ValueError("cross_attention_dim must be specified for CrossAttnDownBlock2D") + return CrossAttnDownBlock2D( + num_layers=num_layers, + transformer_layers_per_block=transformer_layers_per_block, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + downsample_padding=downsample_padding, + cross_attention_dim=cross_attention_dim, + num_attention_heads=num_attention_heads, + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention, + upcast_attention=upcast_attention, + resnet_time_scale_shift=resnet_time_scale_shift, + attention_type=attention_type, + ) + elif down_block_type == "SimpleCrossAttnDownBlock2D": + if cross_attention_dim is None: + raise ValueError("cross_attention_dim must be specified for SimpleCrossAttnDownBlock2D") + return SimpleCrossAttnDownBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + cross_attention_dim=cross_attention_dim, + attention_head_dim=attention_head_dim, + resnet_time_scale_shift=resnet_time_scale_shift, + skip_time_act=resnet_skip_time_act, + output_scale_factor=resnet_out_scale_factor, + only_cross_attention=only_cross_attention, + cross_attention_norm=cross_attention_norm, + ) + elif down_block_type == "SkipDownBlock2D": + return SkipDownBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + downsample_padding=downsample_padding, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + elif down_block_type == "AttnSkipDownBlock2D": + return AttnSkipDownBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + attention_head_dim=attention_head_dim, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + elif down_block_type == "DownEncoderBlock2D": + return DownEncoderBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + downsample_padding=downsample_padding, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + elif down_block_type == "AttnDownEncoderBlock2D": + return AttnDownEncoderBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + downsample_padding=downsample_padding, + attention_head_dim=attention_head_dim, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + elif down_block_type == "KDownBlock2D": + return KDownBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + ) + elif down_block_type == "KCrossAttnDownBlock2D": + return KCrossAttnDownBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + cross_attention_dim=cross_attention_dim, + attention_head_dim=attention_head_dim, + add_self_attention=True if not add_downsample else False, + ) + raise ValueError(f"{down_block_type} does not exist.") + + +def get_up_block( + up_block_type: str, + num_layers: int, + in_channels: int, + out_channels: int, + prev_output_channel: int, + temb_channels: int, + add_upsample: bool, + resnet_eps: float, + resnet_act_fn: str, + resolution_idx: Optional[int] = None, + transformer_layers_per_block: int = 1, + num_attention_heads: Optional[int] = None, + resnet_groups: Optional[int] = None, + cross_attention_dim: Optional[int] = None, + dual_cross_attention: bool = False, + use_linear_projection: bool = False, + only_cross_attention: bool = False, + upcast_attention: bool = False, + resnet_time_scale_shift: str = "default", + attention_type: str = "default", + resnet_skip_time_act: bool = False, + resnet_out_scale_factor: float = 1.0, + cross_attention_norm: Optional[str] = None, + attention_head_dim: Optional[int] = None, + upsample_type: Optional[str] = None, + dropout: float = 0.0, +) -> nn.Module: + # If attn head dim is not defined, we default it to the number of heads + if attention_head_dim is None: + logger.warn( + f"It is recommended to provide `attention_head_dim` when calling `get_up_block`. Defaulting `attention_head_dim` to {num_attention_heads}." + ) + attention_head_dim = num_attention_heads + + up_block_type = up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type + if up_block_type == "UpBlock2D": + return UpBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + prev_output_channel=prev_output_channel, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + elif up_block_type == "ResnetUpsampleBlock2D": + return ResnetUpsampleBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + prev_output_channel=prev_output_channel, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + resnet_time_scale_shift=resnet_time_scale_shift, + skip_time_act=resnet_skip_time_act, + output_scale_factor=resnet_out_scale_factor, + ) + elif up_block_type == "CrossAttnUpBlock2D": + if cross_attention_dim is None: + raise ValueError("cross_attention_dim must be specified for CrossAttnUpBlock2D") + return CrossAttnUpBlock2D( + num_layers=num_layers, + transformer_layers_per_block=transformer_layers_per_block, + in_channels=in_channels, + out_channels=out_channels, + prev_output_channel=prev_output_channel, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + cross_attention_dim=cross_attention_dim, + num_attention_heads=num_attention_heads, + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention, + upcast_attention=upcast_attention, + resnet_time_scale_shift=resnet_time_scale_shift, + attention_type=attention_type, + ) + elif up_block_type == "SimpleCrossAttnUpBlock2D": + if cross_attention_dim is None: + raise ValueError("cross_attention_dim must be specified for SimpleCrossAttnUpBlock2D") + return SimpleCrossAttnUpBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + prev_output_channel=prev_output_channel, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + cross_attention_dim=cross_attention_dim, + attention_head_dim=attention_head_dim, + resnet_time_scale_shift=resnet_time_scale_shift, + skip_time_act=resnet_skip_time_act, + output_scale_factor=resnet_out_scale_factor, + only_cross_attention=only_cross_attention, + cross_attention_norm=cross_attention_norm, + ) + elif up_block_type == "AttnUpBlock2D": + if add_upsample is False: + upsample_type = None + else: + upsample_type = upsample_type or "conv" # default to 'conv' + + return AttnUpBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + prev_output_channel=prev_output_channel, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + attention_head_dim=attention_head_dim, + resnet_time_scale_shift=resnet_time_scale_shift, + upsample_type=upsample_type, + ) + elif up_block_type == "SkipUpBlock2D": + return SkipUpBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + prev_output_channel=prev_output_channel, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + elif up_block_type == "AttnSkipUpBlock2D": + return AttnSkipUpBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + prev_output_channel=prev_output_channel, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + attention_head_dim=attention_head_dim, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + elif up_block_type == "UpDecoderBlock2D": + return UpDecoderBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + resnet_time_scale_shift=resnet_time_scale_shift, + temb_channels=temb_channels, + ) + elif up_block_type == "AttnUpDecoderBlock2D": + return AttnUpDecoderBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + attention_head_dim=attention_head_dim, + resnet_time_scale_shift=resnet_time_scale_shift, + temb_channels=temb_channels, + ) + elif up_block_type == "KUpBlock2D": + return KUpBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + ) + elif up_block_type == "KCrossAttnUpBlock2D": + return KCrossAttnUpBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + cross_attention_dim=cross_attention_dim, + attention_head_dim=attention_head_dim, + ) + + raise ValueError(f"{up_block_type} does not exist.") + + +class AutoencoderTinyBlock(nn.Module): + """ + Tiny Autoencoder block used in [`AutoencoderTiny`]. It is a mini residual module consisting of plain conv + ReLU + blocks. + + Args: + in_channels (`int`): The number of input channels. + out_channels (`int`): The number of output channels. + act_fn (`str`): + ` The activation function to use. Supported values are `"swish"`, `"mish"`, `"gelu"`, and `"relu"`. + + Returns: + `torch.FloatTensor`: A tensor with the same shape as the input tensor, but with the number of channels equal to + `out_channels`. + """ + + def __init__(self, in_channels: int, out_channels: int, act_fn: str): + super().__init__() + act_fn = get_activation(act_fn) + self.conv = nn.Sequential( + nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1), + act_fn, + nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1), + act_fn, + nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1), + ) + self.skip = ( + nn.Conv2d(in_channels, out_channels, kernel_size=1, bias=False) + if in_channels != out_channels + else nn.Identity() + ) + self.fuse = nn.ReLU() + + def forward(self, x: torch.FloatTensor) -> torch.FloatTensor: + return self.fuse(self.conv(x) + self.skip(x)) + + +class UNetMidBlock2D(nn.Module): + """ + A 2D UNet mid-block [`UNetMidBlock2D`] with multiple residual blocks and optional attention blocks. + + Args: + in_channels (`int`): The number of input channels. + temb_channels (`int`): The number of temporal embedding channels. + dropout (`float`, *optional*, defaults to 0.0): The dropout rate. + num_layers (`int`, *optional*, defaults to 1): The number of residual blocks. + resnet_eps (`float`, *optional*, 1e-6 ): The epsilon value for the resnet blocks. + resnet_time_scale_shift (`str`, *optional*, defaults to `default`): + The type of normalization to apply to the time embeddings. This can help to improve the performance of the + model on tasks with long-range temporal dependencies. + resnet_act_fn (`str`, *optional*, defaults to `swish`): The activation function for the resnet blocks. + resnet_groups (`int`, *optional*, defaults to 32): + The number of groups to use in the group normalization layers of the resnet blocks. + attn_groups (`Optional[int]`, *optional*, defaults to None): The number of groups for the attention blocks. + resnet_pre_norm (`bool`, *optional*, defaults to `True`): + Whether to use pre-normalization for the resnet blocks. + add_attention (`bool`, *optional*, defaults to `True`): Whether to add attention blocks. + attention_head_dim (`int`, *optional*, defaults to 1): + Dimension of a single attention head. The number of attention heads is determined based on this value and + the number of input channels. + output_scale_factor (`float`, *optional*, defaults to 1.0): The output scale factor. + + Returns: + `torch.FloatTensor`: The output of the last residual block, which is a tensor of shape `(batch_size, + in_channels, height, width)`. + + """ + + def __init__( + self, + in_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", # default, spatial + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + attn_groups: Optional[int] = None, + resnet_pre_norm: bool = True, + add_attention: bool = True, + attention_head_dim: int = 1, + output_scale_factor: float = 1.0, + ): + super().__init__() + resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32) + self.add_attention = add_attention + + if attn_groups is None: + attn_groups = resnet_groups if resnet_time_scale_shift == "default" else None + + # there is always at least one resnet + if resnet_time_scale_shift == "spatial": + resnets = [ + ResnetBlockCondNorm2D( + in_channels=in_channels, + out_channels=in_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm="spatial", + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + ) + ] + else: + resnets = [ + ResnetBlock2D( + in_channels=in_channels, + out_channels=in_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ] + attentions = [] + + if attention_head_dim is None: + logger.warn( + f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {in_channels}." + ) + attention_head_dim = in_channels + + for _ in range(num_layers): + if self.add_attention: + attentions.append( + Attention( + in_channels, + heads=in_channels // attention_head_dim, + dim_head=attention_head_dim, + rescale_output_factor=output_scale_factor, + eps=resnet_eps, + norm_num_groups=attn_groups, + spatial_norm_dim=temb_channels if resnet_time_scale_shift == "spatial" else None, + residual_connection=True, + bias=True, + upcast_softmax=True, + _from_deprecated_attn_block=True, + ) + ) + else: + attentions.append(None) + + if resnet_time_scale_shift == "spatial": + resnets.append( + ResnetBlockCondNorm2D( + in_channels=in_channels, + out_channels=in_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm="spatial", + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + ) + ) + else: + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=in_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + def forward(self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None) -> torch.FloatTensor: + hidden_states = self.resnets[0](hidden_states, temb) + for attn, resnet in zip(self.attentions, self.resnets[1:]): + if attn is not None: + hidden_states = attn(hidden_states, temb=temb) + hidden_states = resnet(hidden_states, temb) + + return hidden_states + + +class UNetMidBlock2DCrossAttn(nn.Module): + def __init__( + self, + in_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + transformer_layers_per_block: Union[int, Tuple[int]] = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + num_attention_heads: int = 1, + output_scale_factor: float = 1.0, + cross_attention_dim: int = 1280, + dual_cross_attention: bool = False, + use_linear_projection: bool = False, + upcast_attention: bool = False, + attention_type: str = "default", + ): + super().__init__() + + self.has_cross_attention = True + self.num_attention_heads = num_attention_heads + resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32) + + # support for variable transformer layers per block + if isinstance(transformer_layers_per_block, int): + transformer_layers_per_block = [transformer_layers_per_block] * num_layers + + # there is always at least one resnet + resnets = [ + ResnetBlock2D( + in_channels=in_channels, + out_channels=in_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ] + attentions = [] + + for i in range(num_layers): + if not dual_cross_attention: + attentions.append( + Transformer2DModel( + num_attention_heads, + in_channels // num_attention_heads, + in_channels=in_channels, + num_layers=transformer_layers_per_block[i], + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + use_linear_projection=use_linear_projection, + upcast_attention=upcast_attention, + attention_type=attention_type, + ) + ) + else: + attentions.append( + DualTransformer2DModel( + num_attention_heads, + in_channels // num_attention_heads, + in_channels=in_channels, + num_layers=1, + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + ) + ) + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=in_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + self.gradient_checkpointing = False + + def forward( + self, + hidden_states: torch.FloatTensor, + temb: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + ) -> torch.FloatTensor: + lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0 + hidden_states = self.resnets[0](hidden_states, temb, scale=lora_scale) + for attn, resnet in zip(self.attentions, self.resnets[1:]): + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {} + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + return_dict=False, + )[0] + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), + hidden_states, + temb, + **ckpt_kwargs, + ) + else: + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + return_dict=False, + )[0] + hidden_states = resnet(hidden_states, temb, scale=lora_scale) + + return hidden_states + + +class UNetMidBlock2DSimpleCrossAttn(nn.Module): + def __init__( + self, + in_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + output_scale_factor: float = 1.0, + cross_attention_dim: int = 1280, + skip_time_act: bool = False, + only_cross_attention: bool = False, + cross_attention_norm: Optional[str] = None, + ): + super().__init__() + + self.has_cross_attention = True + + self.attention_head_dim = attention_head_dim + resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32) + + self.num_heads = in_channels // self.attention_head_dim + + # there is always at least one resnet + resnets = [ + ResnetBlock2D( + in_channels=in_channels, + out_channels=in_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + ) + ] + attentions = [] + + for _ in range(num_layers): + processor = ( + AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor() + ) + + attentions.append( + Attention( + query_dim=in_channels, + cross_attention_dim=in_channels, + heads=self.num_heads, + dim_head=self.attention_head_dim, + added_kv_proj_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + bias=True, + upcast_softmax=True, + only_cross_attention=only_cross_attention, + cross_attention_norm=cross_attention_norm, + processor=processor, + ) + ) + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=in_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + ) + ) + + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + def forward( + self, + hidden_states: torch.FloatTensor, + temb: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + ) -> torch.FloatTensor: + cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {} + lora_scale = cross_attention_kwargs.get("scale", 1.0) + + if attention_mask is None: + # if encoder_hidden_states is defined: we are doing cross-attn, so we should use cross-attn mask. + mask = None if encoder_hidden_states is None else encoder_attention_mask + else: + # when attention_mask is defined: we don't even check for encoder_attention_mask. + # this is to maintain compatibility with UnCLIP, which uses 'attention_mask' param for cross-attn masks. + # TODO: UnCLIP should express cross-attn mask via encoder_attention_mask param instead of via attention_mask. + # then we can simplify this whole if/else block to: + # mask = attention_mask if encoder_hidden_states is None else encoder_attention_mask + mask = attention_mask + + hidden_states = self.resnets[0](hidden_states, temb, scale=lora_scale) + for attn, resnet in zip(self.attentions, self.resnets[1:]): + # attn + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=mask, + **cross_attention_kwargs, + ) + + # resnet + hidden_states = resnet(hidden_states, temb, scale=lora_scale) + + return hidden_states + + +class AttnDownBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + output_scale_factor: float = 1.0, + downsample_padding: int = 1, + downsample_type: str = "conv", + ): + super().__init__() + resnets = [] + attentions = [] + self.downsample_type = downsample_type + + if attention_head_dim is None: + logger.warn( + f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {out_channels}." + ) + attention_head_dim = out_channels + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + attentions.append( + Attention( + out_channels, + heads=out_channels // attention_head_dim, + dim_head=attention_head_dim, + rescale_output_factor=output_scale_factor, + eps=resnet_eps, + norm_num_groups=resnet_groups, + residual_connection=True, + bias=True, + upcast_softmax=True, + _from_deprecated_attn_block=True, + ) + ) + + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if downsample_type == "conv": + self.downsamplers = nn.ModuleList( + [ + Downsample2D( + out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op" + ) + ] + ) + elif downsample_type == "resnet": + self.downsamplers = nn.ModuleList( + [ + ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + down=True, + ) + ] + ) + else: + self.downsamplers = None + + def forward( + self, + hidden_states: torch.FloatTensor, + temb: Optional[torch.FloatTensor] = None, + upsample_size: Optional[int] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]: + cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {} + + lora_scale = cross_attention_kwargs.get("scale", 1.0) + + output_states = () + + for resnet, attn in zip(self.resnets, self.attentions): + cross_attention_kwargs.update({"scale": lora_scale}) + hidden_states = resnet(hidden_states, temb, scale=lora_scale) + hidden_states = attn(hidden_states, **cross_attention_kwargs) + output_states = output_states + (hidden_states,) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + if self.downsample_type == "resnet": + hidden_states = downsampler(hidden_states, temb=temb, scale=lora_scale) + else: + hidden_states = downsampler(hidden_states, scale=lora_scale) + + output_states += (hidden_states,) + + return hidden_states, output_states + + +class CrossAttnDownBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + transformer_layers_per_block: Union[int, Tuple[int]] = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + num_attention_heads: int = 1, + cross_attention_dim: int = 1280, + output_scale_factor: float = 1.0, + downsample_padding: int = 1, + add_downsample: bool = True, + dual_cross_attention: bool = False, + use_linear_projection: bool = False, + only_cross_attention: bool = False, + upcast_attention: bool = False, + attention_type: str = "default", + ): + super().__init__() + resnets = [] + attentions = [] + + self.has_cross_attention = True + self.num_attention_heads = num_attention_heads + if isinstance(transformer_layers_per_block, int): + transformer_layers_per_block = [transformer_layers_per_block] * num_layers + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + if not dual_cross_attention: + attentions.append( + Transformer2DModel( + num_attention_heads, + out_channels // num_attention_heads, + in_channels=out_channels, + num_layers=transformer_layers_per_block[i], + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention, + upcast_attention=upcast_attention, + attention_type=attention_type, + ) + ) + else: + attentions.append( + DualTransformer2DModel( + num_attention_heads, + out_channels // num_attention_heads, + in_channels=out_channels, + num_layers=1, + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + ) + ) + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if add_downsample: + self.downsamplers = nn.ModuleList( + [ + Downsample2D( + out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op" + ) + ] + ) + else: + self.downsamplers = None + + self.gradient_checkpointing = False + + def forward( + self, + hidden_states: torch.FloatTensor, + temb: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + additional_residuals: Optional[torch.FloatTensor] = None, + ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]: + output_states = () + + lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0 + + blocks = list(zip(self.resnets, self.attentions)) + + for i, (resnet, attn) in enumerate(blocks): + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {} + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), + hidden_states, + temb, + **ckpt_kwargs, + ) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + return_dict=False, + )[0] + else: + hidden_states = resnet(hidden_states, temb, scale=lora_scale) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + return_dict=False, + )[0] + + # apply additional residuals to the output of the last pair of resnet and attention blocks + if i == len(blocks) - 1 and additional_residuals is not None: + hidden_states = hidden_states + additional_residuals + + output_states = output_states + (hidden_states,) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states, scale=lora_scale) + + output_states = output_states + (hidden_states,) + + return hidden_states, output_states + + +class DownBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + output_scale_factor: float = 1.0, + add_downsample: bool = True, + downsample_padding: int = 1, + ): + super().__init__() + resnets = [] + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + self.resnets = nn.ModuleList(resnets) + + if add_downsample: + self.downsamplers = nn.ModuleList( + [ + Downsample2D( + out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op" + ) + ] + ) + else: + self.downsamplers = None + + self.gradient_checkpointing = False + + def forward( + self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0 + ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]: + output_states = () + + for resnet in self.resnets: + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module): + def custom_forward(*inputs): + return module(*inputs) + + return custom_forward + + if is_torch_version(">=", "1.11.0"): + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb, use_reentrant=False + ) + else: + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb + ) + else: + hidden_states = resnet(hidden_states, temb, scale=scale) + + output_states = output_states + (hidden_states,) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states, scale=scale) + + output_states = output_states + (hidden_states,) + + return hidden_states, output_states + + +class DownEncoderBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + output_scale_factor: float = 1.0, + add_downsample: bool = True, + downsample_padding: int = 1, + ): + super().__init__() + resnets = [] + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + if resnet_time_scale_shift == "spatial": + resnets.append( + ResnetBlockCondNorm2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=None, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm="spatial", + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + ) + ) + else: + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=None, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + self.resnets = nn.ModuleList(resnets) + + if add_downsample: + self.downsamplers = nn.ModuleList( + [ + Downsample2D( + out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op" + ) + ] + ) + else: + self.downsamplers = None + + def forward(self, hidden_states: torch.FloatTensor, scale: float = 1.0) -> torch.FloatTensor: + for resnet in self.resnets: + hidden_states = resnet(hidden_states, temb=None, scale=scale) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states, scale) + + return hidden_states + + +class AttnDownEncoderBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + output_scale_factor: float = 1.0, + add_downsample: bool = True, + downsample_padding: int = 1, + ): + super().__init__() + resnets = [] + attentions = [] + + if attention_head_dim is None: + logger.warn( + f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {out_channels}." + ) + attention_head_dim = out_channels + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + if resnet_time_scale_shift == "spatial": + resnets.append( + ResnetBlockCondNorm2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=None, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm="spatial", + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + ) + ) + else: + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=None, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + attentions.append( + Attention( + out_channels, + heads=out_channels // attention_head_dim, + dim_head=attention_head_dim, + rescale_output_factor=output_scale_factor, + eps=resnet_eps, + norm_num_groups=resnet_groups, + residual_connection=True, + bias=True, + upcast_softmax=True, + _from_deprecated_attn_block=True, + ) + ) + + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if add_downsample: + self.downsamplers = nn.ModuleList( + [ + Downsample2D( + out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op" + ) + ] + ) + else: + self.downsamplers = None + + def forward(self, hidden_states: torch.FloatTensor, scale: float = 1.0) -> torch.FloatTensor: + for resnet, attn in zip(self.resnets, self.attentions): + hidden_states = resnet(hidden_states, temb=None, scale=scale) + cross_attention_kwargs = {"scale": scale} + hidden_states = attn(hidden_states, **cross_attention_kwargs) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states, scale) + + return hidden_states + + +class AttnSkipDownBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + output_scale_factor: float = np.sqrt(2.0), + add_downsample: bool = True, + ): + super().__init__() + self.attentions = nn.ModuleList([]) + self.resnets = nn.ModuleList([]) + + if attention_head_dim is None: + logger.warn( + f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {out_channels}." + ) + attention_head_dim = out_channels + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + self.resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=min(in_channels // 4, 32), + groups_out=min(out_channels // 4, 32), + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + self.attentions.append( + Attention( + out_channels, + heads=out_channels // attention_head_dim, + dim_head=attention_head_dim, + rescale_output_factor=output_scale_factor, + eps=resnet_eps, + norm_num_groups=32, + residual_connection=True, + bias=True, + upcast_softmax=True, + _from_deprecated_attn_block=True, + ) + ) + + if add_downsample: + self.resnet_down = ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=min(out_channels // 4, 32), + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + use_in_shortcut=True, + down=True, + kernel="fir", + ) + self.downsamplers = nn.ModuleList([FirDownsample2D(out_channels, out_channels=out_channels)]) + self.skip_conv = nn.Conv2d(3, out_channels, kernel_size=(1, 1), stride=(1, 1)) + else: + self.resnet_down = None + self.downsamplers = None + self.skip_conv = None + + def forward( + self, + hidden_states: torch.FloatTensor, + temb: Optional[torch.FloatTensor] = None, + skip_sample: Optional[torch.FloatTensor] = None, + scale: float = 1.0, + ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...], torch.FloatTensor]: + output_states = () + + for resnet, attn in zip(self.resnets, self.attentions): + hidden_states = resnet(hidden_states, temb, scale=scale) + cross_attention_kwargs = {"scale": scale} + hidden_states = attn(hidden_states, **cross_attention_kwargs) + output_states += (hidden_states,) + + if self.downsamplers is not None: + hidden_states = self.resnet_down(hidden_states, temb, scale=scale) + for downsampler in self.downsamplers: + skip_sample = downsampler(skip_sample) + + hidden_states = self.skip_conv(skip_sample) + hidden_states + + output_states += (hidden_states,) + + return hidden_states, output_states, skip_sample + + +class SkipDownBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_pre_norm: bool = True, + output_scale_factor: float = np.sqrt(2.0), + add_downsample: bool = True, + downsample_padding: int = 1, + ): + super().__init__() + self.resnets = nn.ModuleList([]) + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + self.resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=min(in_channels // 4, 32), + groups_out=min(out_channels // 4, 32), + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + if add_downsample: + self.resnet_down = ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=min(out_channels // 4, 32), + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + use_in_shortcut=True, + down=True, + kernel="fir", + ) + self.downsamplers = nn.ModuleList([FirDownsample2D(out_channels, out_channels=out_channels)]) + self.skip_conv = nn.Conv2d(3, out_channels, kernel_size=(1, 1), stride=(1, 1)) + else: + self.resnet_down = None + self.downsamplers = None + self.skip_conv = None + + def forward( + self, + hidden_states: torch.FloatTensor, + temb: Optional[torch.FloatTensor] = None, + skip_sample: Optional[torch.FloatTensor] = None, + scale: float = 1.0, + ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...], torch.FloatTensor]: + output_states = () + + for resnet in self.resnets: + hidden_states = resnet(hidden_states, temb, scale) + output_states += (hidden_states,) + + if self.downsamplers is not None: + hidden_states = self.resnet_down(hidden_states, temb, scale) + for downsampler in self.downsamplers: + skip_sample = downsampler(skip_sample) + + hidden_states = self.skip_conv(skip_sample) + hidden_states + + output_states += (hidden_states,) + + return hidden_states, output_states, skip_sample + + +class ResnetDownsampleBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + output_scale_factor: float = 1.0, + add_downsample: bool = True, + skip_time_act: bool = False, + ): + super().__init__() + resnets = [] + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + ) + ) + + self.resnets = nn.ModuleList(resnets) + + if add_downsample: + self.downsamplers = nn.ModuleList( + [ + ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + down=True, + ) + ] + ) + else: + self.downsamplers = None + + self.gradient_checkpointing = False + + def forward( + self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0 + ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]: + output_states = () + + for resnet in self.resnets: + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module): + def custom_forward(*inputs): + return module(*inputs) + + return custom_forward + + if is_torch_version(">=", "1.11.0"): + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb, use_reentrant=False + ) + else: + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb + ) + else: + hidden_states = resnet(hidden_states, temb, scale) + + output_states = output_states + (hidden_states,) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states, temb, scale) + + output_states = output_states + (hidden_states,) + + return hidden_states, output_states + + +class SimpleCrossAttnDownBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + cross_attention_dim: int = 1280, + output_scale_factor: float = 1.0, + add_downsample: bool = True, + skip_time_act: bool = False, + only_cross_attention: bool = False, + cross_attention_norm: Optional[str] = None, + ): + super().__init__() + + self.has_cross_attention = True + + resnets = [] + attentions = [] + + self.attention_head_dim = attention_head_dim + self.num_heads = out_channels // self.attention_head_dim + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + ) + ) + + processor = ( + AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor() + ) + + attentions.append( + Attention( + query_dim=out_channels, + cross_attention_dim=out_channels, + heads=self.num_heads, + dim_head=attention_head_dim, + added_kv_proj_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + bias=True, + upcast_softmax=True, + only_cross_attention=only_cross_attention, + cross_attention_norm=cross_attention_norm, + processor=processor, + ) + ) + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if add_downsample: + self.downsamplers = nn.ModuleList( + [ + ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + down=True, + ) + ] + ) + else: + self.downsamplers = None + + self.gradient_checkpointing = False + + def forward( + self, + hidden_states: torch.FloatTensor, + temb: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]: + output_states = () + cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {} + + lora_scale = cross_attention_kwargs.get("scale", 1.0) + + if attention_mask is None: + # if encoder_hidden_states is defined: we are doing cross-attn, so we should use cross-attn mask. + mask = None if encoder_hidden_states is None else encoder_attention_mask + else: + # when attention_mask is defined: we don't even check for encoder_attention_mask. + # this is to maintain compatibility with UnCLIP, which uses 'attention_mask' param for cross-attn masks. + # TODO: UnCLIP should express cross-attn mask via encoder_attention_mask param instead of via attention_mask. + # then we can simplify this whole if/else block to: + # mask = attention_mask if encoder_hidden_states is None else encoder_attention_mask + mask = attention_mask + + for resnet, attn in zip(self.resnets, self.attentions): + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=mask, + **cross_attention_kwargs, + ) + else: + hidden_states = resnet(hidden_states, temb, scale=lora_scale) + + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=mask, + **cross_attention_kwargs, + ) + + output_states = output_states + (hidden_states,) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states, temb, scale=lora_scale) + + output_states = output_states + (hidden_states,) + + return hidden_states, output_states + + +class KDownBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 4, + resnet_eps: float = 1e-5, + resnet_act_fn: str = "gelu", + resnet_group_size: int = 32, + add_downsample: bool = False, + ): + super().__init__() + resnets = [] + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + groups = in_channels // resnet_group_size + groups_out = out_channels // resnet_group_size + + resnets.append( + ResnetBlockCondNorm2D( + in_channels=in_channels, + out_channels=out_channels, + dropout=dropout, + temb_channels=temb_channels, + groups=groups, + groups_out=groups_out, + eps=resnet_eps, + non_linearity=resnet_act_fn, + time_embedding_norm="ada_group", + conv_shortcut_bias=False, + ) + ) + + self.resnets = nn.ModuleList(resnets) + + if add_downsample: + # YiYi's comments- might be able to use FirDownsample2D, look into details later + self.downsamplers = nn.ModuleList([KDownsample2D()]) + else: + self.downsamplers = None + + self.gradient_checkpointing = False + + def forward( + self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0 + ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]: + output_states = () + + for resnet in self.resnets: + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module): + def custom_forward(*inputs): + return module(*inputs) + + return custom_forward + + if is_torch_version(">=", "1.11.0"): + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb, use_reentrant=False + ) + else: + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb + ) + else: + hidden_states = resnet(hidden_states, temb, scale) + + output_states += (hidden_states,) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states) + + return hidden_states, output_states + + +class KCrossAttnDownBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + cross_attention_dim: int, + dropout: float = 0.0, + num_layers: int = 4, + resnet_group_size: int = 32, + add_downsample: bool = True, + attention_head_dim: int = 64, + add_self_attention: bool = False, + resnet_eps: float = 1e-5, + resnet_act_fn: str = "gelu", + ): + super().__init__() + resnets = [] + attentions = [] + + self.has_cross_attention = True + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + groups = in_channels // resnet_group_size + groups_out = out_channels // resnet_group_size + + resnets.append( + ResnetBlockCondNorm2D( + in_channels=in_channels, + out_channels=out_channels, + dropout=dropout, + temb_channels=temb_channels, + groups=groups, + groups_out=groups_out, + eps=resnet_eps, + non_linearity=resnet_act_fn, + time_embedding_norm="ada_group", + conv_shortcut_bias=False, + ) + ) + attentions.append( + KAttentionBlock( + out_channels, + out_channels // attention_head_dim, + attention_head_dim, + cross_attention_dim=cross_attention_dim, + temb_channels=temb_channels, + attention_bias=True, + add_self_attention=add_self_attention, + cross_attention_norm="layer_norm", + group_size=resnet_group_size, + ) + ) + + self.resnets = nn.ModuleList(resnets) + self.attentions = nn.ModuleList(attentions) + + if add_downsample: + self.downsamplers = nn.ModuleList([KDownsample2D()]) + else: + self.downsamplers = None + + self.gradient_checkpointing = False + + def forward( + self, + hidden_states: torch.FloatTensor, + temb: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]: + output_states = () + lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0 + + for resnet, attn in zip(self.resnets, self.attentions): + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {} + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), + hidden_states, + temb, + **ckpt_kwargs, + ) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + emb=temb, + attention_mask=attention_mask, + cross_attention_kwargs=cross_attention_kwargs, + encoder_attention_mask=encoder_attention_mask, + ) + else: + hidden_states = resnet(hidden_states, temb, scale=lora_scale) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + emb=temb, + attention_mask=attention_mask, + cross_attention_kwargs=cross_attention_kwargs, + encoder_attention_mask=encoder_attention_mask, + ) + + if self.downsamplers is None: + output_states += (None,) + else: + output_states += (hidden_states,) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states) + + return hidden_states, output_states + + +class AttnUpBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + prev_output_channel: int, + out_channels: int, + temb_channels: int, + resolution_idx: int = None, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + output_scale_factor: float = 1.0, + upsample_type: str = "conv", + ): + super().__init__() + resnets = [] + attentions = [] + + self.upsample_type = upsample_type + + if attention_head_dim is None: + logger.warn( + f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {out_channels}." + ) + attention_head_dim = out_channels + + for i in range(num_layers): + res_skip_channels = in_channels if (i == num_layers - 1) else out_channels + resnet_in_channels = prev_output_channel if i == 0 else out_channels + + resnets.append( + ResnetBlock2D( + in_channels=resnet_in_channels + res_skip_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + attentions.append( + Attention( + out_channels, + heads=out_channels // attention_head_dim, + dim_head=attention_head_dim, + rescale_output_factor=output_scale_factor, + eps=resnet_eps, + norm_num_groups=resnet_groups, + residual_connection=True, + bias=True, + upcast_softmax=True, + _from_deprecated_attn_block=True, + ) + ) + + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if upsample_type == "conv": + self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)]) + elif upsample_type == "resnet": + self.upsamplers = nn.ModuleList( + [ + ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + up=True, + ) + ] + ) + else: + self.upsamplers = None + + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + upsample_size: Optional[int] = None, + scale: float = 1.0, + ) -> torch.FloatTensor: + for resnet, attn in zip(self.resnets, self.attentions): + # pop res hidden states + res_hidden_states = res_hidden_states_tuple[-1] + res_hidden_states_tuple = res_hidden_states_tuple[:-1] + hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1) + + hidden_states = resnet(hidden_states, temb, scale=scale) + cross_attention_kwargs = {"scale": scale} + hidden_states = attn(hidden_states, **cross_attention_kwargs) + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + if self.upsample_type == "resnet": + hidden_states = upsampler(hidden_states, temb=temb, scale=scale) + else: + hidden_states = upsampler(hidden_states, scale=scale) + + return hidden_states + + +class CrossAttnUpBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + prev_output_channel: int, + temb_channels: int, + resolution_idx: Optional[int] = None, + dropout: float = 0.0, + num_layers: int = 1, + transformer_layers_per_block: Union[int, Tuple[int]] = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + num_attention_heads: int = 1, + cross_attention_dim: int = 1280, + output_scale_factor: float = 1.0, + add_upsample: bool = True, + dual_cross_attention: bool = False, + use_linear_projection: bool = False, + only_cross_attention: bool = False, + upcast_attention: bool = False, + attention_type: str = "default", + ): + super().__init__() + resnets = [] + attentions = [] + + self.has_cross_attention = True + self.num_attention_heads = num_attention_heads + + if isinstance(transformer_layers_per_block, int): + transformer_layers_per_block = [transformer_layers_per_block] * num_layers + + for i in range(num_layers): + res_skip_channels = in_channels if (i == num_layers - 1) else out_channels + resnet_in_channels = prev_output_channel if i == 0 else out_channels + + resnets.append( + ResnetBlock2D( + in_channels=resnet_in_channels + res_skip_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + if not dual_cross_attention: + attentions.append( + Transformer2DModel( + num_attention_heads, + out_channels // num_attention_heads, + in_channels=out_channels, + num_layers=transformer_layers_per_block[i], + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention, + upcast_attention=upcast_attention, + attention_type=attention_type, + ) + ) + else: + attentions.append( + DualTransformer2DModel( + num_attention_heads, + out_channels // num_attention_heads, + in_channels=out_channels, + num_layers=1, + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + ) + ) + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if add_upsample: + self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)]) + else: + self.upsamplers = None + + self.gradient_checkpointing = False + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + upsample_size: Optional[int] = None, + attention_mask: Optional[torch.FloatTensor] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + ) -> torch.FloatTensor: + lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0 + is_freeu_enabled = ( + getattr(self, "s1", None) + and getattr(self, "s2", None) + and getattr(self, "b1", None) + and getattr(self, "b2", None) + ) + + for resnet, attn in zip(self.resnets, self.attentions): + # pop res hidden states + res_hidden_states = res_hidden_states_tuple[-1] + res_hidden_states_tuple = res_hidden_states_tuple[:-1] + + # FreeU: Only operate on the first two stages + if is_freeu_enabled: + hidden_states, res_hidden_states = apply_freeu( + self.resolution_idx, + hidden_states, + res_hidden_states, + s1=self.s1, + s2=self.s2, + b1=self.b1, + b2=self.b2, + ) + + hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1) + + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {} + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), + hidden_states, + temb, + **ckpt_kwargs, + ) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + return_dict=False, + )[0] + else: + hidden_states = resnet(hidden_states, temb, scale=lora_scale) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + return_dict=False, + )[0] + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states, upsample_size, scale=lora_scale) + + return hidden_states + + +class UpBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + prev_output_channel: int, + out_channels: int, + temb_channels: int, + resolution_idx: Optional[int] = None, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + output_scale_factor: float = 1.0, + add_upsample: bool = True, + ): + super().__init__() + resnets = [] + + for i in range(num_layers): + res_skip_channels = in_channels if (i == num_layers - 1) else out_channels + resnet_in_channels = prev_output_channel if i == 0 else out_channels + + resnets.append( + ResnetBlock2D( + in_channels=resnet_in_channels + res_skip_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + self.resnets = nn.ModuleList(resnets) + + if add_upsample: + self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)]) + else: + self.upsamplers = None + + self.gradient_checkpointing = False + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + upsample_size: Optional[int] = None, + scale: float = 1.0, + ) -> torch.FloatTensor: + is_freeu_enabled = ( + getattr(self, "s1", None) + and getattr(self, "s2", None) + and getattr(self, "b1", None) + and getattr(self, "b2", None) + ) + + for resnet in self.resnets: + # pop res hidden states + res_hidden_states = res_hidden_states_tuple[-1] + res_hidden_states_tuple = res_hidden_states_tuple[:-1] + + # FreeU: Only operate on the first two stages + if is_freeu_enabled: + hidden_states, res_hidden_states = apply_freeu( + self.resolution_idx, + hidden_states, + res_hidden_states, + s1=self.s1, + s2=self.s2, + b1=self.b1, + b2=self.b2, + ) + + hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1) + + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module): + def custom_forward(*inputs): + return module(*inputs) + + return custom_forward + + if is_torch_version(">=", "1.11.0"): + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb, use_reentrant=False + ) + else: + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb + ) + else: + hidden_states = resnet(hidden_states, temb, scale=scale) + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states, upsample_size, scale=scale) + + return hidden_states + + +class UpDecoderBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + resolution_idx: Optional[int] = None, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", # default, spatial + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + output_scale_factor: float = 1.0, + add_upsample: bool = True, + temb_channels: Optional[int] = None, + ): + super().__init__() + resnets = [] + + for i in range(num_layers): + input_channels = in_channels if i == 0 else out_channels + + if resnet_time_scale_shift == "spatial": + resnets.append( + ResnetBlockCondNorm2D( + in_channels=input_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm="spatial", + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + ) + ) + else: + resnets.append( + ResnetBlock2D( + in_channels=input_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + self.resnets = nn.ModuleList(resnets) + + if add_upsample: + self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)]) + else: + self.upsamplers = None + + self.resolution_idx = resolution_idx + + def forward( + self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0 + ) -> torch.FloatTensor: + for resnet in self.resnets: + hidden_states = resnet(hidden_states, temb=temb, scale=scale) + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states) + + return hidden_states + + +class AttnUpDecoderBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + resolution_idx: Optional[int] = None, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + output_scale_factor: float = 1.0, + add_upsample: bool = True, + temb_channels: Optional[int] = None, + ): + super().__init__() + resnets = [] + attentions = [] + + if attention_head_dim is None: + logger.warn( + f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `out_channels`: {out_channels}." + ) + attention_head_dim = out_channels + + for i in range(num_layers): + input_channels = in_channels if i == 0 else out_channels + + if resnet_time_scale_shift == "spatial": + resnets.append( + ResnetBlockCondNorm2D( + in_channels=input_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm="spatial", + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + ) + ) + else: + resnets.append( + ResnetBlock2D( + in_channels=input_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + attentions.append( + Attention( + out_channels, + heads=out_channels // attention_head_dim, + dim_head=attention_head_dim, + rescale_output_factor=output_scale_factor, + eps=resnet_eps, + norm_num_groups=resnet_groups if resnet_time_scale_shift != "spatial" else None, + spatial_norm_dim=temb_channels if resnet_time_scale_shift == "spatial" else None, + residual_connection=True, + bias=True, + upcast_softmax=True, + _from_deprecated_attn_block=True, + ) + ) + + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if add_upsample: + self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)]) + else: + self.upsamplers = None + + self.resolution_idx = resolution_idx + + def forward( + self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0 + ) -> torch.FloatTensor: + for resnet, attn in zip(self.resnets, self.attentions): + hidden_states = resnet(hidden_states, temb=temb, scale=scale) + cross_attention_kwargs = {"scale": scale} + hidden_states = attn(hidden_states, temb=temb, **cross_attention_kwargs) + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states, scale=scale) + + return hidden_states + + +class AttnSkipUpBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + prev_output_channel: int, + out_channels: int, + temb_channels: int, + resolution_idx: Optional[int] = None, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + output_scale_factor: float = np.sqrt(2.0), + add_upsample: bool = True, + ): + super().__init__() + self.attentions = nn.ModuleList([]) + self.resnets = nn.ModuleList([]) + + for i in range(num_layers): + res_skip_channels = in_channels if (i == num_layers - 1) else out_channels + resnet_in_channels = prev_output_channel if i == 0 else out_channels + + self.resnets.append( + ResnetBlock2D( + in_channels=resnet_in_channels + res_skip_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=min(resnet_in_channels + res_skip_channels // 4, 32), + groups_out=min(out_channels // 4, 32), + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + if attention_head_dim is None: + logger.warn( + f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `out_channels`: {out_channels}." + ) + attention_head_dim = out_channels + + self.attentions.append( + Attention( + out_channels, + heads=out_channels // attention_head_dim, + dim_head=attention_head_dim, + rescale_output_factor=output_scale_factor, + eps=resnet_eps, + norm_num_groups=32, + residual_connection=True, + bias=True, + upcast_softmax=True, + _from_deprecated_attn_block=True, + ) + ) + + self.upsampler = FirUpsample2D(in_channels, out_channels=out_channels) + if add_upsample: + self.resnet_up = ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=min(out_channels // 4, 32), + groups_out=min(out_channels // 4, 32), + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + use_in_shortcut=True, + up=True, + kernel="fir", + ) + self.skip_conv = nn.Conv2d(out_channels, 3, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) + self.skip_norm = torch.nn.GroupNorm( + num_groups=min(out_channels // 4, 32), num_channels=out_channels, eps=resnet_eps, affine=True + ) + self.act = nn.SiLU() + else: + self.resnet_up = None + self.skip_conv = None + self.skip_norm = None + self.act = None + + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + skip_sample=None, + scale: float = 1.0, + ) -> Tuple[torch.FloatTensor, torch.FloatTensor]: + for resnet in self.resnets: + # pop res hidden states + res_hidden_states = res_hidden_states_tuple[-1] + res_hidden_states_tuple = res_hidden_states_tuple[:-1] + hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1) + + hidden_states = resnet(hidden_states, temb, scale=scale) + + cross_attention_kwargs = {"scale": scale} + hidden_states = self.attentions[0](hidden_states, **cross_attention_kwargs) + + if skip_sample is not None: + skip_sample = self.upsampler(skip_sample) + else: + skip_sample = 0 + + if self.resnet_up is not None: + skip_sample_states = self.skip_norm(hidden_states) + skip_sample_states = self.act(skip_sample_states) + skip_sample_states = self.skip_conv(skip_sample_states) + + skip_sample = skip_sample + skip_sample_states + + hidden_states = self.resnet_up(hidden_states, temb, scale=scale) + + return hidden_states, skip_sample + + +class SkipUpBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + prev_output_channel: int, + out_channels: int, + temb_channels: int, + resolution_idx: Optional[int] = None, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_pre_norm: bool = True, + output_scale_factor: float = np.sqrt(2.0), + add_upsample: bool = True, + upsample_padding: int = 1, + ): + super().__init__() + self.resnets = nn.ModuleList([]) + + for i in range(num_layers): + res_skip_channels = in_channels if (i == num_layers - 1) else out_channels + resnet_in_channels = prev_output_channel if i == 0 else out_channels + + self.resnets.append( + ResnetBlock2D( + in_channels=resnet_in_channels + res_skip_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=min((resnet_in_channels + res_skip_channels) // 4, 32), + groups_out=min(out_channels // 4, 32), + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + self.upsampler = FirUpsample2D(in_channels, out_channels=out_channels) + if add_upsample: + self.resnet_up = ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=min(out_channels // 4, 32), + groups_out=min(out_channels // 4, 32), + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + use_in_shortcut=True, + up=True, + kernel="fir", + ) + self.skip_conv = nn.Conv2d(out_channels, 3, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) + self.skip_norm = torch.nn.GroupNorm( + num_groups=min(out_channels // 4, 32), num_channels=out_channels, eps=resnet_eps, affine=True + ) + self.act = nn.SiLU() + else: + self.resnet_up = None + self.skip_conv = None + self.skip_norm = None + self.act = None + + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + skip_sample=None, + scale: float = 1.0, + ) -> Tuple[torch.FloatTensor, torch.FloatTensor]: + for resnet in self.resnets: + # pop res hidden states + res_hidden_states = res_hidden_states_tuple[-1] + res_hidden_states_tuple = res_hidden_states_tuple[:-1] + hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1) + + hidden_states = resnet(hidden_states, temb, scale=scale) + + if skip_sample is not None: + skip_sample = self.upsampler(skip_sample) + else: + skip_sample = 0 + + if self.resnet_up is not None: + skip_sample_states = self.skip_norm(hidden_states) + skip_sample_states = self.act(skip_sample_states) + skip_sample_states = self.skip_conv(skip_sample_states) + + skip_sample = skip_sample + skip_sample_states + + hidden_states = self.resnet_up(hidden_states, temb, scale=scale) + + return hidden_states, skip_sample + + +class ResnetUpsampleBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + prev_output_channel: int, + out_channels: int, + temb_channels: int, + resolution_idx: Optional[int] = None, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + output_scale_factor: float = 1.0, + add_upsample: bool = True, + skip_time_act: bool = False, + ): + super().__init__() + resnets = [] + + for i in range(num_layers): + res_skip_channels = in_channels if (i == num_layers - 1) else out_channels + resnet_in_channels = prev_output_channel if i == 0 else out_channels + + resnets.append( + ResnetBlock2D( + in_channels=resnet_in_channels + res_skip_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + ) + ) + + self.resnets = nn.ModuleList(resnets) + + if add_upsample: + self.upsamplers = nn.ModuleList( + [ + ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + up=True, + ) + ] + ) + else: + self.upsamplers = None + + self.gradient_checkpointing = False + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + upsample_size: Optional[int] = None, + scale: float = 1.0, + ) -> torch.FloatTensor: + for resnet in self.resnets: + # pop res hidden states + res_hidden_states = res_hidden_states_tuple[-1] + res_hidden_states_tuple = res_hidden_states_tuple[:-1] + hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1) + + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module): + def custom_forward(*inputs): + return module(*inputs) + + return custom_forward + + if is_torch_version(">=", "1.11.0"): + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb, use_reentrant=False + ) + else: + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb + ) + else: + hidden_states = resnet(hidden_states, temb, scale=scale) + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states, temb, scale=scale) + + return hidden_states + + +class SimpleCrossAttnUpBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + prev_output_channel: int, + temb_channels: int, + resolution_idx: Optional[int] = None, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + cross_attention_dim: int = 1280, + output_scale_factor: float = 1.0, + add_upsample: bool = True, + skip_time_act: bool = False, + only_cross_attention: bool = False, + cross_attention_norm: Optional[str] = None, + ): + super().__init__() + resnets = [] + attentions = [] + + self.has_cross_attention = True + self.attention_head_dim = attention_head_dim + + self.num_heads = out_channels // self.attention_head_dim + + for i in range(num_layers): + res_skip_channels = in_channels if (i == num_layers - 1) else out_channels + resnet_in_channels = prev_output_channel if i == 0 else out_channels + + resnets.append( + ResnetBlock2D( + in_channels=resnet_in_channels + res_skip_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + ) + ) + + processor = ( + AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor() + ) + + attentions.append( + Attention( + query_dim=out_channels, + cross_attention_dim=out_channels, + heads=self.num_heads, + dim_head=self.attention_head_dim, + added_kv_proj_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + bias=True, + upcast_softmax=True, + only_cross_attention=only_cross_attention, + cross_attention_norm=cross_attention_norm, + processor=processor, + ) + ) + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if add_upsample: + self.upsamplers = nn.ModuleList( + [ + ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + up=True, + ) + ] + ) + else: + self.upsamplers = None + + self.gradient_checkpointing = False + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + upsample_size: Optional[int] = None, + attention_mask: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + ) -> torch.FloatTensor: + cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {} + + lora_scale = cross_attention_kwargs.get("scale", 1.0) + if attention_mask is None: + # if encoder_hidden_states is defined: we are doing cross-attn, so we should use cross-attn mask. + mask = None if encoder_hidden_states is None else encoder_attention_mask + else: + # when attention_mask is defined: we don't even check for encoder_attention_mask. + # this is to maintain compatibility with UnCLIP, which uses 'attention_mask' param for cross-attn masks. + # TODO: UnCLIP should express cross-attn mask via encoder_attention_mask param instead of via attention_mask. + # then we can simplify this whole if/else block to: + # mask = attention_mask if encoder_hidden_states is None else encoder_attention_mask + mask = attention_mask + + for resnet, attn in zip(self.resnets, self.attentions): + # resnet + # pop res hidden states + res_hidden_states = res_hidden_states_tuple[-1] + res_hidden_states_tuple = res_hidden_states_tuple[:-1] + hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1) + + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=mask, + **cross_attention_kwargs, + ) + else: + hidden_states = resnet(hidden_states, temb, scale=lora_scale) + + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=mask, + **cross_attention_kwargs, + ) + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states, temb, scale=lora_scale) + + return hidden_states + + +class KUpBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + resolution_idx: int, + dropout: float = 0.0, + num_layers: int = 5, + resnet_eps: float = 1e-5, + resnet_act_fn: str = "gelu", + resnet_group_size: Optional[int] = 32, + add_upsample: bool = True, + ): + super().__init__() + resnets = [] + k_in_channels = 2 * out_channels + k_out_channels = in_channels + num_layers = num_layers - 1 + + for i in range(num_layers): + in_channels = k_in_channels if i == 0 else out_channels + groups = in_channels // resnet_group_size + groups_out = out_channels // resnet_group_size + + resnets.append( + ResnetBlockCondNorm2D( + in_channels=in_channels, + out_channels=k_out_channels if (i == num_layers - 1) else out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=groups, + groups_out=groups_out, + dropout=dropout, + non_linearity=resnet_act_fn, + time_embedding_norm="ada_group", + conv_shortcut_bias=False, + ) + ) + + self.resnets = nn.ModuleList(resnets) + + if add_upsample: + self.upsamplers = nn.ModuleList([KUpsample2D()]) + else: + self.upsamplers = None + + self.gradient_checkpointing = False + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + upsample_size: Optional[int] = None, + scale: float = 1.0, + ) -> torch.FloatTensor: + res_hidden_states_tuple = res_hidden_states_tuple[-1] + if res_hidden_states_tuple is not None: + hidden_states = torch.cat([hidden_states, res_hidden_states_tuple], dim=1) + + for resnet in self.resnets: + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module): + def custom_forward(*inputs): + return module(*inputs) + + return custom_forward + + if is_torch_version(">=", "1.11.0"): + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb, use_reentrant=False + ) + else: + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb + ) + else: + hidden_states = resnet(hidden_states, temb, scale=scale) + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states) + + return hidden_states + + +class KCrossAttnUpBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + resolution_idx: int, + dropout: float = 0.0, + num_layers: int = 4, + resnet_eps: float = 1e-5, + resnet_act_fn: str = "gelu", + resnet_group_size: int = 32, + attention_head_dim: int = 1, # attention dim_head + cross_attention_dim: int = 768, + add_upsample: bool = True, + upcast_attention: bool = False, + ): + super().__init__() + resnets = [] + attentions = [] + + is_first_block = in_channels == out_channels == temb_channels + is_middle_block = in_channels != out_channels + add_self_attention = True if is_first_block else False + + self.has_cross_attention = True + self.attention_head_dim = attention_head_dim + + # in_channels, and out_channels for the block (k-unet) + k_in_channels = out_channels if is_first_block else 2 * out_channels + k_out_channels = in_channels + + num_layers = num_layers - 1 + + for i in range(num_layers): + in_channels = k_in_channels if i == 0 else out_channels + groups = in_channels // resnet_group_size + groups_out = out_channels // resnet_group_size + + if is_middle_block and (i == num_layers - 1): + conv_2d_out_channels = k_out_channels + else: + conv_2d_out_channels = None + + resnets.append( + ResnetBlockCondNorm2D( + in_channels=in_channels, + out_channels=out_channels, + conv_2d_out_channels=conv_2d_out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=groups, + groups_out=groups_out, + dropout=dropout, + non_linearity=resnet_act_fn, + time_embedding_norm="ada_group", + conv_shortcut_bias=False, + ) + ) + attentions.append( + KAttentionBlock( + k_out_channels if (i == num_layers - 1) else out_channels, + k_out_channels // attention_head_dim + if (i == num_layers - 1) + else out_channels // attention_head_dim, + attention_head_dim, + cross_attention_dim=cross_attention_dim, + temb_channels=temb_channels, + attention_bias=True, + add_self_attention=add_self_attention, + cross_attention_norm="layer_norm", + upcast_attention=upcast_attention, + ) + ) + + self.resnets = nn.ModuleList(resnets) + self.attentions = nn.ModuleList(attentions) + + if add_upsample: + self.upsamplers = nn.ModuleList([KUpsample2D()]) + else: + self.upsamplers = None + + self.gradient_checkpointing = False + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + upsample_size: Optional[int] = None, + attention_mask: Optional[torch.FloatTensor] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + ) -> torch.FloatTensor: + res_hidden_states_tuple = res_hidden_states_tuple[-1] + if res_hidden_states_tuple is not None: + hidden_states = torch.cat([hidden_states, res_hidden_states_tuple], dim=1) + + lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0 + for resnet, attn in zip(self.resnets, self.attentions): + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {} + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), + hidden_states, + temb, + **ckpt_kwargs, + ) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + emb=temb, + attention_mask=attention_mask, + cross_attention_kwargs=cross_attention_kwargs, + encoder_attention_mask=encoder_attention_mask, + ) + else: + hidden_states = resnet(hidden_states, temb, scale=lora_scale) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + emb=temb, + attention_mask=attention_mask, + cross_attention_kwargs=cross_attention_kwargs, + encoder_attention_mask=encoder_attention_mask, + ) + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states) + + return hidden_states + + +# can potentially later be renamed to `No-feed-forward` attention +class KAttentionBlock(nn.Module): + r""" + A basic Transformer block. + + Parameters: + dim (`int`): The number of channels in the input and output. + num_attention_heads (`int`): The number of heads to use for multi-head attention. + attention_head_dim (`int`): The number of channels in each head. + dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. + cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention. + attention_bias (`bool`, *optional*, defaults to `False`): + Configure if the attention layers should contain a bias parameter. + upcast_attention (`bool`, *optional*, defaults to `False`): + Set to `True` to upcast the attention computation to `float32`. + temb_channels (`int`, *optional*, defaults to 768): + The number of channels in the token embedding. + add_self_attention (`bool`, *optional*, defaults to `False`): + Set to `True` to add self-attention to the block. + cross_attention_norm (`str`, *optional*, defaults to `None`): + The type of normalization to use for the cross attention. Can be `None`, `layer_norm`, or `group_norm`. + group_size (`int`, *optional*, defaults to 32): + The number of groups to separate the channels into for group normalization. + """ + + def __init__( + self, + dim: int, + num_attention_heads: int, + attention_head_dim: int, + dropout: float = 0.0, + cross_attention_dim: Optional[int] = None, + attention_bias: bool = False, + upcast_attention: bool = False, + temb_channels: int = 768, # for ada_group_norm + add_self_attention: bool = False, + cross_attention_norm: Optional[str] = None, + group_size: int = 32, + ): + super().__init__() + self.add_self_attention = add_self_attention + + # 1. Self-Attn + if add_self_attention: + self.norm1 = AdaGroupNorm(temb_channels, dim, max(1, dim // group_size)) + self.attn1 = Attention( + query_dim=dim, + heads=num_attention_heads, + dim_head=attention_head_dim, + dropout=dropout, + bias=attention_bias, + cross_attention_dim=None, + cross_attention_norm=None, + ) + + # 2. Cross-Attn + self.norm2 = AdaGroupNorm(temb_channels, dim, max(1, dim // group_size)) + self.attn2 = Attention( + query_dim=dim, + cross_attention_dim=cross_attention_dim, + heads=num_attention_heads, + dim_head=attention_head_dim, + dropout=dropout, + bias=attention_bias, + upcast_attention=upcast_attention, + cross_attention_norm=cross_attention_norm, + ) + + def _to_3d(self, hidden_states: torch.FloatTensor, height: int, weight: int) -> torch.FloatTensor: + return hidden_states.permute(0, 2, 3, 1).reshape(hidden_states.shape[0], height * weight, -1) + + def _to_4d(self, hidden_states: torch.FloatTensor, height: int, weight: int) -> torch.FloatTensor: + return hidden_states.permute(0, 2, 1).reshape(hidden_states.shape[0], -1, height, weight) + + def forward( + self, + hidden_states: torch.FloatTensor, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + # TODO: mark emb as non-optional (self.norm2 requires it). + # requires assessing impact of change to positional param interface. + emb: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + ) -> torch.FloatTensor: + cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {} + + # 1. Self-Attention + if self.add_self_attention: + norm_hidden_states = self.norm1(hidden_states, emb) + + height, weight = norm_hidden_states.shape[2:] + norm_hidden_states = self._to_3d(norm_hidden_states, height, weight) + + attn_output = self.attn1( + norm_hidden_states, + encoder_hidden_states=None, + attention_mask=attention_mask, + **cross_attention_kwargs, + ) + attn_output = self._to_4d(attn_output, height, weight) + + hidden_states = attn_output + hidden_states + + # 2. Cross-Attention/None + norm_hidden_states = self.norm2(hidden_states, emb) + + height, weight = norm_hidden_states.shape[2:] + norm_hidden_states = self._to_3d(norm_hidden_states, height, weight) + attn_output = self.attn2( + norm_hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=attention_mask if encoder_hidden_states is None else encoder_attention_mask, + **cross_attention_kwargs, + ) + attn_output = self._to_4d(attn_output, height, weight) + + hidden_states = attn_output + hidden_states + + return hidden_states diff --git a/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/models/unets/unet_2d_condition.py b/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/models/unets/unet_2d_condition.py new file mode 100644 index 0000000000000000000000000000000000000000..bbe078d297cf740e02198ff0a8b02133b5d7b00b --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/models/unets/unet_2d_condition.py @@ -0,0 +1,1218 @@ +# Copyright 2023 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from dataclasses import dataclass +from typing import Any, Dict, List, Optional, Tuple, Union + +import torch +import torch.nn as nn +import torch.utils.checkpoint + +from diffusers.configuration_utils import ConfigMixin, register_to_config +from diffusers.loaders import PeftAdapterMixin, UNet2DConditionLoadersMixin +from diffusers.utils import USE_PEFT_BACKEND, BaseOutput, deprecate, logging, scale_lora_layers, unscale_lora_layers +from stablediffusion.layers.activations import get_activation +from stablediffusion.layers.attention_processor import ( + ADDED_KV_ATTENTION_PROCESSORS, + CROSS_ATTENTION_PROCESSORS, + Attention, + AttentionProcessor, + AttnAddedKVProcessor, + AttnProcessor, +) +from stablediffusion.layers.embeddings import ( + GaussianFourierProjection, + GLIGENTextBoundingboxProjection, + ImageHintTimeEmbedding, + ImageProjection, + ImageTimeEmbedding, + TextImageProjection, + TextImageTimeEmbedding, + TextTimeEmbedding, + TimestepEmbedding, + Timesteps, +) +from .unet_2d_blocks import ( + UNetMidBlock2D, + UNetMidBlock2DCrossAttn, + UNetMidBlock2DSimpleCrossAttn, + get_down_block, + get_up_block, +) +from stablediffusion.models.model_utils import ModelMixin + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +@dataclass +class UNet2DConditionOutput(BaseOutput): + """ + The output of [`UNet2DConditionModel`]. + + Args: + sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`): + The hidden states output conditioned on `encoder_hidden_states` input. Output of last layer of model. + """ + + sample: torch.FloatTensor = None + + +class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin, PeftAdapterMixin): + r""" + A conditional 2D UNet model that takes a noisy sample, conditional state, and a timestep and returns a sample + shaped output. + + This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented + for all models (such as downloading or saving). + + Parameters: + sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`): + Height and width of input/output sample. + in_channels (`int`, *optional*, defaults to 4): Number of channels in the input sample. + out_channels (`int`, *optional*, defaults to 4): Number of channels in the output. + center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample. + flip_sin_to_cos (`bool`, *optional*, defaults to `False`): + Whether to flip the sin to cos in the time embedding. + freq_shift (`int`, *optional*, defaults to 0): The frequency shift to apply to the time embedding. + down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`): + The tuple of downsample blocks to use. + mid_block_type (`str`, *optional*, defaults to `"UNetMidBlock2DCrossAttn"`): + Block type for middle of UNet, it can be one of `UNetMidBlock2DCrossAttn`, `UNetMidBlock2D`, or + `UNetMidBlock2DSimpleCrossAttn`. If `None`, the mid block layer is skipped. + up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D")`): + The tuple of upsample blocks to use. + only_cross_attention(`bool` or `Tuple[bool]`, *optional*, default to `False`): + Whether to include self-attention in the basic transformer blocks, see + [`~models.attention.BasicTransformerBlock`]. + block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`): + The tuple of output channels for each block. + layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block. + downsample_padding (`int`, *optional*, defaults to 1): The padding to use for the downsampling convolution. + mid_block_scale_factor (`float`, *optional*, defaults to 1.0): The scale factor to use for the mid block. + dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. + act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use. + norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for the normalization. + If `None`, normalization and activation layers is skipped in post-processing. + norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization. + cross_attention_dim (`int` or `Tuple[int]`, *optional*, defaults to 1280): + The dimension of the cross attention features. + transformer_layers_per_block (`int`, `Tuple[int]`, or `Tuple[Tuple]` , *optional*, defaults to 1): + The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for + [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`], + [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`]. + reverse_transformer_layers_per_block : (`Tuple[Tuple]`, *optional*, defaults to None): + The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`], in the upsampling + blocks of the U-Net. Only relevant if `transformer_layers_per_block` is of type `Tuple[Tuple]` and for + [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`], + [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`]. + encoder_hid_dim (`int`, *optional*, defaults to None): + If `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim` + dimension to `cross_attention_dim`. + encoder_hid_dim_type (`str`, *optional*, defaults to `None`): + If given, the `encoder_hidden_states` and potentially other embeddings are down-projected to text + embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`. + attention_head_dim (`int`, *optional*, defaults to 8): The dimension of the attention heads. + num_attention_heads (`int`, *optional*): + The number of attention heads. If not defined, defaults to `attention_head_dim` + resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config + for ResNet blocks (see [`~models.resnet.ResnetBlock2D`]). Choose from `default` or `scale_shift`. + class_embed_type (`str`, *optional*, defaults to `None`): + The type of class embedding to use which is ultimately summed with the time embeddings. Choose from `None`, + `"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`. + addition_embed_type (`str`, *optional*, defaults to `None`): + Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or + "text". "text" will use the `TextTimeEmbedding` layer. + addition_time_embed_dim: (`int`, *optional*, defaults to `None`): + Dimension for the timestep embeddings. + num_class_embeds (`int`, *optional*, defaults to `None`): + Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing + class conditioning with `class_embed_type` equal to `None`. + time_embedding_type (`str`, *optional*, defaults to `positional`): + The type of position embedding to use for timesteps. Choose from `positional` or `fourier`. + time_embedding_dim (`int`, *optional*, defaults to `None`): + An optional override for the dimension of the projected time embedding. + time_embedding_act_fn (`str`, *optional*, defaults to `None`): + Optional activation function to use only once on the time embeddings before they are passed to the rest of + the UNet. Choose from `silu`, `mish`, `gelu`, and `swish`. + timestep_post_act (`str`, *optional*, defaults to `None`): + The second activation function to use in timestep embedding. Choose from `silu`, `mish` and `gelu`. + time_cond_proj_dim (`int`, *optional*, defaults to `None`): + The dimension of `cond_proj` layer in the timestep embedding. + conv_in_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_in` layer. conv_out_kernel (`int`, + *optional*, default to `3`): The kernel size of `conv_out` layer. projection_class_embeddings_input_dim (`int`, + *optional*): The dimension of the `class_labels` input when + `class_embed_type="projection"`. Required when `class_embed_type="projection"`. + class_embeddings_concat (`bool`, *optional*, defaults to `False`): Whether to concatenate the time + embeddings with the class embeddings. + mid_block_only_cross_attention (`bool`, *optional*, defaults to `None`): + Whether to use cross attention with the mid block when using the `UNetMidBlock2DSimpleCrossAttn`. If + `only_cross_attention` is given as a single boolean and `mid_block_only_cross_attention` is `None`, the + `only_cross_attention` value is used as the value for `mid_block_only_cross_attention`. Default to `False` + otherwise. + """ + + _supports_gradient_checkpointing = True + + @register_to_config + def __init__( + self, + sample_size: Optional[int] = None, + in_channels: int = 4, + out_channels: int = 4, + center_input_sample: bool = False, + flip_sin_to_cos: bool = True, + freq_shift: int = 0, + down_block_types: Tuple[str] = ( + "CrossAttnDownBlock2D", + "CrossAttnDownBlock2D", + "CrossAttnDownBlock2D", + "DownBlock2D", + ), + mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn", + up_block_types: Tuple[str] = ("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D"), + only_cross_attention: Union[bool, Tuple[bool]] = False, + block_out_channels: Tuple[int] = (320, 640, 1280, 1280), + layers_per_block: Union[int, Tuple[int]] = 2, + downsample_padding: int = 1, + mid_block_scale_factor: float = 1, + dropout: float = 0.0, + act_fn: str = "silu", + norm_num_groups: Optional[int] = 32, + norm_eps: float = 1e-5, + cross_attention_dim: Union[int, Tuple[int]] = 1280, + transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple]] = 1, + reverse_transformer_layers_per_block: Optional[Tuple[Tuple[int]]] = None, + encoder_hid_dim: Optional[int] = None, + encoder_hid_dim_type: Optional[str] = None, + attention_head_dim: Union[int, Tuple[int]] = 8, + num_attention_heads: Optional[Union[int, Tuple[int]]] = None, + dual_cross_attention: bool = False, + use_linear_projection: bool = False, + class_embed_type: Optional[str] = None, + addition_embed_type: Optional[str] = None, + addition_time_embed_dim: Optional[int] = None, + num_class_embeds: Optional[int] = None, + upcast_attention: bool = False, + resnet_time_scale_shift: str = "default", + resnet_skip_time_act: bool = False, + resnet_out_scale_factor: int = 1.0, + time_embedding_type: str = "positional", + time_embedding_dim: Optional[int] = None, + time_embedding_act_fn: Optional[str] = None, + timestep_post_act: Optional[str] = None, + time_cond_proj_dim: Optional[int] = None, + conv_in_kernel: int = 3, + conv_out_kernel: int = 3, + projection_class_embeddings_input_dim: Optional[int] = None, + attention_type: str = "default", + class_embeddings_concat: bool = False, + mid_block_only_cross_attention: Optional[bool] = None, + cross_attention_norm: Optional[str] = None, + addition_embed_type_num_heads=64, + ): + super().__init__() + + self.sample_size = sample_size + + if num_attention_heads is not None: + raise ValueError( + "At the moment it is not possible to define the number of attention heads via `num_attention_heads` because of a naming issue as described in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131. Passing `num_attention_heads` will only be supported in diffusers v0.19." + ) + + # If `num_attention_heads` is not defined (which is the case for most models) + # it will default to `attention_head_dim`. This looks weird upon first reading it and it is. + # The reason for this behavior is to correct for incorrectly named variables that were introduced + # when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131 + # Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking + # which is why we correct for the naming here. + num_attention_heads = num_attention_heads or attention_head_dim + + # Check inputs + if len(down_block_types) != len(up_block_types): + raise ValueError( + f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}." + ) + + if len(block_out_channels) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}." + ) + + if not isinstance(only_cross_attention, bool) and len(only_cross_attention) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `only_cross_attention` as `down_block_types`. `only_cross_attention`: {only_cross_attention}. `down_block_types`: {down_block_types}." + ) + + if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}." + ) + + if not isinstance(attention_head_dim, int) and len(attention_head_dim) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `attention_head_dim` as `down_block_types`. `attention_head_dim`: {attention_head_dim}. `down_block_types`: {down_block_types}." + ) + + if isinstance(cross_attention_dim, list) and len(cross_attention_dim) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `cross_attention_dim` as `down_block_types`. `cross_attention_dim`: {cross_attention_dim}. `down_block_types`: {down_block_types}." + ) + + if not isinstance(layers_per_block, int) and len(layers_per_block) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `layers_per_block` as `down_block_types`. `layers_per_block`: {layers_per_block}. `down_block_types`: {down_block_types}." + ) + if isinstance(transformer_layers_per_block, list) and reverse_transformer_layers_per_block is None: + for layer_number_per_block in transformer_layers_per_block: + if isinstance(layer_number_per_block, list): + raise ValueError("Must provide 'reverse_transformer_layers_per_block` if using asymmetrical UNet.") + + # input + conv_in_padding = (conv_in_kernel - 1) // 2 + self.conv_in = nn.Conv2d( + in_channels, block_out_channels[0], kernel_size=conv_in_kernel, padding=conv_in_padding + ) + + # time + if time_embedding_type == "fourier": + time_embed_dim = time_embedding_dim or block_out_channels[0] * 2 + if time_embed_dim % 2 != 0: + raise ValueError(f"`time_embed_dim` should be divisible by 2, but is {time_embed_dim}.") + self.time_proj = GaussianFourierProjection( + time_embed_dim // 2, set_W_to_weight=False, log=False, flip_sin_to_cos=flip_sin_to_cos + ) + timestep_input_dim = time_embed_dim + elif time_embedding_type == "positional": + time_embed_dim = time_embedding_dim or block_out_channels[0] * 4 + + self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift) + timestep_input_dim = block_out_channels[0] + else: + raise ValueError( + f"{time_embedding_type} does not exist. Please make sure to use one of `fourier` or `positional`." + ) + + self.time_embedding = TimestepEmbedding( + timestep_input_dim, + time_embed_dim, + act_fn=act_fn, + post_act_fn=timestep_post_act, + cond_proj_dim=time_cond_proj_dim, + ) + + if encoder_hid_dim_type is None and encoder_hid_dim is not None: + encoder_hid_dim_type = "text_proj" + self.register_to_config(encoder_hid_dim_type=encoder_hid_dim_type) + logger.info("encoder_hid_dim_type defaults to 'text_proj' as `encoder_hid_dim` is defined.") + + if encoder_hid_dim is None and encoder_hid_dim_type is not None: + raise ValueError( + f"`encoder_hid_dim` has to be defined when `encoder_hid_dim_type` is set to {encoder_hid_dim_type}." + ) + + if encoder_hid_dim_type == "text_proj": + self.encoder_hid_proj = nn.Linear(encoder_hid_dim, cross_attention_dim) + elif encoder_hid_dim_type == "text_image_proj": + # image_embed_dim DOESN'T have to be `cross_attention_dim`. To not clutter the __init__ too much + # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use + # case when `addition_embed_type == "text_image_proj"` (Kadinsky 2.1)` + self.encoder_hid_proj = TextImageProjection( + text_embed_dim=encoder_hid_dim, + image_embed_dim=cross_attention_dim, + cross_attention_dim=cross_attention_dim, + ) + elif encoder_hid_dim_type == "image_proj": + # Kandinsky 2.2 + self.encoder_hid_proj = ImageProjection( + image_embed_dim=encoder_hid_dim, + cross_attention_dim=cross_attention_dim, + ) + elif encoder_hid_dim_type is not None: + raise ValueError( + f"encoder_hid_dim_type: {encoder_hid_dim_type} must be None, 'text_proj' or 'text_image_proj'." + ) + else: + self.encoder_hid_proj = None + + # class embedding + if class_embed_type is None and num_class_embeds is not None: + self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim) + elif class_embed_type == "timestep": + self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim, act_fn=act_fn) + elif class_embed_type == "identity": + self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim) + elif class_embed_type == "projection": + if projection_class_embeddings_input_dim is None: + raise ValueError( + "`class_embed_type`: 'projection' requires `projection_class_embeddings_input_dim` be set" + ) + # The projection `class_embed_type` is the same as the timestep `class_embed_type` except + # 1. the `class_labels` inputs are not first converted to sinusoidal embeddings + # 2. it projects from an arbitrary input dimension. + # + # Note that `TimestepEmbedding` is quite general, being mainly linear layers and activations. + # When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings. + # As a result, `TimestepEmbedding` can be passed arbitrary vectors. + self.class_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim) + elif class_embed_type == "simple_projection": + if projection_class_embeddings_input_dim is None: + raise ValueError( + "`class_embed_type`: 'simple_projection' requires `projection_class_embeddings_input_dim` be set" + ) + self.class_embedding = nn.Linear(projection_class_embeddings_input_dim, time_embed_dim) + else: + self.class_embedding = None + + if addition_embed_type == "text": + if encoder_hid_dim is not None: + text_time_embedding_from_dim = encoder_hid_dim + else: + text_time_embedding_from_dim = cross_attention_dim + + self.add_embedding = TextTimeEmbedding( + text_time_embedding_from_dim, time_embed_dim, num_heads=addition_embed_type_num_heads + ) + elif addition_embed_type == "text_image": + # text_embed_dim and image_embed_dim DON'T have to be `cross_attention_dim`. To not clutter the __init__ too much + # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use + # case when `addition_embed_type == "text_image"` (Kadinsky 2.1)` + self.add_embedding = TextImageTimeEmbedding( + text_embed_dim=cross_attention_dim, image_embed_dim=cross_attention_dim, time_embed_dim=time_embed_dim + ) + elif addition_embed_type == "text_time": + self.add_time_proj = Timesteps(addition_time_embed_dim, flip_sin_to_cos, freq_shift) + self.add_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim) + elif addition_embed_type == "image": + # Kandinsky 2.2 + self.add_embedding = ImageTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim) + elif addition_embed_type == "image_hint": + # Kandinsky 2.2 ControlNet + self.add_embedding = ImageHintTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim) + elif addition_embed_type is not None: + raise ValueError(f"addition_embed_type: {addition_embed_type} must be None, 'text' or 'text_image'.") + + if time_embedding_act_fn is None: + self.time_embed_act = None + else: + self.time_embed_act = get_activation(time_embedding_act_fn) + + self.down_blocks = nn.ModuleList([]) + self.up_blocks = nn.ModuleList([]) + + if isinstance(only_cross_attention, bool): + if mid_block_only_cross_attention is None: + mid_block_only_cross_attention = only_cross_attention + + only_cross_attention = [only_cross_attention] * len(down_block_types) + + if mid_block_only_cross_attention is None: + mid_block_only_cross_attention = False + + if isinstance(num_attention_heads, int): + num_attention_heads = (num_attention_heads,) * len(down_block_types) + + if isinstance(attention_head_dim, int): + attention_head_dim = (attention_head_dim,) * len(down_block_types) + + if isinstance(cross_attention_dim, int): + cross_attention_dim = (cross_attention_dim,) * len(down_block_types) + + if isinstance(layers_per_block, int): + layers_per_block = [layers_per_block] * len(down_block_types) + + if isinstance(transformer_layers_per_block, int): + transformer_layers_per_block = [transformer_layers_per_block] * len(down_block_types) + + if class_embeddings_concat: + # The time embeddings are concatenated with the class embeddings. The dimension of the + # time embeddings passed to the down, middle, and up blocks is twice the dimension of the + # regular time embeddings + blocks_time_embed_dim = time_embed_dim * 2 + else: + blocks_time_embed_dim = time_embed_dim + + # down + output_channel = block_out_channels[0] + for i, down_block_type in enumerate(down_block_types): + input_channel = output_channel + output_channel = block_out_channels[i] + is_final_block = i == len(block_out_channels) - 1 + + down_block = get_down_block( + down_block_type, + num_layers=layers_per_block[i], + transformer_layers_per_block=transformer_layers_per_block[i], + in_channels=input_channel, + out_channels=output_channel, + temb_channels=blocks_time_embed_dim, + add_downsample=not is_final_block, + resnet_eps=norm_eps, + resnet_act_fn=act_fn, + resnet_groups=norm_num_groups, + cross_attention_dim=cross_attention_dim[i], + num_attention_heads=num_attention_heads[i], + downsample_padding=downsample_padding, + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention[i], + upcast_attention=upcast_attention, + resnet_time_scale_shift=resnet_time_scale_shift, + attention_type=attention_type, + resnet_skip_time_act=resnet_skip_time_act, + resnet_out_scale_factor=resnet_out_scale_factor, + cross_attention_norm=cross_attention_norm, + attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel, + dropout=dropout, + ) + self.down_blocks.append(down_block) + + # mid + if mid_block_type == "UNetMidBlock2DCrossAttn": + self.mid_block = UNetMidBlock2DCrossAttn( + transformer_layers_per_block=transformer_layers_per_block[-1], + in_channels=block_out_channels[-1], + temb_channels=blocks_time_embed_dim, + dropout=dropout, + resnet_eps=norm_eps, + resnet_act_fn=act_fn, + output_scale_factor=mid_block_scale_factor, + resnet_time_scale_shift=resnet_time_scale_shift, + cross_attention_dim=cross_attention_dim[-1], + num_attention_heads=num_attention_heads[-1], + resnet_groups=norm_num_groups, + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, + upcast_attention=upcast_attention, + attention_type=attention_type, + ) + elif mid_block_type == "UNetMidBlock2DSimpleCrossAttn": + self.mid_block = UNetMidBlock2DSimpleCrossAttn( + in_channels=block_out_channels[-1], + temb_channels=blocks_time_embed_dim, + dropout=dropout, + resnet_eps=norm_eps, + resnet_act_fn=act_fn, + output_scale_factor=mid_block_scale_factor, + cross_attention_dim=cross_attention_dim[-1], + attention_head_dim=attention_head_dim[-1], + resnet_groups=norm_num_groups, + resnet_time_scale_shift=resnet_time_scale_shift, + skip_time_act=resnet_skip_time_act, + only_cross_attention=mid_block_only_cross_attention, + cross_attention_norm=cross_attention_norm, + ) + elif mid_block_type == "UNetMidBlock2D": + self.mid_block = UNetMidBlock2D( + in_channels=block_out_channels[-1], + temb_channels=blocks_time_embed_dim, + dropout=dropout, + num_layers=0, + resnet_eps=norm_eps, + resnet_act_fn=act_fn, + output_scale_factor=mid_block_scale_factor, + resnet_groups=norm_num_groups, + resnet_time_scale_shift=resnet_time_scale_shift, + add_attention=False, + ) + elif mid_block_type is None: + self.mid_block = None + else: + raise ValueError(f"unknown mid_block_type : {mid_block_type}") + + # count how many layers upsample the images + self.num_upsamplers = 0 + + # up + reversed_block_out_channels = list(reversed(block_out_channels)) + reversed_num_attention_heads = list(reversed(num_attention_heads)) + reversed_layers_per_block = list(reversed(layers_per_block)) + reversed_cross_attention_dim = list(reversed(cross_attention_dim)) + reversed_transformer_layers_per_block = ( + list(reversed(transformer_layers_per_block)) + if reverse_transformer_layers_per_block is None + else reverse_transformer_layers_per_block + ) + only_cross_attention = list(reversed(only_cross_attention)) + + output_channel = reversed_block_out_channels[0] + for i, up_block_type in enumerate(up_block_types): + is_final_block = i == len(block_out_channels) - 1 + + prev_output_channel = output_channel + output_channel = reversed_block_out_channels[i] + input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)] + + # add upsample block for all BUT final layer + if not is_final_block: + add_upsample = True + self.num_upsamplers += 1 + else: + add_upsample = False + + up_block = get_up_block( + up_block_type, + num_layers=reversed_layers_per_block[i] + 1, + transformer_layers_per_block=reversed_transformer_layers_per_block[i], + in_channels=input_channel, + out_channels=output_channel, + prev_output_channel=prev_output_channel, + temb_channels=blocks_time_embed_dim, + add_upsample=add_upsample, + resnet_eps=norm_eps, + resnet_act_fn=act_fn, + resolution_idx=i, + resnet_groups=norm_num_groups, + cross_attention_dim=reversed_cross_attention_dim[i], + num_attention_heads=reversed_num_attention_heads[i], + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention[i], + upcast_attention=upcast_attention, + resnet_time_scale_shift=resnet_time_scale_shift, + attention_type=attention_type, + resnet_skip_time_act=resnet_skip_time_act, + resnet_out_scale_factor=resnet_out_scale_factor, + cross_attention_norm=cross_attention_norm, + attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel, + dropout=dropout, + ) + self.up_blocks.append(up_block) + prev_output_channel = output_channel + + # out + if norm_num_groups is not None: + self.conv_norm_out = nn.GroupNorm( + num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=norm_eps + ) + + self.conv_act = get_activation(act_fn) + + else: + self.conv_norm_out = None + self.conv_act = None + + conv_out_padding = (conv_out_kernel - 1) // 2 + self.conv_out = nn.Conv2d( + block_out_channels[0], out_channels, kernel_size=conv_out_kernel, padding=conv_out_padding + ) + + if attention_type in ["gated", "gated-text-image"]: + positive_len = 768 + if isinstance(cross_attention_dim, int): + positive_len = cross_attention_dim + elif isinstance(cross_attention_dim, tuple) or isinstance(cross_attention_dim, list): + positive_len = cross_attention_dim[0] + + feature_type = "text-only" if attention_type == "gated" else "text-image" + self.position_net = GLIGENTextBoundingboxProjection( + positive_len=positive_len, out_dim=cross_attention_dim, feature_type=feature_type + ) + + @property + def attn_processors(self) -> Dict[str, AttentionProcessor]: + r""" + Returns: + `dict` of attention processors: A dictionary containing all attention processors used in the model with + indexed by its weight name. + """ + # set recursively + processors = {} + + def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]): + if hasattr(module, "get_processor"): + processors[f"{name}.processor"] = module.get_processor(return_deprecated_lora=True) + + for sub_name, child in module.named_children(): + fn_recursive_add_processors(f"{name}.{sub_name}", child, processors) + + return processors + + for name, module in self.named_children(): + fn_recursive_add_processors(name, module, processors) + + return processors + + def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]): + r""" + Sets the attention processor to use to compute attention. + + Parameters: + processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`): + The instantiated processor class or a dictionary of processor classes that will be set as the processor + for **all** `Attention` layers. + + If `processor` is a dict, the key needs to define the path to the corresponding cross attention + processor. This is strongly recommended when setting trainable attention processors. + + """ + count = len(self.attn_processors.keys()) + + if isinstance(processor, dict) and len(processor) != count: + raise ValueError( + f"A dict of processors was passed, but the number of processors {len(processor)} does not match the" + f" number of attention layers: {count}. Please make sure to pass {count} processor classes." + ) + + def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor): + if hasattr(module, "set_processor"): + if not isinstance(processor, dict): + module.set_processor(processor) + else: + module.set_processor(processor.pop(f"{name}.processor")) + + for sub_name, child in module.named_children(): + fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor) + + for name, module in self.named_children(): + fn_recursive_attn_processor(name, module, processor) + + def set_default_attn_processor(self): + """ + Disables custom attention processors and sets the default attention implementation. + """ + if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()): + processor = AttnAddedKVProcessor() + elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()): + processor = AttnProcessor() + else: + raise ValueError( + f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}" + ) + + self.set_attn_processor(processor) + + def set_attention_slice(self, slice_size): + r""" + Enable sliced attention computation. + + When this option is enabled, the attention module splits the input tensor in slices to compute attention in + several steps. This is useful for saving some memory in exchange for a small decrease in speed. + + Args: + slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`): + When `"auto"`, input to the attention heads is halved, so attention is computed in two steps. If + `"max"`, maximum amount of memory is saved by running only one slice at a time. If a number is + provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim` + must be a multiple of `slice_size`. + """ + sliceable_head_dims = [] + + def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module): + if hasattr(module, "set_attention_slice"): + sliceable_head_dims.append(module.sliceable_head_dim) + + for child in module.children(): + fn_recursive_retrieve_sliceable_dims(child) + + # retrieve number of attention layers + for module in self.children(): + fn_recursive_retrieve_sliceable_dims(module) + + num_sliceable_layers = len(sliceable_head_dims) + + if slice_size == "auto": + # half the attention head size is usually a good trade-off between + # speed and memory + slice_size = [dim // 2 for dim in sliceable_head_dims] + elif slice_size == "max": + # make smallest slice possible + slice_size = num_sliceable_layers * [1] + + slice_size = num_sliceable_layers * [slice_size] if not isinstance(slice_size, list) else slice_size + + if len(slice_size) != len(sliceable_head_dims): + raise ValueError( + f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different" + f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}." + ) + + for i in range(len(slice_size)): + size = slice_size[i] + dim = sliceable_head_dims[i] + if size is not None and size > dim: + raise ValueError(f"size {size} has to be smaller or equal to {dim}.") + + # Recursively walk through all the children. + # Any children which exposes the set_attention_slice method + # gets the message + def fn_recursive_set_attention_slice(module: torch.nn.Module, slice_size: List[int]): + if hasattr(module, "set_attention_slice"): + module.set_attention_slice(slice_size.pop()) + + for child in module.children(): + fn_recursive_set_attention_slice(child, slice_size) + + reversed_slice_size = list(reversed(slice_size)) + for module in self.children(): + fn_recursive_set_attention_slice(module, reversed_slice_size) + + def _set_gradient_checkpointing(self, module, value=False): + if hasattr(module, "gradient_checkpointing"): + module.gradient_checkpointing = value + + def enable_freeu(self, s1, s2, b1, b2): + r"""Enables the FreeU mechanism from https://arxiv.org/abs/2309.11497. + + The suffixes after the scaling factors represent the stage blocks where they are being applied. + + Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of values that + are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL. + + Args: + s1 (`float`): + Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to + mitigate the "oversmoothing effect" in the enhanced denoising process. + s2 (`float`): + Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to + mitigate the "oversmoothing effect" in the enhanced denoising process. + b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features. + b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features. + """ + for i, upsample_block in enumerate(self.up_blocks): + setattr(upsample_block, "s1", s1) + setattr(upsample_block, "s2", s2) + setattr(upsample_block, "b1", b1) + setattr(upsample_block, "b2", b2) + + def disable_freeu(self): + """Disables the FreeU mechanism.""" + freeu_keys = {"s1", "s2", "b1", "b2"} + for i, upsample_block in enumerate(self.up_blocks): + for k in freeu_keys: + if hasattr(upsample_block, k) or getattr(upsample_block, k, None) is not None: + setattr(upsample_block, k, None) + + def fuse_qkv_projections(self): + """ + Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, + key, value) are fused. For cross-attention modules, key and value projection matrices are fused. + + + + This API is 🧪 experimental. + + + """ + self.original_attn_processors = None + + for _, attn_processor in self.attn_processors.items(): + if "Added" in str(attn_processor.__class__.__name__): + raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.") + + self.original_attn_processors = self.attn_processors + + for module in self.modules(): + if isinstance(module, Attention): + module.fuse_projections(fuse=True) + + def unfuse_qkv_projections(self): + """Disables the fused QKV projection if enabled. + + + + This API is 🧪 experimental. + + + + """ + if self.original_attn_processors is not None: + self.set_attn_processor(self.original_attn_processors) + + def unload_lora(self): + """Unloads LoRA weights.""" + deprecate( + "unload_lora", + "0.28.0", + "Calling `unload_lora()` is deprecated and will be removed in a future version. Please install `peft` and then call `disable_adapters().", + ) + for module in self.modules(): + if hasattr(module, "set_lora_layer"): + module.set_lora_layer(None) + + def forward( + self, + sample: torch.FloatTensor, + timestep: Union[torch.Tensor, float, int], + encoder_hidden_states: torch.Tensor, + class_labels: Optional[torch.Tensor] = None, + timestep_cond: Optional[torch.Tensor] = None, + attention_mask: Optional[torch.Tensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None, + down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None, + mid_block_additional_residual: Optional[torch.Tensor] = None, + down_intrablock_additional_residuals: Optional[Tuple[torch.Tensor]] = None, + encoder_attention_mask: Optional[torch.Tensor] = None, + return_dict: bool = True, + ) -> Union[UNet2DConditionOutput, Tuple]: + r""" + The [`UNet2DConditionModel`] forward method. + + Args: + sample (`torch.FloatTensor`): + The noisy input tensor with the following shape `(batch, channel, height, width)`. + timestep (`torch.FloatTensor` or `float` or `int`): The number of timesteps to denoise an input. + encoder_hidden_states (`torch.FloatTensor`): + The encoder hidden states with shape `(batch, sequence_length, feature_dim)`. + class_labels (`torch.Tensor`, *optional*, defaults to `None`): + Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings. + timestep_cond: (`torch.Tensor`, *optional*, defaults to `None`): + Conditional embeddings for timestep. If provided, the embeddings will be summed with the samples passed + through the `self.time_embedding` layer to obtain the timestep embeddings. + attention_mask (`torch.Tensor`, *optional*, defaults to `None`): + An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask + is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large + negative values to the attention scores corresponding to "discard" tokens. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + added_cond_kwargs: (`dict`, *optional*): + A kwargs dictionary containing additional embeddings that if specified are added to the embeddings that + are passed along to the UNet blocks. + down_block_additional_residuals: (`tuple` of `torch.Tensor`, *optional*): + A tuple of tensors that if specified are added to the residuals of down unet blocks. + mid_block_additional_residual: (`torch.Tensor`, *optional*): + A tensor that if specified is added to the residual of the middle unet block. + encoder_attention_mask (`torch.Tensor`): + A cross-attention mask of shape `(batch, sequence_length)` is applied to `encoder_hidden_states`. If + `True` the mask is kept, otherwise if `False` it is discarded. Mask will be converted into a bias, + which adds large negative values to the attention scores corresponding to "discard" tokens. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] instead of a plain + tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttnProcessor`]. + added_cond_kwargs: (`dict`, *optional*): + A kwargs dictionary containin additional embeddings that if specified are added to the embeddings that + are passed along to the UNet blocks. + down_block_additional_residuals (`tuple` of `torch.Tensor`, *optional*): + additional residuals to be added to UNet long skip connections from down blocks to up blocks for + example from ControlNet side model(s) + mid_block_additional_residual (`torch.Tensor`, *optional*): + additional residual to be added to UNet mid block output, for example from ControlNet side model + down_intrablock_additional_residuals (`tuple` of `torch.Tensor`, *optional*): + additional residuals to be added within UNet down blocks, for example from T2I-Adapter side model(s) + + Returns: + [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] or `tuple`: + If `return_dict` is True, an [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] is returned, otherwise + a `tuple` is returned where the first element is the sample tensor. + """ + # By default samples have to be AT least a multiple of the overall upsampling factor. + # The overall upsampling factor is equal to 2 ** (# num of upsampling layers). + # However, the upsampling interpolation output size can be forced to fit any upsampling size + # on the fly if necessary. + default_overall_up_factor = 2**self.num_upsamplers + + # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor` + forward_upsample_size = False + upsample_size = None + + for dim in sample.shape[-2:]: + if dim % default_overall_up_factor != 0: + # Forward upsample size to force interpolation output size. + forward_upsample_size = True + break + + # ensure attention_mask is a bias, and give it a singleton query_tokens dimension + # expects mask of shape: + # [batch, key_tokens] + # adds singleton query_tokens dimension: + # [batch, 1, key_tokens] + # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes: + # [batch, heads, query_tokens, key_tokens] (e.g. torch sdp attn) + # [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn) + if attention_mask is not None: + # assume that mask is expressed as: + # (1 = keep, 0 = discard) + # convert mask into a bias that can be added to attention scores: + # (keep = +0, discard = -10000.0) + attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0 + attention_mask = attention_mask.unsqueeze(1) + + # convert encoder_attention_mask to a bias the same way we do for attention_mask + if encoder_attention_mask is not None: + encoder_attention_mask = (1 - encoder_attention_mask.to(sample.dtype)) * -10000.0 + encoder_attention_mask = encoder_attention_mask.unsqueeze(1) + + # 0. center input if necessary + if self.config.center_input_sample: + sample = 2 * sample - 1.0 + + # 1. time + timesteps = timestep + if not torch.is_tensor(timesteps): + # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can + # This would be a good case for the `match` statement (Python 3.10+) + is_mps = sample.device.type == "mps" + if isinstance(timestep, float): + dtype = torch.float32 if is_mps else torch.float64 + else: + dtype = torch.int32 if is_mps else torch.int64 + timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device) + elif len(timesteps.shape) == 0: + timesteps = timesteps[None].to(sample.device) + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timesteps = timesteps.expand(sample.shape[0]) + + t_emb = self.time_proj(timesteps) + + # `Timesteps` does not contain any weights and will always return f32 tensors + # but time_embedding might actually be running in fp16. so we need to cast here. + # there might be better ways to encapsulate this. + t_emb = t_emb.to(dtype=sample.dtype) + + emb = self.time_embedding(t_emb, timestep_cond) + aug_emb = None + + if self.class_embedding is not None: + if class_labels is None: + raise ValueError("class_labels should be provided when num_class_embeds > 0") + + if self.config.class_embed_type == "timestep": + class_labels = self.time_proj(class_labels) + + # `Timesteps` does not contain any weights and will always return f32 tensors + # there might be better ways to encapsulate this. + class_labels = class_labels.to(dtype=sample.dtype) + + class_emb = self.class_embedding(class_labels).to(dtype=sample.dtype) + + if self.config.class_embeddings_concat: + emb = torch.cat([emb, class_emb], dim=-1) + else: + emb = emb + class_emb + + if self.config.addition_embed_type == "text": + aug_emb = self.add_embedding(encoder_hidden_states) + elif self.config.addition_embed_type == "text_image": + # Kandinsky 2.1 - style + if "image_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `addition_embed_type` set to 'text_image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`" + ) + + image_embs = added_cond_kwargs.get("image_embeds") + text_embs = added_cond_kwargs.get("text_embeds", encoder_hidden_states) + aug_emb = self.add_embedding(text_embs, image_embs) + elif self.config.addition_embed_type == "text_time": + # SDXL - style + if "text_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`" + ) + text_embeds = added_cond_kwargs.get("text_embeds") + if "time_ids" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`" + ) + time_ids = added_cond_kwargs.get("time_ids") + time_embeds = self.add_time_proj(time_ids.flatten()) + time_embeds = time_embeds.reshape((text_embeds.shape[0], -1)) + add_embeds = torch.concat([text_embeds, time_embeds], dim=-1) + add_embeds = add_embeds.to(emb.dtype) + aug_emb = self.add_embedding(add_embeds) + elif self.config.addition_embed_type == "image": + # Kandinsky 2.2 - style + if "image_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `addition_embed_type` set to 'image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`" + ) + image_embs = added_cond_kwargs.get("image_embeds") + aug_emb = self.add_embedding(image_embs) + elif self.config.addition_embed_type == "image_hint": + # Kandinsky 2.2 - style + if "image_embeds" not in added_cond_kwargs or "hint" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `addition_embed_type` set to 'image_hint' which requires the keyword arguments `image_embeds` and `hint` to be passed in `added_cond_kwargs`" + ) + image_embs = added_cond_kwargs.get("image_embeds") + hint = added_cond_kwargs.get("hint") + aug_emb, hint = self.add_embedding(image_embs, hint) + sample = torch.cat([sample, hint], dim=1) + + emb = emb + aug_emb if aug_emb is not None else emb + + if self.time_embed_act is not None: + emb = self.time_embed_act(emb) + + if self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_proj": + encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states) + elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_image_proj": + # Kadinsky 2.1 - style + if "image_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'text_image_proj' which requires the keyword argument `image_embeds` to be passed in `added_conditions`" + ) + + image_embeds = added_cond_kwargs.get("image_embeds") + encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states, image_embeds) + elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "image_proj": + # Kandinsky 2.2 - style + if "image_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'image_proj' which requires the keyword argument `image_embeds` to be passed in `added_conditions`" + ) + image_embeds = added_cond_kwargs.get("image_embeds") + encoder_hidden_states = self.encoder_hid_proj(image_embeds) + elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "ip_image_proj": + if "image_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'ip_image_proj' which requires the keyword argument `image_embeds` to be passed in `added_conditions`" + ) + image_embeds = added_cond_kwargs.get("image_embeds") + image_embeds = self.encoder_hid_proj(image_embeds) + encoder_hidden_states = (encoder_hidden_states, image_embeds) + + # 2. pre-process + sample = self.conv_in(sample) + + # 2.5 GLIGEN position net + if cross_attention_kwargs is not None and cross_attention_kwargs.get("gligen", None) is not None: + cross_attention_kwargs = cross_attention_kwargs.copy() + gligen_args = cross_attention_kwargs.pop("gligen") + cross_attention_kwargs["gligen"] = {"objs": self.position_net(**gligen_args)} + + # 3. down + lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0 + if USE_PEFT_BACKEND: + # weight the lora layers by setting `lora_scale` for each PEFT layer + scale_lora_layers(self, lora_scale) + + is_controlnet = mid_block_additional_residual is not None and down_block_additional_residuals is not None + # using new arg down_intrablock_additional_residuals for T2I-Adapters, to distinguish from controlnets + is_adapter = down_intrablock_additional_residuals is not None + # maintain backward compatibility for legacy usage, where + # T2I-Adapter and ControlNet both use down_block_additional_residuals arg + # but can only use one or the other + if not is_adapter and mid_block_additional_residual is None and down_block_additional_residuals is not None: + deprecate( + "T2I should not use down_block_additional_residuals", + "1.3.0", + "Passing intrablock residual connections with `down_block_additional_residuals` is deprecated \ + and will be removed in diffusers 1.3.0. `down_block_additional_residuals` should only be used \ + for ControlNet. Please make sure use `down_intrablock_additional_residuals` instead. ", + standard_warn=False, + ) + down_intrablock_additional_residuals = down_block_additional_residuals + is_adapter = True + + down_block_res_samples = (sample,) + for downsample_block in self.down_blocks: + if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention: + # For t2i-adapter CrossAttnDownBlock2D + additional_residuals = {} + if is_adapter and len(down_intrablock_additional_residuals) > 0: + additional_residuals["additional_residuals"] = down_intrablock_additional_residuals.pop(0) + + sample, res_samples = downsample_block( + hidden_states=sample, + temb=emb, + encoder_hidden_states=encoder_hidden_states, + attention_mask=attention_mask, + cross_attention_kwargs=cross_attention_kwargs, + encoder_attention_mask=encoder_attention_mask, + **additional_residuals, + ) + else: + sample, res_samples = downsample_block(hidden_states=sample, temb=emb, scale=lora_scale) + if is_adapter and len(down_intrablock_additional_residuals) > 0: + sample += down_intrablock_additional_residuals.pop(0) + + down_block_res_samples += res_samples + + if is_controlnet: + new_down_block_res_samples = () + + for down_block_res_sample, down_block_additional_residual in zip( + down_block_res_samples, down_block_additional_residuals + ): + down_block_res_sample = down_block_res_sample + down_block_additional_residual + new_down_block_res_samples = new_down_block_res_samples + (down_block_res_sample,) + + down_block_res_samples = new_down_block_res_samples + + # 4. mid + if self.mid_block is not None: + if hasattr(self.mid_block, "has_cross_attention") and self.mid_block.has_cross_attention: + sample = self.mid_block( + sample, + emb, + encoder_hidden_states=encoder_hidden_states, + attention_mask=attention_mask, + cross_attention_kwargs=cross_attention_kwargs, + encoder_attention_mask=encoder_attention_mask, + ) + else: + sample = self.mid_block(sample, emb) + + # To support T2I-Adapter-XL + if ( + is_adapter + and len(down_intrablock_additional_residuals) > 0 + and sample.shape == down_intrablock_additional_residuals[0].shape + ): + sample += down_intrablock_additional_residuals.pop(0) + + if is_controlnet: + sample = sample + mid_block_additional_residual + + # 5. up + for i, upsample_block in enumerate(self.up_blocks): + is_final_block = i == len(self.up_blocks) - 1 + + res_samples = down_block_res_samples[-len(upsample_block.resnets) :] + down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)] + + # if we have not reached the final block and need to forward the + # upsample size, we do it here + if not is_final_block and forward_upsample_size: + upsample_size = down_block_res_samples[-1].shape[2:] + + if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention: + sample = upsample_block( + hidden_states=sample, + temb=emb, + res_hidden_states_tuple=res_samples, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + upsample_size=upsample_size, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + ) + else: + sample = upsample_block( + hidden_states=sample, + temb=emb, + res_hidden_states_tuple=res_samples, + upsample_size=upsample_size, + scale=lora_scale, + ) + + # 6. post-process + if self.conv_norm_out: + sample = self.conv_norm_out(sample) + sample = self.conv_act(sample) + sample = self.conv_out(sample) + + if USE_PEFT_BACKEND: + # remove `lora_scale` from each PEFT layer + unscale_lora_layers(self, lora_scale) + + if not return_dict: + return (sample,) + + return UNet2DConditionOutput(sample=sample) diff --git a/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/pipeline/__init__.py b/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/pipeline/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..efd186f76b5e7023389c7d4c5f6358e44492d874 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/pipeline/__init__.py @@ -0,0 +1 @@ +from .pipeline_stable_diffusion import StableDiffusionPipeline, DiffusionPipeline \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/pipeline/pipeline_stable_diffusion.py b/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/pipeline/pipeline_stable_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..5d00932bcf2274801db295625c428aade682b856 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/pipeline/pipeline_stable_diffusion.py @@ -0,0 +1,1088 @@ +# Copyright 2023 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable, Dict, List, Optional, Union + +import torch +from packaging import version +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection + +from diffusers.configuration_utils import FrozenDict +from diffusers.image_processor import PipelineImageInput, VaeImageProcessor +from diffusers.loaders import FromSingleFileMixin, IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin +from diffusers.models import AutoencoderKL, ImageProjection, UNet2DConditionModel +from diffusers.models.attention_processor import FusedAttnProcessor2_0 +from diffusers.models.lora import adjust_lora_scale_text_encoder +from diffusers.schedulers import KarrasDiffusionSchedulers +from diffusers.utils import ( + USE_PEFT_BACKEND, + deprecate, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from diffusers.utils.torch_utils import randn_tensor +from diffusers.pipelines.pipeline_utils import DiffusionPipeline +from diffusers.pipelines.stable_diffusion.pipeline_output import StableDiffusionPipelineOutput +from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import StableDiffusionPipeline + + >>> pipe = StableDiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16) + >>> pipe = pipe.to("cuda") + + >>> prompt = "a photo of an astronaut riding a horse on mars" + >>> image = pipe(prompt).images[0] + ``` +""" + + +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + """ + Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4 + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +def retrieve_timesteps( + scheduler, + num_inference_steps: Optional[int] = None, + device: Optional[Union[str, torch.device]] = None, + timesteps: Optional[List[int]] = None, + **kwargs, +): + """ + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, + `timesteps` must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`List[int]`, *optional*): + Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default + timestep spacing strategy of the scheduler is used. If `timesteps` is passed, `num_inference_steps` + must be `None`. + + Returns: + `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class StableDiffusionPipeline( + DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin, IPAdapterMixin, FromSingleFileMixin +): + r""" + Pipeline for text-to-image generation using Stable Diffusion. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details + about a model's potential harms. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + """ + + model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae" + _optional_components = ["safety_checker", "feature_extractor", "image_encoder"] + _exclude_from_cpu_offload = ["safety_checker"] + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + image_encoder: CLIPVisionModelWithProjection = None, + requires_safety_checker: bool = True, + ): + super().__init__() + + if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`" + f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure " + "to update the config accordingly as leaving `steps_offset` might led to incorrect results" + " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub," + " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`" + " file" + ) + deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["steps_offset"] = 1 + scheduler._internal_dict = FrozenDict(new_config) + + if hasattr(scheduler.config, "clip_sample") and scheduler.config.clip_sample is True: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`." + " `clip_sample` should be set to False in the configuration file. Please make sure to update the" + " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in" + " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very" + " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file" + ) + deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["clip_sample"] = False + scheduler._internal_dict = FrozenDict(new_config) + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + is_unet_version_less_0_9_0 = hasattr(unet.config, "_diffusers_version") and version.parse( + version.parse(unet.config._diffusers_version).base_version + ) < version.parse("0.9.0.dev0") + is_unet_sample_size_less_64 = hasattr(unet.config, "sample_size") and unet.config.sample_size < 64 + if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64: + deprecation_message = ( + "The configuration file of the unet has set the default `sample_size` to smaller than" + " 64 which seems highly unlikely. If your checkpoint is a fine-tuned version of any of the" + " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-" + " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- runwayml/stable-diffusion-v1-5" + " \n- runwayml/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the" + " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`" + " in the config might lead to incorrect results in future versions. If you have downloaded this" + " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for" + " the `unet/config.json` file" + ) + deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(unet.config) + new_config["sample_size"] = 64 + unet._internal_dict = FrozenDict(new_config) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + image_encoder=image_encoder, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + def enable_vae_slicing(self): + r""" + Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to + compute decoding in several steps. This is useful to save some memory and allow larger batch sizes. + """ + self.vae.enable_slicing() + + def disable_vae_slicing(self): + r""" + Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to + computing decoding in one step. + """ + self.vae.disable_slicing() + + def enable_vae_tiling(self): + r""" + Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to + compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow + processing larger images. + """ + self.vae.enable_tiling() + + def disable_vae_tiling(self): + r""" + Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to + computing decoding in one step. + """ + self.vae.disable_tiling() + + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.FloatTensor] = None, + negative_prompt_embeds: Optional[torch.FloatTensor] = None, + lora_scale: Optional[float] = None, + **kwargs, + ): + deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple." + deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False) + + prompt_embeds_tuple = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + **kwargs, + ) + + # concatenate for backwards comp + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + + return prompt_embeds + + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.FloatTensor] = None, + negative_prompt_embeds: Optional[torch.FloatTensor] = None, + lora_scale: Optional[float] = None, + clip_skip: Optional[int] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, LoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: procecss multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: List[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: procecss multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + def prepare_ip_adapter_image_embeds(self, ip_adapter_image, device, num_images_per_prompt): + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + image_embeds = [] + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers + ): + output_hidden_state = not isinstance(image_proj_layer, ImageProjection) + single_image_embeds, single_negative_image_embeds = self.encode_image( + single_ip_adapter_image, device, 1, output_hidden_state + ) + single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0) + single_negative_image_embeds = torch.stack([single_negative_image_embeds] * num_images_per_prompt, dim=0) + + if self.do_classifier_free_guidance: + single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds]) + single_image_embeds = single_image_embeds.to(device) + + image_embeds.append(single_image_embeds) + + return image_embeds + + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + def decode_latents(self, latents): + deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead" + deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False) + + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + height, + width, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + def enable_freeu(self, s1: float, s2: float, b1: float, b2: float): + r"""Enables the FreeU mechanism as in https://arxiv.org/abs/2309.11497. + + The suffixes after the scaling factors represent the stages where they are being applied. + + Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of the values + that are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL. + + Args: + s1 (`float`): + Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to + mitigate "oversmoothing effect" in the enhanced denoising process. + s2 (`float`): + Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to + mitigate "oversmoothing effect" in the enhanced denoising process. + b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features. + b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features. + """ + if not hasattr(self, "unet"): + raise ValueError("The pipeline must have `unet` for using FreeU.") + self.unet.enable_freeu(s1=s1, s2=s2, b1=b1, b2=b2) + + def disable_freeu(self): + """Disables the FreeU mechanism if enabled.""" + self.unet.disable_freeu() + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.fuse_qkv_projections + def fuse_qkv_projections(self, unet: bool = True, vae: bool = True): + """ + Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, + key, value) are fused. For cross-attention modules, key and value projection matrices are fused. + + + + This API is 🧪 experimental. + + + + Args: + unet (`bool`, defaults to `True`): To apply fusion on the UNet. + vae (`bool`, defaults to `True`): To apply fusion on the VAE. + """ + self.fusing_unet = False + self.fusing_vae = False + + if unet: + self.fusing_unet = True + self.unet.fuse_qkv_projections() + self.unet.set_attn_processor(FusedAttnProcessor2_0()) + + if vae: + if not isinstance(self.vae, AutoencoderKL): + raise ValueError("`fuse_qkv_projections()` is only supported for the VAE of type `AutoencoderKL`.") + + self.fusing_vae = True + self.vae.fuse_qkv_projections() + self.vae.set_attn_processor(FusedAttnProcessor2_0()) + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.unfuse_qkv_projections + def unfuse_qkv_projections(self, unet: bool = True, vae: bool = True): + """Disable QKV projection fusion if enabled. + + + + This API is 🧪 experimental. + + + + Args: + unet (`bool`, defaults to `True`): To apply fusion on the UNet. + vae (`bool`, defaults to `True`): To apply fusion on the VAE. + + """ + if unet: + if not self.fusing_unet: + logger.warning("The UNet was not initially fused for QKV projections. Doing nothing.") + else: + self.unet.unfuse_qkv_projections() + self.fusing_unet = False + + if vae: + if not self.fusing_vae: + logger.warning("The VAE was not initially fused for QKV projections. Doing nothing.") + else: + self.vae.unfuse_qkv_projections() + self.fusing_vae = False + + # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding + def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32): + """ + See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 + + Args: + timesteps (`torch.Tensor`): + generate embedding vectors at these timesteps + embedding_dim (`int`, *optional*, defaults to 512): + dimension of the embeddings to generate + dtype: + data type of the generated embeddings + + Returns: + `torch.FloatTensor`: Embedding vectors with shape `(len(timesteps), embedding_dim)` + """ + assert len(w.shape) == 1 + w = w * 1000.0 + + half_dim = embedding_dim // 2 + emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) + emb = w.to(dtype)[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0, 1)) + assert emb.shape == (w.shape[0], embedding_dim) + return emb + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def guidance_rescale(self): + return self._guidance_rescale + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None + + @property + def cross_attention_kwargs(self): + return self._cross_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Union[str, List[str]] = None, + height: Optional[int] = None, + width: Optional[int] = None, + num_inference_steps: int = 50, + timesteps: List[int] = None, + guidance_scale: float = 7.5, + negative_prompt: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.FloatTensor] = None, + prompt_embeds: Optional[torch.FloatTensor] = None, + negative_prompt_embeds: Optional[torch.FloatTensor] = None, + ip_adapter_image: Optional[PipelineImageInput] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + guidance_rescale: float = 0.0, + clip_skip: Optional[int] = None, + callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None, + callback_on_step_end_tensor_inputs: List[str] = ["latents"], + **kwargs, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`List[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies + to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + guidance_rescale (`float`, *optional*, defaults to 0.0): + Guidance rescale factor from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://arxiv.org/pdf/2305.08891.pdf). Guidance rescale factor should fix overexposure when + using zero terminal SNR. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + """ + + callback = kwargs.pop("callback", None) + callback_steps = kwargs.pop("callback_steps", None) + + if callback is not None: + deprecate( + "callback", + "1.0.0", + "Passing `callback` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`", + ) + if callback_steps is not None: + deprecate( + "callback_steps", + "1.0.0", + "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`", + ) + + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + # to deal with lora scaling and other possible forward hooks + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + height, + width, + callback_steps, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + callback_on_step_end_tensor_inputs, + ) + + self._guidance_scale = guidance_scale + self._guidance_rescale = guidance_rescale + self._clip_skip = clip_skip + self._cross_attention_kwargs = cross_attention_kwargs + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Encode input prompt + lora_scale = ( + self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None + ) + + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + self.do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + clip_skip=self.clip_skip, + ) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + if ip_adapter_image is not None: + image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, device, batch_size * num_images_per_prompt + ) + + # 4. Prepare timesteps + timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps) + + # 5. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 6.1 Add image embeds for IP-Adapter + added_cond_kwargs = {"image_embeds": image_embeds} if ip_adapter_image is not None else None + + # 6.2 Optionally get Guidance Scale Embedding + timestep_cond = None + if self.unet.config.time_cond_proj_dim is not None: + guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt) + timestep_cond = self.get_guidance_scale_embedding( + guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim + ).to(device=device, dtype=latents.dtype) + + # 7. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + timestep_cond=timestep_cond, + cross_attention_kwargs=self.cross_attention_kwargs, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + + if self.do_classifier_free_guidance and self.guidance_rescale > 0.0: + # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf + noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, generator=generator)[ + 0 + ] + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + else: + image = latents + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/schedulers/__init__.py b/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/schedulers/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..6f853e7d075eb2528f2ef9e4e230a082af79fb80 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/schedulers/__init__.py @@ -0,0 +1 @@ +from .scheduling_pndm import PNDMScheduler, SchedulerMixin \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/schedulers/scheduling_pndm.py b/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/schedulers/scheduling_pndm.py new file mode 100644 index 0000000000000000000000000000000000000000..7d50d01e9602ee2bfa4b612baadcfdc8128deff1 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/schedulers/scheduling_pndm.py @@ -0,0 +1,478 @@ +# Copyright 2023 Zhejiang University Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# DISCLAIMER: This file is strongly influenced by https://github.com/ermongroup/ddim + +import math +from typing import List, Optional, Tuple, Union + +import numpy as np +import torch + +from diffusers.configuration_utils import ConfigMixin, register_to_config +from diffusers.schedulers.scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin, SchedulerOutput + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps, + max_beta=0.999, + alpha_transform_type="cosine", +): + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + + Args: + num_diffusion_timesteps (`int`): the number of betas to produce. + max_beta (`float`): the maximum beta to use; use values lower than 1 to + prevent singularities. + alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar. + Choose from `cosine` or `exp` + + Returns: + betas (`np.ndarray`): the betas used by the scheduler to step the model outputs + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_tranform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +class PNDMScheduler(SchedulerMixin, ConfigMixin): + """ + `PNDMScheduler` uses pseudo numerical methods for diffusion models such as the Runge-Kutta and linear multi-step + method. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to 0.0001): + The starting `beta` value of inference. + beta_end (`float`, defaults to 0.02): + The final `beta` value. + beta_schedule (`str`, defaults to `"linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear`, `scaled_linear`, or `squaredcos_cap_v2`. + trained_betas (`np.ndarray`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + skip_prk_steps (`bool`, defaults to `False`): + Allows the scheduler to skip the Runge-Kutta steps defined in the original paper as being required before + PLMS steps. + set_alpha_to_one (`bool`, defaults to `False`): + Each diffusion step uses the alphas product value at that step and at the previous one. For the final step + there is no previous alpha. When this option is `True` the previous alpha product is fixed to `1`, + otherwise it uses the alpha value at step 0. + prediction_type (`str`, defaults to `epsilon`, *optional*): + Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process) + or `v_prediction` (see section 2.4 of [Imagen Video](https://imagen.research.google/video/paper.pdf) + paper). + timestep_spacing (`str`, defaults to `"leading"`): + The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + steps_offset (`int`, defaults to 0): + An offset added to the inference steps. You can use a combination of `offset=1` and + `set_alpha_to_one=False` to make the last step use step 0 for the previous alpha product like in Stable + Diffusion. + """ + + _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.0001, + beta_end: float = 0.02, + beta_schedule: str = "linear", + trained_betas: Optional[Union[np.ndarray, List[float]]] = None, + skip_prk_steps: bool = False, + set_alpha_to_one: bool = False, + prediction_type: str = "epsilon", + timestep_spacing: str = "leading", + steps_offset: int = 0, + ): + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2 + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + else: + raise NotImplementedError(f"{beta_schedule} does is not implemented for {self.__class__}") + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + + self.final_alpha_cumprod = torch.tensor(1.0) if set_alpha_to_one else self.alphas_cumprod[0] + + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + # For now we only support F-PNDM, i.e. the runge-kutta method + # For more information on the algorithm please take a look at the paper: https://arxiv.org/pdf/2202.09778.pdf + # mainly at formula (9), (12), (13) and the Algorithm 2. + self.pndm_order = 4 + + # running values + self.cur_model_output = 0 + self.counter = 0 + self.cur_sample = None + self.ets = [] + + # setable values + self.num_inference_steps = None + self._timesteps = np.arange(0, num_train_timesteps)[::-1].copy() + self.prk_timesteps = None + self.plms_timesteps = None + self.timesteps = None + + def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + """ + + self.num_inference_steps = num_inference_steps + # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://arxiv.org/abs/2305.08891 + if self.config.timestep_spacing == "linspace": + self._timesteps = ( + np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps).round().astype(np.int64) + ) + elif self.config.timestep_spacing == "leading": + step_ratio = self.config.num_train_timesteps // self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + self._timesteps = (np.arange(0, num_inference_steps) * step_ratio).round() + self._timesteps += self.config.steps_offset + elif self.config.timestep_spacing == "trailing": + step_ratio = self.config.num_train_timesteps / self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + self._timesteps = np.round(np.arange(self.config.num_train_timesteps, 0, -step_ratio))[::-1].astype( + np.int64 + ) + self._timesteps -= 1 + else: + raise ValueError( + f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'." + ) + + if self.config.skip_prk_steps: + # for some models like stable diffusion the prk steps can/should be skipped to + # produce better results. When using PNDM with `self.config.skip_prk_steps` the implementation + # is based on crowsonkb's PLMS sampler implementation: https://github.com/CompVis/latent-diffusion/pull/51 + self.prk_timesteps = np.array([]) + self.plms_timesteps = np.concatenate([self._timesteps[:-1], self._timesteps[-2:-1], self._timesteps[-1:]])[ + ::-1 + ].copy() + else: + prk_timesteps = np.array(self._timesteps[-self.pndm_order :]).repeat(2) + np.tile( + np.array([0, self.config.num_train_timesteps // num_inference_steps // 2]), self.pndm_order + ) + self.prk_timesteps = (prk_timesteps[:-1].repeat(2)[1:-1])[::-1].copy() + self.plms_timesteps = self._timesteps[:-3][ + ::-1 + ].copy() # we copy to avoid having negative strides which are not supported by torch.from_numpy + + timesteps = np.concatenate([self.prk_timesteps, self.plms_timesteps]).astype(np.int64) + self.timesteps = torch.from_numpy(timesteps).to(device) + + self.ets = [] + self.counter = 0 + self.cur_model_output = 0 + + def step( + self, + model_output: torch.FloatTensor, + timestep: int, + sample: torch.FloatTensor, + return_dict: bool = True, + ) -> Union[SchedulerOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise), and calls [`~PNDMScheduler.step_prk`] + or [`~PNDMScheduler.step_plms`] depending on the internal variable `counter`. + + Args: + model_output (`torch.FloatTensor`): + The direct output from learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + sample (`torch.FloatTensor`): + A current instance of a sample created by the diffusion process. + return_dict (`bool`): + Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`. + + Returns: + [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + + """ + if self.counter < len(self.prk_timesteps) and not self.config.skip_prk_steps: + return self.step_prk(model_output=model_output, timestep=timestep, sample=sample, return_dict=return_dict) + else: + return self.step_plms(model_output=model_output, timestep=timestep, sample=sample, return_dict=return_dict) + + def step_prk( + self, + model_output: torch.FloatTensor, + timestep: int, + sample: torch.FloatTensor, + return_dict: bool = True, + ) -> Union[SchedulerOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with + the Runge-Kutta method. It performs four forward passes to approximate the solution to the differential + equation. + + Args: + model_output (`torch.FloatTensor`): + The direct output from learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + sample (`torch.FloatTensor`): + A current instance of a sample created by the diffusion process. + return_dict (`bool`): + Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or tuple. + + Returns: + [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + + """ + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + diff_to_prev = 0 if self.counter % 2 else self.config.num_train_timesteps // self.num_inference_steps // 2 + prev_timestep = timestep - diff_to_prev + timestep = self.prk_timesteps[self.counter // 4 * 4] + + if self.counter % 4 == 0: + self.cur_model_output += 1 / 6 * model_output + self.ets.append(model_output) + self.cur_sample = sample + elif (self.counter - 1) % 4 == 0: + self.cur_model_output += 1 / 3 * model_output + elif (self.counter - 2) % 4 == 0: + self.cur_model_output += 1 / 3 * model_output + elif (self.counter - 3) % 4 == 0: + model_output = self.cur_model_output + 1 / 6 * model_output + self.cur_model_output = 0 + + # cur_sample should not be `None` + cur_sample = self.cur_sample if self.cur_sample is not None else sample + + prev_sample = self._get_prev_sample(cur_sample, timestep, prev_timestep, model_output) + self.counter += 1 + + if not return_dict: + return (prev_sample,) + + return SchedulerOutput(prev_sample=prev_sample) + + def step_plms( + self, + model_output: torch.FloatTensor, + timestep: int, + sample: torch.FloatTensor, + return_dict: bool = True, + ) -> Union[SchedulerOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with + the linear multistep method. It performs one forward pass multiple times to approximate the solution. + + Args: + model_output (`torch.FloatTensor`): + The direct output from learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + sample (`torch.FloatTensor`): + A current instance of a sample created by the diffusion process. + return_dict (`bool`): + Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or tuple. + + Returns: + [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + + """ + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + if not self.config.skip_prk_steps and len(self.ets) < 3: + raise ValueError( + f"{self.__class__} can only be run AFTER scheduler has been run " + "in 'prk' mode for at least 12 iterations " + "See: https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/pipeline_pndm.py " + "for more information." + ) + + prev_timestep = timestep - self.config.num_train_timesteps // self.num_inference_steps + + if self.counter != 1: + self.ets = self.ets[-3:] + self.ets.append(model_output) + else: + prev_timestep = timestep + timestep = timestep + self.config.num_train_timesteps // self.num_inference_steps + + if len(self.ets) == 1 and self.counter == 0: + model_output = model_output + self.cur_sample = sample + elif len(self.ets) == 1 and self.counter == 1: + model_output = (model_output + self.ets[-1]) / 2 + sample = self.cur_sample + self.cur_sample = None + elif len(self.ets) == 2: + model_output = (3 * self.ets[-1] - self.ets[-2]) / 2 + elif len(self.ets) == 3: + model_output = (23 * self.ets[-1] - 16 * self.ets[-2] + 5 * self.ets[-3]) / 12 + else: + model_output = (1 / 24) * (55 * self.ets[-1] - 59 * self.ets[-2] + 37 * self.ets[-3] - 9 * self.ets[-4]) + + prev_sample = self._get_prev_sample(sample, timestep, prev_timestep, model_output) + self.counter += 1 + + if not return_dict: + return (prev_sample,) + + return SchedulerOutput(prev_sample=prev_sample) + + def scale_model_input(self, sample: torch.FloatTensor, *args, **kwargs) -> torch.FloatTensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.FloatTensor`): + The input sample. + + Returns: + `torch.FloatTensor`: + A scaled input sample. + """ + return sample + + def _get_prev_sample(self, sample, timestep, prev_timestep, model_output): + # See formula (9) of PNDM paper https://arxiv.org/pdf/2202.09778.pdf + # this function computes x_(t−δ) using the formula of (9) + # Note that x_t needs to be added to both sides of the equation + + # Notation ( -> + # alpha_prod_t -> α_t + # alpha_prod_t_prev -> α_(t−δ) + # beta_prod_t -> (1 - α_t) + # beta_prod_t_prev -> (1 - α_(t−δ)) + # sample -> x_t + # model_output -> e_θ(x_t, t) + # prev_sample -> x_(t−δ) + alpha_prod_t = self.alphas_cumprod[timestep] + alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod + beta_prod_t = 1 - alpha_prod_t + beta_prod_t_prev = 1 - alpha_prod_t_prev + + if self.config.prediction_type == "v_prediction": + model_output = (alpha_prod_t**0.5) * model_output + (beta_prod_t**0.5) * sample + elif self.config.prediction_type != "epsilon": + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon` or `v_prediction`" + ) + + # corresponds to (α_(t−δ) - α_t) divided by + # denominator of x_t in formula (9) and plus 1 + # Note: (α_(t−δ) - α_t) / (sqrt(α_t) * (sqrt(α_(t−δ)) + sqr(α_t))) = + # sqrt(α_(t−δ)) / sqrt(α_t)) + sample_coeff = (alpha_prod_t_prev / alpha_prod_t) ** (0.5) + + # corresponds to denominator of e_θ(x_t, t) in formula (9) + model_output_denom_coeff = alpha_prod_t * beta_prod_t_prev ** (0.5) + ( + alpha_prod_t * beta_prod_t * alpha_prod_t_prev + ) ** (0.5) + + # full formula (9) + prev_sample = ( + sample_coeff * sample - (alpha_prod_t_prev - alpha_prod_t) * model_output / model_output_denom_coeff + ) + + return prev_sample + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.add_noise + def add_noise( + self, + original_samples: torch.FloatTensor, + noise: torch.FloatTensor, + timesteps: torch.IntTensor, + ) -> torch.FloatTensor: + # Make sure alphas_cumprod and timestep have same device and dtype as original_samples + # Move the self.alphas_cumprod to device to avoid redundant CPU to GPU data movement + # for the subsequent add_noise calls + self.alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device) + alphas_cumprod = self.alphas_cumprod.to(dtype=original_samples.dtype) + timesteps = timesteps.to(original_samples.device) + + sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(original_samples.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise + return noisy_samples + + def __len__(self): + return self.config.num_train_timesteps diff --git a/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/vae/__init__.py b/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/vae/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..58bbb8f14ee8e394705159b0ebf877de963c143a --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/vae/__init__.py @@ -0,0 +1 @@ +from .autoencoder_kl import AutoencoderKL \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/vae/autoencoder_kl.py b/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/vae/autoencoder_kl.py new file mode 100644 index 0000000000000000000000000000000000000000..0a9d02770b66dd1874224d14fad5f6d46f7f0370 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-1.5/stablediffusion/vae/autoencoder_kl.py @@ -0,0 +1,488 @@ +# Copyright 2023 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from typing import Dict, Optional, Tuple, Union + +import torch +import torch.nn as nn + +from diffusers.configuration_utils import ConfigMixin, register_to_config +from diffusers.loaders import FromOriginalVAEMixin +from diffusers.utils.accelerate_utils import apply_forward_hook +from ..layers.attention_processor import ( + ADDED_KV_ATTENTION_PROCESSORS, + CROSS_ATTENTION_PROCESSORS, + Attention, + AttentionProcessor, + AttnAddedKVProcessor, + AttnProcessor, +) +from diffusers.models.modeling_outputs import AutoencoderKLOutput +from diffusers.models.autoencoders.vae import Decoder, DecoderOutput, DiagonalGaussianDistribution, Encoder + +from stablediffusion.models.model_utils import ModelMixin + + +class AutoencoderKL(ModelMixin, ConfigMixin, FromOriginalVAEMixin): + r""" + A VAE model with KL loss for encoding images into latents and decoding latent representations into images. + + This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented + for all models (such as downloading or saving). + + Parameters: + in_channels (int, *optional*, defaults to 3): Number of channels in the input image. + out_channels (int, *optional*, defaults to 3): Number of channels in the output. + down_block_types (`Tuple[str]`, *optional*, defaults to `("DownEncoderBlock2D",)`): + Tuple of downsample block types. + up_block_types (`Tuple[str]`, *optional*, defaults to `("UpDecoderBlock2D",)`): + Tuple of upsample block types. + block_out_channels (`Tuple[int]`, *optional*, defaults to `(64,)`): + Tuple of block output channels. + act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use. + latent_channels (`int`, *optional*, defaults to 4): Number of channels in the latent space. + sample_size (`int`, *optional*, defaults to `32`): Sample input size. + scaling_factor (`float`, *optional*, defaults to 0.18215): + The component-wise standard deviation of the trained latent space computed using the first batch of the + training set. This is used to scale the latent space to have unit variance when training the diffusion + model. The latents are scaled with the formula `z = z * scaling_factor` before being passed to the + diffusion model. When decoding, the latents are scaled back to the original scale with the formula: `z = 1 + / scaling_factor * z`. For more details, refer to sections 4.3.2 and D.1 of the [High-Resolution Image + Synthesis with Latent Diffusion Models](https://arxiv.org/abs/2112.10752) paper. + force_upcast (`bool`, *optional*, default to `True`): + If enabled it will force the VAE to run in float32 for high image resolution pipelines, such as SD-XL. VAE + can be fine-tuned / trained to a lower range without loosing too much precision in which case + `force_upcast` can be set to `False` - see: https://huggingface.co/madebyollin/sdxl-vae-fp16-fix + """ + + _supports_gradient_checkpointing = True + + @register_to_config + def __init__( + self, + in_channels: int = 3, + out_channels: int = 3, + down_block_types: Tuple[str] = ("DownEncoderBlock2D",), + up_block_types: Tuple[str] = ("UpDecoderBlock2D",), + block_out_channels: Tuple[int] = (64,), + layers_per_block: int = 1, + act_fn: str = "silu", + latent_channels: int = 4, + norm_num_groups: int = 32, + sample_size: int = 32, + scaling_factor: float = 0.18215, + force_upcast: float = True, + ): + super().__init__() + + # pass init params to Encoder + self.encoder = Encoder( + in_channels=in_channels, + out_channels=latent_channels, + down_block_types=down_block_types, + block_out_channels=block_out_channels, + layers_per_block=layers_per_block, + act_fn=act_fn, + norm_num_groups=norm_num_groups, + double_z=True, + ) + + # pass init params to Decoder + self.decoder = Decoder( + in_channels=latent_channels, + out_channels=out_channels, + up_block_types=up_block_types, + block_out_channels=block_out_channels, + layers_per_block=layers_per_block, + norm_num_groups=norm_num_groups, + act_fn=act_fn, + ) + + self.quant_conv = nn.Conv2d(2 * latent_channels, 2 * latent_channels, 1) + self.post_quant_conv = nn.Conv2d(latent_channels, latent_channels, 1) + + self.use_slicing = False + self.use_tiling = False + + # only relevant if vae tiling is enabled + self.tile_sample_min_size = self.config.sample_size + sample_size = ( + self.config.sample_size[0] + if isinstance(self.config.sample_size, (list, tuple)) + else self.config.sample_size + ) + self.tile_latent_min_size = int(sample_size / (2 ** (len(self.config.block_out_channels) - 1))) + self.tile_overlap_factor = 0.25 + + def _set_gradient_checkpointing(self, module, value=False): + if isinstance(module, (Encoder, Decoder)): + module.gradient_checkpointing = value + + def enable_tiling(self, use_tiling: bool = True): + r""" + Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to + compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow + processing larger images. + """ + self.use_tiling = use_tiling + + def disable_tiling(self): + r""" + Disable tiled VAE decoding. If `enable_tiling` was previously enabled, this method will go back to computing + decoding in one step. + """ + self.enable_tiling(False) + + def enable_slicing(self): + r""" + Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to + compute decoding in several steps. This is useful to save some memory and allow larger batch sizes. + """ + self.use_slicing = True + + def disable_slicing(self): + r""" + Disable sliced VAE decoding. If `enable_slicing` was previously enabled, this method will go back to computing + decoding in one step. + """ + self.use_slicing = False + + @property + # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors + def attn_processors(self) -> Dict[str, AttentionProcessor]: + r""" + Returns: + `dict` of attention processors: A dictionary containing all attention processors used in the model with + indexed by its weight name. + """ + # set recursively + processors = {} + + def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]): + if hasattr(module, "get_processor"): + processors[f"{name}.processor"] = module.get_processor(return_deprecated_lora=True) + + for sub_name, child in module.named_children(): + fn_recursive_add_processors(f"{name}.{sub_name}", child, processors) + + return processors + + for name, module in self.named_children(): + fn_recursive_add_processors(name, module, processors) + + return processors + + # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor + def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]): + r""" + Sets the attention processor to use to compute attention. + + Parameters: + processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`): + The instantiated processor class or a dictionary of processor classes that will be set as the processor + for **all** `Attention` layers. + + If `processor` is a dict, the key needs to define the path to the corresponding cross attention + processor. This is strongly recommended when setting trainable attention processors. + + """ + count = len(self.attn_processors.keys()) + + if isinstance(processor, dict) and len(processor) != count: + raise ValueError( + f"A dict of processors was passed, but the number of processors {len(processor)} does not match the" + f" number of attention layers: {count}. Please make sure to pass {count} processor classes." + ) + + def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor): + if hasattr(module, "set_processor"): + if not isinstance(processor, dict): + module.set_processor(processor) + else: + module.set_processor(processor.pop(f"{name}.processor")) + + for sub_name, child in module.named_children(): + fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor) + + for name, module in self.named_children(): + fn_recursive_attn_processor(name, module, processor) + + # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor + def set_default_attn_processor(self): + """ + Disables custom attention processors and sets the default attention implementation. + """ + if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()): + processor = AttnAddedKVProcessor() + elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()): + processor = AttnProcessor() + else: + raise ValueError( + f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}" + ) + + self.set_attn_processor(processor) + + @apply_forward_hook + def encode( + self, x: torch.FloatTensor, return_dict: bool = True + ) -> Union[AutoencoderKLOutput, Tuple[DiagonalGaussianDistribution]]: + """ + Encode a batch of images into latents. + + Args: + x (`torch.FloatTensor`): Input batch of images. + return_dict (`bool`, *optional*, defaults to `True`): + Whether to return a [`~models.autoencoder_kl.AutoencoderKLOutput`] instead of a plain tuple. + + Returns: + The latent representations of the encoded images. If `return_dict` is True, a + [`~models.autoencoder_kl.AutoencoderKLOutput`] is returned, otherwise a plain `tuple` is returned. + """ + if self.use_tiling and (x.shape[-1] > self.tile_sample_min_size or x.shape[-2] > self.tile_sample_min_size): + return self.tiled_encode(x, return_dict=return_dict) + + if self.use_slicing and x.shape[0] > 1: + encoded_slices = [self.encoder(x_slice) for x_slice in x.split(1)] + h = torch.cat(encoded_slices) + else: + h = self.encoder(x) + + moments = self.quant_conv(h) + posterior = DiagonalGaussianDistribution(moments) + + if not return_dict: + return (posterior,) + + return AutoencoderKLOutput(latent_dist=posterior) + + def _decode(self, z: torch.FloatTensor, return_dict: bool = True) -> Union[DecoderOutput, torch.FloatTensor]: + if self.use_tiling and (z.shape[-1] > self.tile_latent_min_size or z.shape[-2] > self.tile_latent_min_size): + return self.tiled_decode(z, return_dict=return_dict) + + z = self.post_quant_conv(z) + dec = self.decoder(z) + + if not return_dict: + return (dec,) + + return DecoderOutput(sample=dec) + + @apply_forward_hook + def decode( + self, z: torch.FloatTensor, return_dict: bool = True, generator=None + ) -> Union[DecoderOutput, torch.FloatTensor]: + """ + Decode a batch of images. + + Args: + z (`torch.FloatTensor`): Input batch of latent vectors. + return_dict (`bool`, *optional*, defaults to `True`): + Whether to return a [`~models.vae.DecoderOutput`] instead of a plain tuple. + + Returns: + [`~models.vae.DecoderOutput`] or `tuple`: + If return_dict is True, a [`~models.vae.DecoderOutput`] is returned, otherwise a plain `tuple` is + returned. + + """ + if self.use_slicing and z.shape[0] > 1: + decoded_slices = [self._decode(z_slice).sample for z_slice in z.split(1)] + decoded = torch.cat(decoded_slices) + else: + decoded = self._decode(z).sample + + if not return_dict: + return (decoded,) + + return DecoderOutput(sample=decoded) + + def blend_v(self, a: torch.Tensor, b: torch.Tensor, blend_extent: int) -> torch.Tensor: + blend_extent = min(a.shape[2], b.shape[2], blend_extent) + for y in range(blend_extent): + b[:, :, y, :] = a[:, :, -blend_extent + y, :] * (1 - y / blend_extent) + b[:, :, y, :] * (y / blend_extent) + return b + + def blend_h(self, a: torch.Tensor, b: torch.Tensor, blend_extent: int) -> torch.Tensor: + blend_extent = min(a.shape[3], b.shape[3], blend_extent) + for x in range(blend_extent): + b[:, :, :, x] = a[:, :, :, -blend_extent + x] * (1 - x / blend_extent) + b[:, :, :, x] * (x / blend_extent) + return b + + def tiled_encode(self, x: torch.FloatTensor, return_dict: bool = True) -> AutoencoderKLOutput: + r"""Encode a batch of images using a tiled encoder. + + When this option is enabled, the VAE will split the input tensor into tiles to compute encoding in several + steps. This is useful to keep memory use constant regardless of image size. The end result of tiled encoding is + different from non-tiled encoding because each tile uses a different encoder. To avoid tiling artifacts, the + tiles overlap and are blended together to form a smooth output. You may still see tile-sized changes in the + output, but they should be much less noticeable. + + Args: + x (`torch.FloatTensor`): Input batch of images. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~models.autoencoder_kl.AutoencoderKLOutput`] instead of a plain tuple. + + Returns: + [`~models.autoencoder_kl.AutoencoderKLOutput`] or `tuple`: + If return_dict is True, a [`~models.autoencoder_kl.AutoencoderKLOutput`] is returned, otherwise a plain + `tuple` is returned. + """ + overlap_size = int(self.tile_sample_min_size * (1 - self.tile_overlap_factor)) + blend_extent = int(self.tile_latent_min_size * self.tile_overlap_factor) + row_limit = self.tile_latent_min_size - blend_extent + + # Split the image into 512x512 tiles and encode them separately. + rows = [] + for i in range(0, x.shape[2], overlap_size): + row = [] + for j in range(0, x.shape[3], overlap_size): + tile = x[:, :, i : i + self.tile_sample_min_size, j : j + self.tile_sample_min_size] + tile = self.encoder(tile) + tile = self.quant_conv(tile) + row.append(tile) + rows.append(row) + result_rows = [] + for i, row in enumerate(rows): + result_row = [] + for j, tile in enumerate(row): + # blend the above tile and the left tile + # to the current tile and add the current tile to the result row + if i > 0: + tile = self.blend_v(rows[i - 1][j], tile, blend_extent) + if j > 0: + tile = self.blend_h(row[j - 1], tile, blend_extent) + result_row.append(tile[:, :, :row_limit, :row_limit]) + result_rows.append(torch.cat(result_row, dim=3)) + + moments = torch.cat(result_rows, dim=2) + posterior = DiagonalGaussianDistribution(moments) + + if not return_dict: + return (posterior,) + + return AutoencoderKLOutput(latent_dist=posterior) + + def tiled_decode(self, z: torch.FloatTensor, return_dict: bool = True) -> Union[DecoderOutput, torch.FloatTensor]: + r""" + Decode a batch of images using a tiled decoder. + + Args: + z (`torch.FloatTensor`): Input batch of latent vectors. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~models.vae.DecoderOutput`] instead of a plain tuple. + + Returns: + [`~models.vae.DecoderOutput`] or `tuple`: + If return_dict is True, a [`~models.vae.DecoderOutput`] is returned, otherwise a plain `tuple` is + returned. + """ + overlap_size = int(self.tile_latent_min_size * (1 - self.tile_overlap_factor)) + blend_extent = int(self.tile_sample_min_size * self.tile_overlap_factor) + row_limit = self.tile_sample_min_size - blend_extent + + # Split z into overlapping 64x64 tiles and decode them separately. + # The tiles have an overlap to avoid seams between tiles. + rows = [] + for i in range(0, z.shape[2], overlap_size): + row = [] + for j in range(0, z.shape[3], overlap_size): + tile = z[:, :, i : i + self.tile_latent_min_size, j : j + self.tile_latent_min_size] + tile = self.post_quant_conv(tile) + decoded = self.decoder(tile) + row.append(decoded) + rows.append(row) + result_rows = [] + for i, row in enumerate(rows): + result_row = [] + for j, tile in enumerate(row): + # blend the above tile and the left tile + # to the current tile and add the current tile to the result row + if i > 0: + tile = self.blend_v(rows[i - 1][j], tile, blend_extent) + if j > 0: + tile = self.blend_h(row[j - 1], tile, blend_extent) + result_row.append(tile[:, :, :row_limit, :row_limit]) + result_rows.append(torch.cat(result_row, dim=3)) + + dec = torch.cat(result_rows, dim=2) + if not return_dict: + return (dec,) + + return DecoderOutput(sample=dec) + + def forward( + self, + sample: torch.FloatTensor, + sample_posterior: bool = False, + return_dict: bool = True, + generator: Optional[torch.Generator] = None, + ) -> Union[DecoderOutput, torch.FloatTensor]: + r""" + Args: + sample (`torch.FloatTensor`): Input sample. + sample_posterior (`bool`, *optional*, defaults to `False`): + Whether to sample from the posterior. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`DecoderOutput`] instead of a plain tuple. + """ + x = sample + posterior = self.encode(x).latent_dist + if sample_posterior: + z = posterior.sample(generator=generator) + else: + z = posterior.mode() + dec = self.decode(z).sample + + if not return_dict: + return (dec,) + + return DecoderOutput(sample=dec) + + # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.fuse_qkv_projections + def fuse_qkv_projections(self): + """ + Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, + key, value) are fused. For cross-attention modules, key and value projection matrices are fused. + + + + This API is 🧪 experimental. + + + """ + self.original_attn_processors = None + + for _, attn_processor in self.attn_processors.items(): + if "Added" in str(attn_processor.__class__.__name__): + raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.") + + self.original_attn_processors = self.attn_processors + + for module in self.modules(): + if isinstance(module, Attention): + module.fuse_projections(fuse=True) + + # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.unfuse_qkv_projections + def unfuse_qkv_projections(self): + """Disables the fused QKV projection if enabled. + + + + This API is 🧪 experimental. + + + + """ + if self.original_attn_processors is not None: + self.set_attn_processor(self.original_attn_processors) diff --git a/MindIE/MultiModal/StableDiffusion-2.1/README.md b/MindIE/MultiModal/StableDiffusion-2.1/README.md new file mode 100644 index 0000000000000000000000000000000000000000..ed45b3b49bb61a403e79b190a8ad91ebc8a8af57 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-2.1/README.md @@ -0,0 +1,129 @@ +--- +license: apache-2.0 +pipeline_tag: text-to-image +frameworks: + - PyTorch +library_name: openmind +hardwares: + - NPU +language: + - en +--- +## 一、准备运行环境 + + **表 1** 版本配套表 + + | 配套 | 版本 | 环境准备指导 | + | ----- | ----- |-----| + | Python | 3.10.2 | - | + | torch | 2.1.0 | - | + +### 1.1 获取CANN&MindIE安装包&环境准备 +- 设备支持 +Atlas 800I A2推理设备：支持的卡数为1或2 +Atlas 300I Duo推理卡：支持的卡数为1，可双芯并行 +- [Atlas 800I A2](https://www.hiascend.com/developer/download/community/result?module=pt+ie+cann&product=4&model=32) +- [Atlas 300I Duo](https://www.hiascend.com/developer/download/community/result?module=pt+ie+cann&product=2&model=17) +- [环境准备指导](https://www.hiascend.com/document/detail/zh/CANNCommunityEdition/80RC2alpha002/softwareinst/instg/instg_0001.html) + +### 1.2 CANN安装 +```shell +# 增加软件包可执行权限，{version}表示软件版本号，{arch}表示CPU架构，{soc}表示昇腾AI处理器的版本。 +chmod +x ./Ascend-cann-toolkit_{version}_linux-{arch}.run +chmod +x ./Ascend-cann-kernels-{soc}_{version}_linux.run +# 校验软件包安装文件的一致性和完整性 +./Ascend-cann-toolkit_{version}_linux-{arch}.run --check +./Ascend-cann-kernels-{soc}_{version}_linux.run --check +# 安装 +./Ascend-cann-toolkit_{version}_linux-{arch}.run --install +./Ascend-cann-kernels-{soc}_{version}_linux.run --install + +# 设置环境变量 +source /usr/local/Ascend/ascend-toolkit/set_env.sh +``` + + +### 1.3 MindIE安装 +```shell +# 增加软件包可执行权限，{version}表示软件版本号，{arch}表示CPU架构。 +chmod +x ./Ascend-mindie_${version}_linux-${arch}.run +./Ascend-mindie_${version}_linux-${arch}.run --check + +# 方式一：默认路径安装 +./Ascend-mindie_${version}_linux-${arch}.run --install +# 设置环境变量 +cd /usr/local/Ascend/mindie && source set_env.sh + +# 方式二：指定路径安装 +./Ascend-mindie_${version}_linux-${arch}.run --install-path=${AieInstallPath} +# 设置环境变量 +cd ${AieInstallPath}/mindie && source set_env.sh +``` + +### 1.4 Torch_npu安装 +安装pytorch框架 版本2.1.0 +[安装包下载](https://download.pytorch.org/whl/cpu/torch/) + +使用pip安装 +```shell +# {version}表示软件版本号，{arch}表示CPU架构。 +pip install torch-${version}-cp310-cp310-linux_${arch}.whl +``` +下载 pytorch_v{pytorchversion}_py{pythonversion}.tar.gz +```shell +tar -xzvf pytorch_v{pytorchversion}_py{pythonversion}.tar.gz +# 解压后，会有whl包 +pip install torch_npu-{pytorchversion}.xxxx.{arch}.whl +``` +## 二、下载本仓库 + +### 2.1 下载到本地 +```shell +git clone https://gitee.com/ascend/ModelZoo-PyTorch.git +``` + +### 2.2 环境依赖安装 +```bash +pip3 install -r requirements.txt +``` + +## 三、Stable-Diffusion-2.1 使用 + +### 3.1 权重及配置文件说明 +下载权重和配置文件 +```shell +# stable-diffusion-v2.1: +git clone https://huggingface.co/stabilityai/stable-diffusion-2-1 +``` +### 3.2 修改配置文件 +将model_index.json中所有的`diffusers`字段修改为`stablediffusion` + +### 3.3 单卡功能测试 +设置权重路径 +```shell +model_base='./stable-diffusion-2-1-base' +``` +执行命令： +```shell +python3 inference_stablediffusion.py \ + --model ${model_base} \ + --prompt_file ./prompts/prompts.txt \ + --steps 50 \ + --save_dir ./results \ + --device 0 +``` +参数说明： +- --model：模型权重路径。 +- --prompt_file：提示词文件。 +- --steps: 图片生成迭代次数。 +- --save_dir：生成图片的存放目录。 +- --device：推理设备ID。 + + +### 3.4 模型推理性能 + +性能参考下列数据。 + +| 硬件形态 | 迭代次数 | 平均耗时| +| :------: |:----:|:----:| +| Atlas 800I A2(8*32G) | 50 | 2.619s | \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-2.1/inference_stablediffusion.py b/MindIE/MultiModal/StableDiffusion-2.1/inference_stablediffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..6dbc957391cc02ce148d1bbf449ef4d071fca53b --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-2.1/inference_stablediffusion.py @@ -0,0 +1,93 @@ +import torch +import torch_npu +import time +import os +import argparse +from stablediffusion import StableDiffusionPipeline + + +def parse_arguments(): + parser = argparse.ArgumentParser() + parser.add_argument( + "--prompt_file", + type=str, + default="./prompts/prompts.txt", + help="The prompts file to guide images generation.", + ) + parser.add_argument( + "--negative_prompt", + type=str, + default="", + help="The prompt or prompts to guide what to not include in image generation.", + ) + parser.add_argument( + "--steps", + type=int, + default=50, + help="The number of denoising steps.", + ) + parser.add_argument( + "--model", + type=str, + default="./stable-diffusion-v1.5", + help="The path of stable-diffusion.", + ) + parser.add_argument( + "--device", + type=int, + default=0, + help="NPU device id.", + ) + parser.add_argument( + "--save_dir", + type=str, + default="./results", + help="Path to save result audio files.", + ) + return parser.parse_args() + + +def main(): + args = parse_arguments() + save_dir = args.save_dir + if not os.path.exists(save_dir): + os.makedirs(save_dir) + + torch_npu.npu.set_device(args.device) + torch.manual_seed(1) + npu_stream = torch_npu.npu.Stream() + + pipe = StableDiffusionPipeline.from_pretrained(args.model, torch_dtype=torch.float16) + pipe.to("npu") + + total_time = 0 + prompts_num = 0 + average_time = 0 + skip = 2 + with os.fdopen(os.open(args.prompt_file, os.O_RDONLY), "r") as f: + for i, prompt in enumerate(f): + with torch.no_grad(): + npu_stream.synchronize() + begin = time.time() + image = pipe( + prompt=prompt, + negative_prompt=args.negative_prompt, + num_inference_steps=args.steps, + ).images[0] + npu_stream.synchronize() + end = time.time() + if i > skip - 1: + total_time += end - begin + prompts_num = i+1 + image_save_path = os.path.join(save_dir, f"images_{i}.png") + image.save(image_save_path) + if prompts_num > skip: + average_time = total_time / (prompts_num-skip) + else: + raise ValueError("Infer average time skip first two prompts, ensure that prompts.txt \ + contains more than three prompts") + print(f"Infer average time: {average_time:.3f}s\n") + + +if __name__ == "__main__": + main() \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-2.1/prompts/prompts.txt b/MindIE/MultiModal/StableDiffusion-2.1/prompts/prompts.txt new file mode 100644 index 0000000000000000000000000000000000000000..a375a0bb63931d0d5da6c6d91df1e14f870f47d0 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-2.1/prompts/prompts.txt @@ -0,0 +1,16 @@ +Beautiful illustration of The ocean. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Islands in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Seaports in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of The waves. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Grassland. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Wheat. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Hut Tong. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of The boat. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Pine trees. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Bamboo. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of The temple. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Cloud in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Sun in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Spring. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Lotus. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Snow piles. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-2.1/requirements.txt b/MindIE/MultiModal/StableDiffusion-2.1/requirements.txt new file mode 100644 index 0000000000000000000000000000000000000000..6d39028a97778341c84978d9643b966f25895618 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-2.1/requirements.txt @@ -0,0 +1,5 @@ +setuptools==57.5.0 +torch==2.1.0 +diffusers==0.26.3 +transformers==4.46.0 +open_clip_torch==2.20.0 \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/__init__.py b/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..529145f57160863a45c35608f5ffcb6fbbc23e99 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/__init__.py @@ -0,0 +1,5 @@ +from .pipeline import StableDiffusionPipeline, DiffusionPipeline +from .models import UNet2DConditionModel +from .models.model_utils import ModelMixin +from .schedulers import PNDMScheduler, SchedulerMixin +from .vae import AutoencoderKL \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/layers/__init__.py b/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/layers/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/layers/activations.py b/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/layers/activations.py new file mode 100644 index 0000000000000000000000000000000000000000..c045238004962da70fad9a62b8b4a25f200a1e95 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/layers/activations.py @@ -0,0 +1,123 @@ +# coding=utf-8 +# Copyright 2023 HuggingFace Inc. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import torch +import torch.nn.functional as F +from torch import nn + +from diffusers.utils import USE_PEFT_BACKEND +from diffusers.models.lora import LoRACompatibleLinear + + +ACTIVATION_FUNCTIONS = { + "swish": nn.SiLU(), + "silu": nn.SiLU(), + "mish": nn.Mish(), + "gelu": nn.GELU(), + "relu": nn.ReLU(), +} + + +def get_activation(act_fn: str) -> nn.Module: + """Helper function to get activation function from string. + + Args: + act_fn (str): Name of activation function. + + Returns: + nn.Module: Activation function. + """ + + act_fn = act_fn.lower() + if act_fn in ACTIVATION_FUNCTIONS: + return ACTIVATION_FUNCTIONS[act_fn] + else: + raise ValueError(f"Unsupported activation function: {act_fn}") + + +class GELU(nn.Module): + r""" + GELU activation function with tanh approximation support with `approximate="tanh"`. + + Parameters: + dim_in (`int`): The number of channels in the input. + dim_out (`int`): The number of channels in the output. + approximate (`str`, *optional*, defaults to `"none"`): If `"tanh"`, use tanh approximation. + bias (`bool`, defaults to True): Whether to use a bias in the linear layer. + """ + + def __init__(self, dim_in: int, dim_out: int, approximate: str = "none", bias: bool = True): + super().__init__() + self.proj = nn.Linear(dim_in, dim_out, bias=bias) + self.approximate = approximate + + def gelu(self, gate: torch.Tensor) -> torch.Tensor: + if gate.device.type != "mps": + return F.gelu(gate, approximate=self.approximate) + # mps: gelu is not implemented for float16 + return F.gelu(gate.to(dtype=torch.float32), approximate=self.approximate).to(dtype=gate.dtype) + + def forward(self, hidden_states): + hidden_states = self.proj(hidden_states) + hidden_states = self.gelu(hidden_states) + return hidden_states + + +class GEGLU(nn.Module): + r""" + A [variant](https://arxiv.org/abs/2002.05202) of the gated linear unit activation function. + + Parameters: + dim_in (`int`): The number of channels in the input. + dim_out (`int`): The number of channels in the output. + bias (`bool`, defaults to True): Whether to use a bias in the linear layer. + """ + + def __init__(self, dim_in: int, dim_out: int, bias: bool = True): + super().__init__() + linear_cls = LoRACompatibleLinear if not USE_PEFT_BACKEND else nn.Linear + + self.proj = linear_cls(dim_in, dim_out * 2, bias=bias) + + def gelu(self, gate: torch.Tensor) -> torch.Tensor: + if gate.device.type != "mps": + return F.gelu(gate) + # mps: gelu is not implemented for float16 + return F.gelu(gate.to(dtype=torch.float32)).to(dtype=gate.dtype) + + def forward(self, hidden_states, scale: float = 1.0): + args = () if USE_PEFT_BACKEND else (scale,) + hidden_states, gate = self.proj(hidden_states, *args).chunk(2, dim=-1) + return hidden_states * self.gelu(gate) + + +class ApproximateGELU(nn.Module): + r""" + The approximate form of the Gaussian Error Linear Unit (GELU). For more details, see section 2 of this + [paper](https://arxiv.org/abs/1606.08415). + + Parameters: + dim_in (`int`): The number of channels in the input. + dim_out (`int`): The number of channels in the output. + bias (`bool`, defaults to True): Whether to use a bias in the linear layer. + """ + + def __init__(self, dim_in: int, dim_out: int, bias: bool = True): + super().__init__() + self.proj = nn.Linear(dim_in, dim_out, bias=bias) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + x = self.proj(x) + return x * torch.sigmoid(1.702 * x) diff --git a/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/layers/attention.py b/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/layers/attention.py new file mode 100644 index 0000000000000000000000000000000000000000..9b2d570658f9af3e08ca74ffe65f9add1065f1a3 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/layers/attention.py @@ -0,0 +1,676 @@ +# Copyright 2023 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from typing import Any, Dict, Optional + +import torch +import torch.nn.functional as F +from torch import nn + +from diffusers.utils import USE_PEFT_BACKEND +from diffusers.utils.torch_utils import maybe_allow_in_graph +from diffusers.models.lora import LoRACompatibleLinear + +from .activations import GEGLU, GELU, ApproximateGELU +from .attention_processor import Attention +from .embeddings import SinusoidalPositionalEmbedding +from .normalization import AdaLayerNorm, AdaLayerNormContinuous, AdaLayerNormZero, RMSNorm + + +def _chunked_feed_forward( + ff: nn.Module, hidden_states: torch.Tensor, chunk_dim: int, chunk_size: int, lora_scale: Optional[float] = None +): + # "feed_forward_chunk_size" can be used to save memory + if hidden_states.shape[chunk_dim] % chunk_size != 0: + raise ValueError( + f"`hidden_states` dimension to be chunked: {hidden_states.shape[chunk_dim]} has to be divisible by chunk size: {chunk_size}. Make sure to set an appropriate `chunk_size` when calling `unet.enable_forward_chunking`." + ) + + num_chunks = hidden_states.shape[chunk_dim] // chunk_size + if lora_scale is None: + ff_output = torch.cat( + [ff(hid_slice) for hid_slice in hidden_states.chunk(num_chunks, dim=chunk_dim)], + dim=chunk_dim, + ) + else: + # TOOD(Patrick): LoRA scale can be removed once PEFT refactor is complete + ff_output = torch.cat( + [ff(hid_slice, scale=lora_scale) for hid_slice in hidden_states.chunk(num_chunks, dim=chunk_dim)], + dim=chunk_dim, + ) + + return ff_output + + +@maybe_allow_in_graph +class GatedSelfAttentionDense(nn.Module): + r""" + A gated self-attention dense layer that combines visual features and object features. + + Parameters: + query_dim (`int`): The number of channels in the query. + context_dim (`int`): The number of channels in the context. + n_heads (`int`): The number of heads to use for attention. + d_head (`int`): The number of channels in each head. + """ + + def __init__(self, query_dim: int, context_dim: int, n_heads: int, d_head: int): + super().__init__() + + # we need a linear projection since we need cat visual feature and obj feature + self.linear = nn.Linear(context_dim, query_dim) + + self.attn = Attention(query_dim=query_dim, heads=n_heads, dim_head=d_head) + self.ff = FeedForward(query_dim, activation_fn="geglu") + + self.norm1 = nn.LayerNorm(query_dim) + self.norm2 = nn.LayerNorm(query_dim) + + self.register_parameter("alpha_attn", nn.Parameter(torch.tensor(0.0))) + self.register_parameter("alpha_dense", nn.Parameter(torch.tensor(0.0))) + + self.enabled = True + + def forward(self, x: torch.Tensor, objs: torch.Tensor) -> torch.Tensor: + if not self.enabled: + return x + + n_visual = x.shape[1] + objs = self.linear(objs) + + x = x + self.alpha_attn.tanh() * self.attn(self.norm1(torch.cat([x, objs], dim=1)))[:, :n_visual, :] + x = x + self.alpha_dense.tanh() * self.ff(self.norm2(x)) + + return x + + +@maybe_allow_in_graph +class BasicTransformerBlock(nn.Module): + r""" + A basic Transformer block. + + Parameters: + dim (`int`): The number of channels in the input and output. + num_attention_heads (`int`): The number of heads to use for multi-head attention. + attention_head_dim (`int`): The number of channels in each head. + dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. + cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention. + activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward. + num_embeds_ada_norm (: + obj: `int`, *optional*): The number of diffusion steps used during training. See `Transformer2DModel`. + attention_bias (: + obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter. + only_cross_attention (`bool`, *optional*): + Whether to use only cross-attention layers. In this case two cross attention layers are used. + double_self_attention (`bool`, *optional*): + Whether to use two self-attention layers. In this case no cross attention layers are used. + upcast_attention (`bool`, *optional*): + Whether to upcast the attention computation to float32. This is useful for mixed precision training. + norm_elementwise_affine (`bool`, *optional*, defaults to `True`): + Whether to use learnable elementwise affine parameters for normalization. + norm_type (`str`, *optional*, defaults to `"layer_norm"`): + The normalization layer to use. Can be `"layer_norm"`, `"ada_norm"` or `"ada_norm_zero"`. + final_dropout (`bool` *optional*, defaults to False): + Whether to apply a final dropout after the last feed-forward layer. + attention_type (`str`, *optional*, defaults to `"default"`): + The type of attention to use. Can be `"default"` or `"gated"` or `"gated-text-image"`. + positional_embeddings (`str`, *optional*, defaults to `None`): + The type of positional embeddings to apply to. + num_positional_embeddings (`int`, *optional*, defaults to `None`): + The maximum number of positional embeddings to apply. + """ + + def __init__( + self, + dim: int, + num_attention_heads: int, + attention_head_dim: int, + dropout=0.0, + cross_attention_dim: Optional[int] = None, + activation_fn: str = "geglu", + num_embeds_ada_norm: Optional[int] = None, + attention_bias: bool = False, + only_cross_attention: bool = False, + double_self_attention: bool = False, + upcast_attention: bool = False, + norm_elementwise_affine: bool = True, + norm_type: str = "layer_norm", # 'layer_norm', 'ada_norm', 'ada_norm_zero', 'ada_norm_single', 'layer_norm_i2vgen' + norm_eps: float = 1e-5, + final_dropout: bool = False, + attention_type: str = "default", + positional_embeddings: Optional[str] = None, + num_positional_embeddings: Optional[int] = None, + ada_norm_continous_conditioning_embedding_dim: Optional[int] = None, + ada_norm_bias: Optional[int] = None, + ff_inner_dim: Optional[int] = None, + ff_bias: bool = True, + attention_out_bias: bool = True, + ): + super().__init__() + self.only_cross_attention = only_cross_attention + + self.use_ada_layer_norm_zero = (num_embeds_ada_norm is not None) and norm_type == "ada_norm_zero" + self.use_ada_layer_norm = (num_embeds_ada_norm is not None) and norm_type == "ada_norm" + self.use_ada_layer_norm_single = norm_type == "ada_norm_single" + self.use_layer_norm = norm_type == "layer_norm" + self.use_ada_layer_norm_continuous = norm_type == "ada_norm_continuous" + + if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None: + raise ValueError( + f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to" + f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}." + ) + + self.norm_type = norm_type + self.num_embeds_ada_norm = num_embeds_ada_norm + + if positional_embeddings and (num_positional_embeddings is None): + raise ValueError( + "If `positional_embedding` type is defined, `num_positition_embeddings` must also be defined." + ) + + if positional_embeddings == "sinusoidal": + self.pos_embed = SinusoidalPositionalEmbedding(dim, max_seq_length=num_positional_embeddings) + else: + self.pos_embed = None + + # Define 3 blocks. Each block has its own normalization layer. + # 1. Self-Attn + if norm_type == "ada_norm": + self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm) + elif norm_type == "ada_norm_zero": + self.norm1 = AdaLayerNormZero(dim, num_embeds_ada_norm) + elif norm_type == "ada_norm_continuous": + self.norm1 = AdaLayerNormContinuous( + dim, + ada_norm_continous_conditioning_embedding_dim, + norm_elementwise_affine, + norm_eps, + ada_norm_bias, + "rms_norm", + ) + else: + self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps) + + self.attn1 = Attention( + query_dim=dim, + heads=num_attention_heads, + dim_head=attention_head_dim, + dropout=dropout, + bias=attention_bias, + cross_attention_dim=cross_attention_dim if only_cross_attention else None, + upcast_attention=upcast_attention, + out_bias=attention_out_bias, + ) + + # 2. Cross-Attn + if cross_attention_dim is not None or double_self_attention: + # We currently only use AdaLayerNormZero for self attention where there will only be one attention block. + # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during + # the second cross attention block. + if norm_type == "ada_norm": + self.norm2 = AdaLayerNorm(dim, num_embeds_ada_norm) + elif norm_type == "ada_norm_continuous": + self.norm2 = AdaLayerNormContinuous( + dim, + ada_norm_continous_conditioning_embedding_dim, + norm_elementwise_affine, + norm_eps, + ada_norm_bias, + "rms_norm", + ) + else: + self.norm2 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine) + + self.attn2 = Attention( + query_dim=dim, + cross_attention_dim=cross_attention_dim if not double_self_attention else None, + heads=num_attention_heads, + dim_head=attention_head_dim, + dropout=dropout, + bias=attention_bias, + upcast_attention=upcast_attention, + out_bias=attention_out_bias, + ) # is self-attn if encoder_hidden_states is none + else: + self.norm2 = None + self.attn2 = None + + # 3. Feed-forward + if norm_type == "ada_norm_continuous": + self.norm3 = AdaLayerNormContinuous( + dim, + ada_norm_continous_conditioning_embedding_dim, + norm_elementwise_affine, + norm_eps, + ada_norm_bias, + "layer_norm", + ) + + elif norm_type in ["ada_norm_zero", "ada_norm", "layer_norm", "ada_norm_continuous"]: + self.norm3 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine) + elif norm_type == "layer_norm_i2vgen": + self.norm3 = None + + self.ff = FeedForward( + dim, + dropout=dropout, + activation_fn=activation_fn, + final_dropout=final_dropout, + inner_dim=ff_inner_dim, + bias=ff_bias, + ) + + # 4. Fuser + if attention_type == "gated" or attention_type == "gated-text-image": + self.fuser = GatedSelfAttentionDense(dim, cross_attention_dim, num_attention_heads, attention_head_dim) + + # 5. Scale-shift for PixArt-Alpha. + if norm_type == "ada_norm_single": + self.scale_shift_table = nn.Parameter(torch.randn(6, dim) / dim**0.5) + + # let chunk size default to None + self._chunk_size = None + self._chunk_dim = 0 + + def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int = 0): + # Sets chunk feed-forward + self._chunk_size = chunk_size + self._chunk_dim = dim + + def forward( + self, + hidden_states: torch.FloatTensor, + attention_mask: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + timestep: Optional[torch.LongTensor] = None, + cross_attention_kwargs: Dict[str, Any] = None, + class_labels: Optional[torch.LongTensor] = None, + added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None, + ) -> torch.FloatTensor: + # Notice that normalization is always applied before the real computation in the following blocks. + # 0. Self-Attention + batch_size = hidden_states.shape[0] + + if self.norm_type == "ada_norm": + norm_hidden_states = self.norm1(hidden_states, timestep) + elif self.norm_type == "ada_norm_zero": + norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1( + hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype + ) + elif self.norm_type in ["layer_norm", "layer_norm_i2vgen"]: + norm_hidden_states = self.norm1(hidden_states) + elif self.norm_type == "ada_norm_continuous": + norm_hidden_states = self.norm1(hidden_states, added_cond_kwargs["pooled_text_emb"]) + elif self.norm_type == "ada_norm_single": + shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = ( + self.scale_shift_table[None] + timestep.reshape(batch_size, 6, -1) + ).chunk(6, dim=1) + norm_hidden_states = self.norm1(hidden_states) + norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa + norm_hidden_states = norm_hidden_states.squeeze(1) + else: + raise ValueError("Incorrect norm used") + + if self.pos_embed is not None: + norm_hidden_states = self.pos_embed(norm_hidden_states) + + # 1. Retrieve lora scale. + lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0 + + # 2. Prepare GLIGEN inputs + cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {} + gligen_kwargs = cross_attention_kwargs.pop("gligen", None) + + attn_output = self.attn1( + norm_hidden_states, + encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None, + attention_mask=attention_mask, + **cross_attention_kwargs, + ) + if self.norm_type == "ada_norm_zero": + attn_output = gate_msa.unsqueeze(1) * attn_output + elif self.norm_type == "ada_norm_single": + attn_output = gate_msa * attn_output + + hidden_states = attn_output + hidden_states + if hidden_states.ndim == 4: + hidden_states = hidden_states.squeeze(1) + + # 2.5 GLIGEN Control + if gligen_kwargs is not None: + hidden_states = self.fuser(hidden_states, gligen_kwargs["objs"]) + + # 3. Cross-Attention + if self.attn2 is not None: + if self.norm_type == "ada_norm": + norm_hidden_states = self.norm2(hidden_states, timestep) + elif self.norm_type in ["ada_norm_zero", "layer_norm", "layer_norm_i2vgen"]: + norm_hidden_states = self.norm2(hidden_states) + elif self.norm_type == "ada_norm_single": + # For PixArt norm2 isn't applied here: + # https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L70C1-L76C103 + norm_hidden_states = hidden_states + elif self.norm_type == "ada_norm_continuous": + norm_hidden_states = self.norm2(hidden_states, added_cond_kwargs["pooled_text_emb"]) + else: + raise ValueError("Incorrect norm") + + if self.pos_embed is not None and self.norm_type != "ada_norm_single": + norm_hidden_states = self.pos_embed(norm_hidden_states) + + attn_output = self.attn2( + norm_hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=encoder_attention_mask, + **cross_attention_kwargs, + ) + hidden_states = attn_output + hidden_states + + # 4. Feed-forward + # i2vgen doesn't have this norm 🤷‍♂️ + if self.norm_type == "ada_norm_continuous": + norm_hidden_states = self.norm3(hidden_states, added_cond_kwargs["pooled_text_emb"]) + elif not self.norm_type == "ada_norm_single": + norm_hidden_states = self.norm3(hidden_states) + + if self.norm_type == "ada_norm_zero": + norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None] + + if self.norm_type == "ada_norm_single": + norm_hidden_states = self.norm2(hidden_states) + norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp + + if self._chunk_size is not None: + # "feed_forward_chunk_size" can be used to save memory + ff_output = _chunked_feed_forward( + self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size, lora_scale=lora_scale + ) + else: + ff_output = self.ff(norm_hidden_states, scale=lora_scale) + + if self.norm_type == "ada_norm_zero": + ff_output = gate_mlp.unsqueeze(1) * ff_output + elif self.norm_type == "ada_norm_single": + ff_output = gate_mlp * ff_output + + hidden_states = ff_output + hidden_states + if hidden_states.ndim == 4: + hidden_states = hidden_states.squeeze(1) + + return hidden_states + + +@maybe_allow_in_graph +class TemporalBasicTransformerBlock(nn.Module): + r""" + A basic Transformer block for video like data. + + Parameters: + dim (`int`): The number of channels in the input and output. + time_mix_inner_dim (`int`): The number of channels for temporal attention. + num_attention_heads (`int`): The number of heads to use for multi-head attention. + attention_head_dim (`int`): The number of channels in each head. + cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention. + """ + + def __init__( + self, + dim: int, + time_mix_inner_dim: int, + num_attention_heads: int, + attention_head_dim: int, + cross_attention_dim: Optional[int] = None, + ): + super().__init__() + self.is_res = dim == time_mix_inner_dim + + self.norm_in = nn.LayerNorm(dim) + + # Define 3 blocks. Each block has its own normalization layer. + # 1. Self-Attn + self.norm_in = nn.LayerNorm(dim) + self.ff_in = FeedForward( + dim, + dim_out=time_mix_inner_dim, + activation_fn="geglu", + ) + + self.norm1 = nn.LayerNorm(time_mix_inner_dim) + self.attn1 = Attention( + query_dim=time_mix_inner_dim, + heads=num_attention_heads, + dim_head=attention_head_dim, + cross_attention_dim=None, + ) + + # 2. Cross-Attn + if cross_attention_dim is not None: + # We currently only use AdaLayerNormZero for self attention where there will only be one attention block. + # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during + # the second cross attention block. + self.norm2 = nn.LayerNorm(time_mix_inner_dim) + self.attn2 = Attention( + query_dim=time_mix_inner_dim, + cross_attention_dim=cross_attention_dim, + heads=num_attention_heads, + dim_head=attention_head_dim, + ) # is self-attn if encoder_hidden_states is none + else: + self.norm2 = None + self.attn2 = None + + # 3. Feed-forward + self.norm3 = nn.LayerNorm(time_mix_inner_dim) + self.ff = FeedForward(time_mix_inner_dim, activation_fn="geglu") + + # let chunk size default to None + self._chunk_size = None + self._chunk_dim = None + + def set_chunk_feed_forward(self, chunk_size: Optional[int], **kwargs): + # Sets chunk feed-forward + self._chunk_size = chunk_size + # chunk dim should be hardcoded to 1 to have better speed vs. memory trade-off + self._chunk_dim = 1 + + def forward( + self, + hidden_states: torch.FloatTensor, + num_frames: int, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + ) -> torch.FloatTensor: + # Notice that normalization is always applied before the real computation in the following blocks. + # 0. Self-Attention + batch_size = hidden_states.shape[0] + + batch_frames, seq_length, channels = hidden_states.shape + batch_size = batch_frames // num_frames + + hidden_states = hidden_states[None, :].reshape(batch_size, num_frames, seq_length, channels) + hidden_states = hidden_states.permute(0, 2, 1, 3) + hidden_states = hidden_states.reshape(batch_size * seq_length, num_frames, channels) + + residual = hidden_states + hidden_states = self.norm_in(hidden_states) + + if self._chunk_size is not None: + hidden_states = _chunked_feed_forward(self.ff_in, hidden_states, self._chunk_dim, self._chunk_size) + else: + hidden_states = self.ff_in(hidden_states) + + if self.is_res: + hidden_states = hidden_states + residual + + norm_hidden_states = self.norm1(hidden_states) + attn_output = self.attn1(norm_hidden_states, encoder_hidden_states=None) + hidden_states = attn_output + hidden_states + + # 3. Cross-Attention + if self.attn2 is not None: + norm_hidden_states = self.norm2(hidden_states) + attn_output = self.attn2(norm_hidden_states, encoder_hidden_states=encoder_hidden_states) + hidden_states = attn_output + hidden_states + + # 4. Feed-forward + norm_hidden_states = self.norm3(hidden_states) + + if self._chunk_size is not None: + ff_output = _chunked_feed_forward(self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size) + else: + ff_output = self.ff(norm_hidden_states) + + if self.is_res: + hidden_states = ff_output + hidden_states + else: + hidden_states = ff_output + + hidden_states = hidden_states[None, :].reshape(batch_size, seq_length, num_frames, channels) + hidden_states = hidden_states.permute(0, 2, 1, 3) + hidden_states = hidden_states.reshape(batch_size * num_frames, seq_length, channels) + + return hidden_states + + +class SkipFFTransformerBlock(nn.Module): + def __init__( + self, + dim: int, + num_attention_heads: int, + attention_head_dim: int, + kv_input_dim: int, + kv_input_dim_proj_use_bias: bool, + dropout=0.0, + cross_attention_dim: Optional[int] = None, + attention_bias: bool = False, + attention_out_bias: bool = True, + ): + super().__init__() + if kv_input_dim != dim: + self.kv_mapper = nn.Linear(kv_input_dim, dim, kv_input_dim_proj_use_bias) + else: + self.kv_mapper = None + + self.norm1 = RMSNorm(dim, 1e-06) + + self.attn1 = Attention( + query_dim=dim, + heads=num_attention_heads, + dim_head=attention_head_dim, + dropout=dropout, + bias=attention_bias, + cross_attention_dim=cross_attention_dim, + out_bias=attention_out_bias, + ) + + self.norm2 = RMSNorm(dim, 1e-06) + + self.attn2 = Attention( + query_dim=dim, + cross_attention_dim=cross_attention_dim, + heads=num_attention_heads, + dim_head=attention_head_dim, + dropout=dropout, + bias=attention_bias, + out_bias=attention_out_bias, + ) + + def forward(self, hidden_states, encoder_hidden_states, cross_attention_kwargs): + cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {} + + if self.kv_mapper is not None: + encoder_hidden_states = self.kv_mapper(F.silu(encoder_hidden_states)) + + norm_hidden_states = self.norm1(hidden_states) + + attn_output = self.attn1( + norm_hidden_states, + encoder_hidden_states=encoder_hidden_states, + **cross_attention_kwargs, + ) + + hidden_states = attn_output + hidden_states + + norm_hidden_states = self.norm2(hidden_states) + + attn_output = self.attn2( + norm_hidden_states, + encoder_hidden_states=encoder_hidden_states, + **cross_attention_kwargs, + ) + + hidden_states = attn_output + hidden_states + + return hidden_states + + +class FeedForward(nn.Module): + r""" + A feed-forward layer. + + Parameters: + dim (`int`): The number of channels in the input. + dim_out (`int`, *optional*): The number of channels in the output. If not given, defaults to `dim`. + mult (`int`, *optional*, defaults to 4): The multiplier to use for the hidden dimension. + dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. + activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward. + final_dropout (`bool` *optional*, defaults to False): Apply a final dropout. + bias (`bool`, defaults to True): Whether to use a bias in the linear layer. + """ + + def __init__( + self, + dim: int, + dim_out: Optional[int] = None, + mult: int = 4, + dropout: float = 0.0, + activation_fn: str = "geglu", + final_dropout: bool = False, + inner_dim=None, + bias: bool = True, + ): + super().__init__() + if inner_dim is None: + inner_dim = int(dim * mult) + dim_out = dim_out if dim_out is not None else dim + linear_cls = LoRACompatibleLinear if not USE_PEFT_BACKEND else nn.Linear + + if activation_fn == "gelu": + act_fn = GELU(dim, inner_dim, bias=bias) + if activation_fn == "gelu-approximate": + act_fn = GELU(dim, inner_dim, approximate="tanh", bias=bias) + elif activation_fn == "geglu": + act_fn = GEGLU(dim, inner_dim, bias=bias) + elif activation_fn == "geglu-approximate": + act_fn = ApproximateGELU(dim, inner_dim, bias=bias) + + self.net = nn.ModuleList([]) + # project in + self.net.append(act_fn) + # project dropout + self.net.append(nn.Dropout(dropout)) + # project out + self.net.append(linear_cls(inner_dim, dim_out, bias=bias)) + # FF as used in Vision Transformer, MLP-Mixer, etc. have a final dropout + if final_dropout: + self.net.append(nn.Dropout(dropout)) + + def forward(self, hidden_states: torch.Tensor, scale: float = 1.0) -> torch.Tensor: + compatible_cls = (GEGLU,) if USE_PEFT_BACKEND else (GEGLU, LoRACompatibleLinear) + for module in self.net: + if isinstance(module, compatible_cls): + hidden_states = module(hidden_states, scale) + else: + hidden_states = module(hidden_states) + return hidden_states diff --git a/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/layers/attention_processor.py b/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/layers/attention_processor.py new file mode 100644 index 0000000000000000000000000000000000000000..ffc4fbfb232b83b85ae946ce2544eab02efb94b2 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/layers/attention_processor.py @@ -0,0 +1,1019 @@ +# Copyright 2023 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from importlib import import_module +from typing import Callable, Optional, Union + +import torch +import torch_npu +import torch.nn.functional as F +from torch import nn + +from diffusers.utils import USE_PEFT_BACKEND, deprecate, logging +from diffusers.utils.import_utils import is_xformers_available +from diffusers.models.attention_processor import ( + CustomDiffusionAttnProcessor, FusedAttnProcessor2_0, CustomDiffusionXFormersAttnProcessor, CustomDiffusionAttnProcessor2_0, + SpatialNorm, SlicedAttnProcessor, SlicedAttnAddedKVProcessor, LoRAAttnProcessor, LoRAAttnProcessor2_0, + LoRAXFormersAttnProcessor, LoRAAttnAddedKVProcessor, IPAdapterAttnProcessor, IPAdapterAttnProcessor2_0, + CustomDiffusionAttnProcessor, XFormersAttnAddedKVProcessor, XFormersAttnProcessor, +) +from diffusers.models.lora import LoRACompatibleLinear, LoRALinearLayer + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +class Attention(nn.Module): + r""" + A cross attention layer. + + Parameters: + query_dim (`int`): + The number of channels in the query. + cross_attention_dim (`int`, *optional*): + The number of channels in the encoder_hidden_states. If not given, defaults to `query_dim`. + heads (`int`, *optional*, defaults to 8): + The number of heads to use for multi-head attention. + dim_head (`int`, *optional*, defaults to 64): + The number of channels in each head. + dropout (`float`, *optional*, defaults to 0.0): + The dropout probability to use. + bias (`bool`, *optional*, defaults to False): + Set to `True` for the query, key, and value linear layers to contain a bias parameter. + upcast_attention (`bool`, *optional*, defaults to False): + Set to `True` to upcast the attention computation to `float32`. + upcast_softmax (`bool`, *optional*, defaults to False): + Set to `True` to upcast the softmax computation to `float32`. + cross_attention_norm (`str`, *optional*, defaults to `None`): + The type of normalization to use for the cross attention. Can be `None`, `layer_norm`, or `group_norm`. + cross_attention_norm_num_groups (`int`, *optional*, defaults to 32): + The number of groups to use for the group norm in the cross attention. + added_kv_proj_dim (`int`, *optional*, defaults to `None`): + The number of channels to use for the added key and value projections. If `None`, no projection is used. + norm_num_groups (`int`, *optional*, defaults to `None`): + The number of groups to use for the group norm in the attention. + spatial_norm_dim (`int`, *optional*, defaults to `None`): + The number of channels to use for the spatial normalization. + out_bias (`bool`, *optional*, defaults to `True`): + Set to `True` to use a bias in the output linear layer. + scale_qk (`bool`, *optional*, defaults to `True`): + Set to `True` to scale the query and key by `1 / sqrt(dim_head)`. + only_cross_attention (`bool`, *optional*, defaults to `False`): + Set to `True` to only use cross attention and not added_kv_proj_dim. Can only be set to `True` if + `added_kv_proj_dim` is not `None`. + eps (`float`, *optional*, defaults to 1e-5): + An additional value added to the denominator in group normalization that is used for numerical stability. + rescale_output_factor (`float`, *optional*, defaults to 1.0): + A factor to rescale the output by dividing it with this value. + residual_connection (`bool`, *optional*, defaults to `False`): + Set to `True` to add the residual connection to the output. + _from_deprecated_attn_block (`bool`, *optional*, defaults to `False`): + Set to `True` if the attention block is loaded from a deprecated state dict. + processor (`AttnProcessor`, *optional*, defaults to `None`): + The attention processor to use. If `None`, defaults to `AttnProcessor2_0` if `torch 2.x` is used and + `AttnProcessor` otherwise. + """ + + def __init__( + self, + query_dim: int, + cross_attention_dim: Optional[int] = None, + heads: int = 8, + dim_head: int = 64, + dropout: float = 0.0, + bias: bool = False, + upcast_attention: bool = False, + upcast_softmax: bool = False, + cross_attention_norm: Optional[str] = None, + cross_attention_norm_num_groups: int = 32, + added_kv_proj_dim: Optional[int] = None, + norm_num_groups: Optional[int] = None, + spatial_norm_dim: Optional[int] = None, + out_bias: bool = True, + scale_qk: bool = True, + only_cross_attention: bool = False, + eps: float = 1e-5, + rescale_output_factor: float = 1.0, + residual_connection: bool = False, + _from_deprecated_attn_block: bool = False, + processor: Optional["AttnProcessor"] = None, + out_dim: int = None, + ): + super().__init__() + self.inner_dim = out_dim if out_dim is not None else dim_head * heads + self.query_dim = query_dim + self.cross_attention_dim = cross_attention_dim if cross_attention_dim is not None else query_dim + self.upcast_attention = upcast_attention + self.upcast_softmax = upcast_softmax + self.rescale_output_factor = rescale_output_factor + self.residual_connection = residual_connection + self.dropout = dropout + self.fused_projections = False + self.out_dim = out_dim if out_dim is not None else query_dim + + # we make use of this private variable to know whether this class is loaded + # with an deprecated state dict so that we can convert it on the fly + self._from_deprecated_attn_block = _from_deprecated_attn_block + + self.scale_qk = scale_qk + self.scale = dim_head**-0.5 if self.scale_qk else 1.0 + + self.heads = out_dim // dim_head if out_dim is not None else heads + # for slice_size > 0 the attention score computation + # is split across the batch axis to save memory + # You can set slice_size with `set_attention_slice` + self.sliceable_head_dim = heads + + self.added_kv_proj_dim = added_kv_proj_dim + self.only_cross_attention = only_cross_attention + + if self.added_kv_proj_dim is None and self.only_cross_attention: + raise ValueError( + "`only_cross_attention` can only be set to True if `added_kv_proj_dim` is not None. Make sure to set either `only_cross_attention=False` or define `added_kv_proj_dim`." + ) + + if norm_num_groups is not None: + self.group_norm = nn.GroupNorm(num_channels=query_dim, num_groups=norm_num_groups, eps=eps, affine=True) + else: + self.group_norm = None + + if spatial_norm_dim is not None: + self.spatial_norm = SpatialNorm(f_channels=query_dim, zq_channels=spatial_norm_dim) + else: + self.spatial_norm = None + + if cross_attention_norm is None: + self.norm_cross = None + elif cross_attention_norm == "layer_norm": + self.norm_cross = nn.LayerNorm(self.cross_attention_dim) + elif cross_attention_norm == "group_norm": + if self.added_kv_proj_dim is not None: + # The given `encoder_hidden_states` are initially of shape + # (batch_size, seq_len, added_kv_proj_dim) before being projected + # to (batch_size, seq_len, cross_attention_dim). The norm is applied + # before the projection, so we need to use `added_kv_proj_dim` as + # the number of channels for the group norm. + norm_cross_num_channels = added_kv_proj_dim + else: + norm_cross_num_channels = self.cross_attention_dim + + self.norm_cross = nn.GroupNorm( + num_channels=norm_cross_num_channels, num_groups=cross_attention_norm_num_groups, eps=1e-5, affine=True + ) + else: + raise ValueError( + f"unknown cross_attention_norm: {cross_attention_norm}. Should be None, 'layer_norm' or 'group_norm'" + ) + + if USE_PEFT_BACKEND: + linear_cls = nn.Linear + else: + linear_cls = LoRACompatibleLinear + + self.linear_cls = linear_cls + self.to_q = linear_cls(query_dim, self.inner_dim, bias=bias) + + if not self.only_cross_attention: + # only relevant for the `AddedKVProcessor` classes + self.to_k = linear_cls(self.cross_attention_dim, self.inner_dim, bias=bias) + self.to_v = linear_cls(self.cross_attention_dim, self.inner_dim, bias=bias) + else: + self.to_k = None + self.to_v = None + + if self.added_kv_proj_dim is not None: + self.add_k_proj = linear_cls(added_kv_proj_dim, self.inner_dim) + self.add_v_proj = linear_cls(added_kv_proj_dim, self.inner_dim) + + self.to_out = nn.ModuleList([]) + self.to_out.append(linear_cls(self.inner_dim, self.out_dim, bias=out_bias)) + self.to_out.append(nn.Dropout(dropout)) + + # set attention processor + # We use the AttnProcessor2_0 by default when torch 2.x is used which uses + # torch.nn.functional.scaled_dot_product_attention for native Flash/memory_efficient_attention + # but only if it has the default `scale` argument. TODO remove scale_qk check when we move to torch 2.1 + if processor is None: + processor = ( + AttnProcessor2_0() if hasattr(F, "scaled_dot_product_attention") and self.scale_qk else AttnProcessor() + ) + self.set_processor(processor) + + def set_use_memory_efficient_attention_xformers( + self, use_memory_efficient_attention_xformers: bool, attention_op: Optional[Callable] = None + ) -> None: + r""" + Set whether to use memory efficient attention from `xformers` or not. + + Args: + use_memory_efficient_attention_xformers (`bool`): + Whether to use memory efficient attention from `xformers` or not. + attention_op (`Callable`, *optional*): + The attention operation to use. Defaults to `None` which uses the default attention operation from + `xformers`. + """ + is_lora = hasattr(self, "processor") and isinstance( + self.processor, + LORA_ATTENTION_PROCESSORS, + ) + is_custom_diffusion = hasattr(self, "processor") and isinstance( + self.processor, + (CustomDiffusionAttnProcessor, CustomDiffusionXFormersAttnProcessor, CustomDiffusionAttnProcessor2_0), + ) + is_added_kv_processor = hasattr(self, "processor") and isinstance( + self.processor, + ( + AttnAddedKVProcessor, + AttnAddedKVProcessor2_0, + SlicedAttnAddedKVProcessor, + XFormersAttnAddedKVProcessor, + LoRAAttnAddedKVProcessor, + ), + ) + + if use_memory_efficient_attention_xformers: + if is_added_kv_processor and (is_lora or is_custom_diffusion): + raise NotImplementedError( + f"Memory efficient attention is currently not supported for LoRA or custom diffusion for attention processor type {self.processor}" + ) + if not is_xformers_available(): + raise ModuleNotFoundError( + ( + "Refer to https://github.com/facebookresearch/xformers for more information on how to install" + " xformers" + ), + name="xformers", + ) + elif not torch.cuda.is_available(): + raise ValueError( + "torch.cuda.is_available() should be True but is False. xformers' memory efficient attention is" + " only available for GPU " + ) + else: + try: + # Make sure we can run the memory efficient attention + _ = xformers.ops.memory_efficient_attention( + torch.randn((1, 2, 40), device="cuda"), + torch.randn((1, 2, 40), device="cuda"), + torch.randn((1, 2, 40), device="cuda"), + ) + except Exception as e: + raise e + + if is_lora: + # TODO (sayakpaul): should we throw a warning if someone wants to use the xformers + # variant when using PT 2.0 now that we have LoRAAttnProcessor2_0? + processor = LoRAXFormersAttnProcessor( + hidden_size=self.processor.hidden_size, + cross_attention_dim=self.processor.cross_attention_dim, + rank=self.processor.rank, + attention_op=attention_op, + ) + processor.load_state_dict(self.processor.state_dict()) + processor.to(self.processor.to_q_lora.up.weight.device) + elif is_custom_diffusion: + processor = CustomDiffusionXFormersAttnProcessor( + train_kv=self.processor.train_kv, + train_q_out=self.processor.train_q_out, + hidden_size=self.processor.hidden_size, + cross_attention_dim=self.processor.cross_attention_dim, + attention_op=attention_op, + ) + processor.load_state_dict(self.processor.state_dict()) + if hasattr(self.processor, "to_k_custom_diffusion"): + processor.to(self.processor.to_k_custom_diffusion.weight.device) + elif is_added_kv_processor: + # TODO(Patrick, Suraj, William) - currently xformers doesn't work for UnCLIP + # which uses this type of cross attention ONLY because the attention mask of format + # [0, ..., -10.000, ..., 0, ...,] is not supported + # throw warning + logger.info( + "Memory efficient attention with `xformers` might currently not work correctly if an attention mask is required for the attention operation." + ) + processor = XFormersAttnAddedKVProcessor(attention_op=attention_op) + else: + processor = XFormersAttnProcessor(attention_op=attention_op) + else: + if is_lora: + attn_processor_class = ( + LoRAAttnProcessor2_0 if hasattr(F, "scaled_dot_product_attention") else LoRAAttnProcessor + ) + processor = attn_processor_class( + hidden_size=self.processor.hidden_size, + cross_attention_dim=self.processor.cross_attention_dim, + rank=self.processor.rank, + ) + processor.load_state_dict(self.processor.state_dict()) + processor.to(self.processor.to_q_lora.up.weight.device) + elif is_custom_diffusion: + attn_processor_class = ( + CustomDiffusionAttnProcessor2_0 + if hasattr(F, "scaled_dot_product_attention") + else CustomDiffusionAttnProcessor + ) + processor = attn_processor_class( + train_kv=self.processor.train_kv, + train_q_out=self.processor.train_q_out, + hidden_size=self.processor.hidden_size, + cross_attention_dim=self.processor.cross_attention_dim, + ) + processor.load_state_dict(self.processor.state_dict()) + if hasattr(self.processor, "to_k_custom_diffusion"): + processor.to(self.processor.to_k_custom_diffusion.weight.device) + else: + # set attention processor + # We use the AttnProcessor2_0 by default when torch 2.x is used which uses + # torch.nn.functional.scaled_dot_product_attention for native Flash/memory_efficient_attention + # but only if it has the default `scale` argument. TODO remove scale_qk check when we move to torch 2.1 + processor = ( + AttnProcessor2_0() + if hasattr(F, "scaled_dot_product_attention") and self.scale_qk + else AttnProcessor() + ) + + self.set_processor(processor) + + def set_attention_slice(self, slice_size: int) -> None: + r""" + Set the slice size for attention computation. + + Args: + slice_size (`int`): + The slice size for attention computation. + """ + if slice_size is not None and slice_size > self.sliceable_head_dim: + raise ValueError(f"slice_size {slice_size} has to be smaller or equal to {self.sliceable_head_dim}.") + + if slice_size is not None and self.added_kv_proj_dim is not None: + processor = SlicedAttnAddedKVProcessor(slice_size) + elif slice_size is not None: + processor = SlicedAttnProcessor(slice_size) + elif self.added_kv_proj_dim is not None: + processor = AttnAddedKVProcessor() + else: + # set attention processor + # We use the AttnProcessor2_0 by default when torch 2.x is used which uses + # torch.nn.functional.scaled_dot_product_attention for native Flash/memory_efficient_attention + # but only if it has the default `scale` argument. TODO remove scale_qk check when we move to torch 2.1 + processor = ( + AttnProcessor2_0() if hasattr(F, "scaled_dot_product_attention") and self.scale_qk else AttnProcessor() + ) + + self.set_processor(processor) + + def set_processor(self, processor: "AttnProcessor") -> None: + r""" + Set the attention processor to use. + + Args: + processor (`AttnProcessor`): + The attention processor to use. + """ + # if current processor is in `self._modules` and if passed `processor` is not, we need to + # pop `processor` from `self._modules` + if ( + hasattr(self, "processor") + and isinstance(self.processor, torch.nn.Module) + and not isinstance(processor, torch.nn.Module) + ): + logger.info(f"You are removing possibly trained weights of {self.processor} with {processor}") + self._modules.pop("processor") + + self.processor = processor + + def get_processor(self, return_deprecated_lora: bool = False) -> "AttentionProcessor": + r""" + Get the attention processor in use. + + Args: + return_deprecated_lora (`bool`, *optional*, defaults to `False`): + Set to `True` to return the deprecated LoRA attention processor. + + Returns: + "AttentionProcessor": The attention processor in use. + """ + if not return_deprecated_lora: + return self.processor + + # TODO(Sayak, Patrick). The rest of the function is needed to ensure backwards compatible + # serialization format for LoRA Attention Processors. It should be deleted once the integration + # with PEFT is completed. + is_lora_activated = { + name: module.lora_layer is not None + for name, module in self.named_modules() + if hasattr(module, "lora_layer") + } + + # 1. if no layer has a LoRA activated we can return the processor as usual + if not any(is_lora_activated.values()): + return self.processor + + # If doesn't apply LoRA do `add_k_proj` or `add_v_proj` + is_lora_activated.pop("add_k_proj", None) + is_lora_activated.pop("add_v_proj", None) + # 2. else it is not posssible that only some layers have LoRA activated + if not all(is_lora_activated.values()): + raise ValueError( + f"Make sure that either all layers or no layers have LoRA activated, but have {is_lora_activated}" + ) + + # 3. And we need to merge the current LoRA layers into the corresponding LoRA attention processor + non_lora_processor_cls_name = self.processor.__class__.__name__ + lora_processor_cls = getattr(import_module(__name__), "LoRA" + non_lora_processor_cls_name) + + hidden_size = self.inner_dim + + # now create a LoRA attention processor from the LoRA layers + if lora_processor_cls in [LoRAAttnProcessor, LoRAAttnProcessor2_0, LoRAXFormersAttnProcessor]: + kwargs = { + "cross_attention_dim": self.cross_attention_dim, + "rank": self.to_q.lora_layer.rank, + "network_alpha": self.to_q.lora_layer.network_alpha, + "q_rank": self.to_q.lora_layer.rank, + "q_hidden_size": self.to_q.lora_layer.out_features, + "k_rank": self.to_k.lora_layer.rank, + "k_hidden_size": self.to_k.lora_layer.out_features, + "v_rank": self.to_v.lora_layer.rank, + "v_hidden_size": self.to_v.lora_layer.out_features, + "out_rank": self.to_out[0].lora_layer.rank, + "out_hidden_size": self.to_out[0].lora_layer.out_features, + } + + if hasattr(self.processor, "attention_op"): + kwargs["attention_op"] = self.processor.attention_op + + lora_processor = lora_processor_cls(hidden_size, **kwargs) + lora_processor.to_q_lora.load_state_dict(self.to_q.lora_layer.state_dict()) + lora_processor.to_k_lora.load_state_dict(self.to_k.lora_layer.state_dict()) + lora_processor.to_v_lora.load_state_dict(self.to_v.lora_layer.state_dict()) + lora_processor.to_out_lora.load_state_dict(self.to_out[0].lora_layer.state_dict()) + elif lora_processor_cls == LoRAAttnAddedKVProcessor: + lora_processor = lora_processor_cls( + hidden_size, + cross_attention_dim=self.add_k_proj.weight.shape[0], + rank=self.to_q.lora_layer.rank, + network_alpha=self.to_q.lora_layer.network_alpha, + ) + lora_processor.to_q_lora.load_state_dict(self.to_q.lora_layer.state_dict()) + lora_processor.to_k_lora.load_state_dict(self.to_k.lora_layer.state_dict()) + lora_processor.to_v_lora.load_state_dict(self.to_v.lora_layer.state_dict()) + lora_processor.to_out_lora.load_state_dict(self.to_out[0].lora_layer.state_dict()) + + # only save if used + if self.add_k_proj.lora_layer is not None: + lora_processor.add_k_proj_lora.load_state_dict(self.add_k_proj.lora_layer.state_dict()) + lora_processor.add_v_proj_lora.load_state_dict(self.add_v_proj.lora_layer.state_dict()) + else: + lora_processor.add_k_proj_lora = None + lora_processor.add_v_proj_lora = None + else: + raise ValueError(f"{lora_processor_cls} does not exist.") + + return lora_processor + + def forward( + self, + hidden_states: torch.FloatTensor, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + **cross_attention_kwargs, + ) -> torch.Tensor: + r""" + The forward method of the `Attention` class. + + Args: + hidden_states (`torch.Tensor`): + The hidden states of the query. + encoder_hidden_states (`torch.Tensor`, *optional*): + The hidden states of the encoder. + attention_mask (`torch.Tensor`, *optional*): + The attention mask to use. If `None`, no mask is applied. + **cross_attention_kwargs: + Additional keyword arguments to pass along to the cross attention. + + Returns: + `torch.Tensor`: The output of the attention layer. + """ + # The `Attention` class can call different attention processors / attention functions + # here we simply pass along all tensors to the selected processor class + # For standard processors that are defined here, `**cross_attention_kwargs` is empty + return self.processor( + self, + hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=attention_mask, + **cross_attention_kwargs, + ) + + def batch_to_head_dim(self, tensor: torch.Tensor) -> torch.Tensor: + r""" + Reshape the tensor from `[batch_size, seq_len, dim]` to `[batch_size // heads, seq_len, dim * heads]`. `heads` + is the number of heads initialized while constructing the `Attention` class. + + Args: + tensor (`torch.Tensor`): The tensor to reshape. + + Returns: + `torch.Tensor`: The reshaped tensor. + """ + head_size = self.heads + batch_size, seq_len, dim = tensor.shape + tensor = tensor.reshape(batch_size // head_size, head_size, seq_len, dim) + tensor = tensor.permute(0, 2, 1, 3).reshape(batch_size // head_size, seq_len, dim * head_size) + return tensor + + def head_to_batch_dim(self, tensor: torch.Tensor, out_dim: int = 3) -> torch.Tensor: + r""" + Reshape the tensor from `[batch_size, seq_len, dim]` to `[batch_size, seq_len, heads, dim // heads]` `heads` is + the number of heads initialized while constructing the `Attention` class. + + Args: + tensor (`torch.Tensor`): The tensor to reshape. + out_dim (`int`, *optional*, defaults to `3`): The output dimension of the tensor. If `3`, the tensor is + reshaped to `[batch_size * heads, seq_len, dim // heads]`. + + Returns: + `torch.Tensor`: The reshaped tensor. + """ + head_size = self.heads + batch_size, seq_len, dim = tensor.shape + tensor = tensor.reshape(batch_size, seq_len, head_size, dim // head_size) + tensor = tensor.permute(0, 2, 1, 3) + + if out_dim == 3: + tensor = tensor.reshape(batch_size * head_size, seq_len, dim // head_size) + + return tensor + + def get_attention_scores( + self, query: torch.Tensor, key: torch.Tensor, attention_mask: torch.Tensor = None + ) -> torch.Tensor: + r""" + Compute the attention scores. + + Args: + query (`torch.Tensor`): The query tensor. + key (`torch.Tensor`): The key tensor. + attention_mask (`torch.Tensor`, *optional*): The attention mask to use. If `None`, no mask is applied. + + Returns: + `torch.Tensor`: The attention probabilities/scores. + """ + dtype = query.dtype + if self.upcast_attention: + query = query.float() + key = key.float() + + if attention_mask is None: + baddbmm_input = torch.empty( + query.shape[0], query.shape[1], key.shape[1], dtype=query.dtype, device=query.device + ) + beta = 0 + else: + baddbmm_input = attention_mask + beta = 1 + + attention_scores = torch.baddbmm( + baddbmm_input, + query, + key.transpose(-1, -2), + beta=beta, + alpha=self.scale, + ) + del baddbmm_input + + if self.upcast_softmax: + attention_scores = attention_scores.float() + + attention_probs = attention_scores.softmax(dim=-1) + del attention_scores + + attention_probs = attention_probs.to(dtype) + + return attention_probs + + def prepare_attention_mask( + self, attention_mask: torch.Tensor, target_length: int, batch_size: int, out_dim: int = 3 + ) -> torch.Tensor: + r""" + Prepare the attention mask for the attention computation. + + Args: + attention_mask (`torch.Tensor`): + The attention mask to prepare. + target_length (`int`): + The target length of the attention mask. This is the length of the attention mask after padding. + batch_size (`int`): + The batch size, which is used to repeat the attention mask. + out_dim (`int`, *optional*, defaults to `3`): + The output dimension of the attention mask. Can be either `3` or `4`. + + Returns: + `torch.Tensor`: The prepared attention mask. + """ + head_size = self.heads + if attention_mask is None: + return attention_mask + + current_length: int = attention_mask.shape[-1] + if current_length != target_length: + if attention_mask.device.type == "mps": + # HACK: MPS: Does not support padding by greater than dimension of input tensor. + # Instead, we can manually construct the padding tensor. + padding_shape = (attention_mask.shape[0], attention_mask.shape[1], target_length) + padding = torch.zeros(padding_shape, dtype=attention_mask.dtype, device=attention_mask.device) + attention_mask = torch.cat([attention_mask, padding], dim=2) + else: + # TODO: for pipelines such as stable-diffusion, padding cross-attn mask: + # we want to instead pad by (0, remaining_length), where remaining_length is: + # remaining_length: int = target_length - current_length + # TODO: re-enable tests/models/test_models_unet_2d_condition.py#test_model_xattn_padding + attention_mask = F.pad(attention_mask, (0, target_length), value=0.0) + + if out_dim == 3: + if attention_mask.shape[0] < batch_size * head_size: + attention_mask = attention_mask.repeat_interleave(head_size, dim=0) + elif out_dim == 4: + attention_mask = attention_mask.unsqueeze(1) + attention_mask = attention_mask.repeat_interleave(head_size, dim=1) + + return attention_mask + + def norm_encoder_hidden_states(self, encoder_hidden_states: torch.Tensor) -> torch.Tensor: + r""" + Normalize the encoder hidden states. Requires `self.norm_cross` to be specified when constructing the + `Attention` class. + + Args: + encoder_hidden_states (`torch.Tensor`): Hidden states of the encoder. + + Returns: + `torch.Tensor`: The normalized encoder hidden states. + """ + assert self.norm_cross is not None, "self.norm_cross must be defined to call self.norm_encoder_hidden_states" + + if isinstance(self.norm_cross, nn.LayerNorm): + encoder_hidden_states = self.norm_cross(encoder_hidden_states) + elif isinstance(self.norm_cross, nn.GroupNorm): + # Group norm norms along the channels dimension and expects + # input to be in the shape of (N, C, *). In this case, we want + # to norm along the hidden dimension, so we need to move + # (batch_size, sequence_length, hidden_size) -> + # (batch_size, hidden_size, sequence_length) + encoder_hidden_states = encoder_hidden_states.transpose(1, 2) + encoder_hidden_states = self.norm_cross(encoder_hidden_states) + encoder_hidden_states = encoder_hidden_states.transpose(1, 2) + else: + assert False + + return encoder_hidden_states + + @torch.no_grad() + def fuse_projections(self, fuse=True): + is_cross_attention = self.cross_attention_dim != self.query_dim + device = self.to_q.weight.data.device + dtype = self.to_q.weight.data.dtype + + if not is_cross_attention: + # fetch weight matrices. + concatenated_weights = torch.cat([self.to_q.weight.data, self.to_k.weight.data, self.to_v.weight.data]) + in_features = concatenated_weights.shape[1] + out_features = concatenated_weights.shape[0] + + # create a new single projection layer and copy over the weights. + self.to_qkv = self.linear_cls(in_features, out_features, bias=False, device=device, dtype=dtype) + self.to_qkv.weight.copy_(concatenated_weights) + + else: + concatenated_weights = torch.cat([self.to_k.weight.data, self.to_v.weight.data]) + in_features = concatenated_weights.shape[1] + out_features = concatenated_weights.shape[0] + + self.to_kv = self.linear_cls(in_features, out_features, bias=False, device=device, dtype=dtype) + self.to_kv.weight.copy_(concatenated_weights) + + self.fused_projections = fuse + + +class AttnProcessor: + r""" + Default processor for performing attention-related computations. + """ + + def __call__( + self, + attn: Attention, + hidden_states: torch.FloatTensor, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + temb: Optional[torch.FloatTensor] = None, + scale: float = 1.0, + ) -> torch.Tensor: + residual = hidden_states + + args = () if USE_PEFT_BACKEND else (scale,) + + if attn.spatial_norm is not None: + hidden_states = attn.spatial_norm(hidden_states, temb) + + input_ndim = hidden_states.ndim + + if input_ndim == 4: + batch_size, channel, height, width = hidden_states.shape + hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2) + + batch_size, sequence_length, _ = ( + hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape + ) + attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) + + if attn.group_norm is not None: + hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2) + + query = attn.to_q(hidden_states, *args) + + if encoder_hidden_states is None: + encoder_hidden_states = hidden_states + elif attn.norm_cross: + encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states) + + key = attn.to_k(encoder_hidden_states, *args) + value = attn.to_v(encoder_hidden_states, *args) + + query = attn.head_to_batch_dim(query) + key = attn.head_to_batch_dim(key) + value = attn.head_to_batch_dim(value) + + attention_probs = attn.get_attention_scores(query, key, attention_mask) + hidden_states = torch.bmm(attention_probs, value) + hidden_states = attn.batch_to_head_dim(hidden_states) + + # linear proj + hidden_states = attn.to_out[0](hidden_states, *args) + # dropout + hidden_states = attn.to_out[1](hidden_states) + + if input_ndim == 4: + hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width) + + if attn.residual_connection: + hidden_states = hidden_states + residual + + hidden_states = hidden_states / attn.rescale_output_factor + + return hidden_states + + +class AttnAddedKVProcessor: + r""" + Processor for performing attention-related computations with extra learnable key and value matrices for the text + encoder. + """ + + def __call__( + self, + attn: Attention, + hidden_states: torch.FloatTensor, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + scale: float = 1.0, + ) -> torch.Tensor: + residual = hidden_states + + args = () if USE_PEFT_BACKEND else (scale,) + + hidden_states = hidden_states.view(hidden_states.shape[0], hidden_states.shape[1], -1).transpose(1, 2) + batch_size, sequence_length, _ = hidden_states.shape + + attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) + + if encoder_hidden_states is None: + encoder_hidden_states = hidden_states + elif attn.norm_cross: + encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states) + + hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2) + + query = attn.to_q(hidden_states, *args) + query = attn.head_to_batch_dim(query) + + encoder_hidden_states_key_proj = attn.add_k_proj(encoder_hidden_states, *args) + encoder_hidden_states_value_proj = attn.add_v_proj(encoder_hidden_states, *args) + encoder_hidden_states_key_proj = attn.head_to_batch_dim(encoder_hidden_states_key_proj) + encoder_hidden_states_value_proj = attn.head_to_batch_dim(encoder_hidden_states_value_proj) + + if not attn.only_cross_attention: + key = attn.to_k(hidden_states, *args) + value = attn.to_v(hidden_states, *args) + key = attn.head_to_batch_dim(key) + value = attn.head_to_batch_dim(value) + key = torch.cat([encoder_hidden_states_key_proj, key], dim=1) + value = torch.cat([encoder_hidden_states_value_proj, value], dim=1) + else: + key = encoder_hidden_states_key_proj + value = encoder_hidden_states_value_proj + + attention_probs = attn.get_attention_scores(query, key, attention_mask) + hidden_states = torch.bmm(attention_probs, value) + hidden_states = attn.batch_to_head_dim(hidden_states) + + # linear proj + hidden_states = attn.to_out[0](hidden_states, *args) + # dropout + hidden_states = attn.to_out[1](hidden_states) + + hidden_states = hidden_states.transpose(-1, -2).reshape(residual.shape) + hidden_states = hidden_states + residual + + return hidden_states + + +class AttnAddedKVProcessor2_0: + r""" + Processor for performing scaled dot-product attention (enabled by default if you're using PyTorch 2.0), with extra + learnable key and value matrices for the text encoder. + """ + + def __init__(self): + if not hasattr(F, "scaled_dot_product_attention"): + raise ImportError( + "AttnAddedKVProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0." + ) + + def __call__( + self, + attn: Attention, + hidden_states: torch.FloatTensor, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + scale: float = 1.0, + ) -> torch.Tensor: + residual = hidden_states + + args = () if USE_PEFT_BACKEND else (scale,) + + hidden_states = hidden_states.view(hidden_states.shape[0], hidden_states.shape[1], -1).transpose(1, 2) + batch_size, sequence_length, _ = hidden_states.shape + + attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size, out_dim=4) + + if encoder_hidden_states is None: + encoder_hidden_states = hidden_states + elif attn.norm_cross: + encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states) + + hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2) + + query = attn.to_q(hidden_states, *args) + query = attn.head_to_batch_dim(query, out_dim=4) + + encoder_hidden_states_key_proj = attn.add_k_proj(encoder_hidden_states) + encoder_hidden_states_value_proj = attn.add_v_proj(encoder_hidden_states) + encoder_hidden_states_key_proj = attn.head_to_batch_dim(encoder_hidden_states_key_proj, out_dim=4) + encoder_hidden_states_value_proj = attn.head_to_batch_dim(encoder_hidden_states_value_proj, out_dim=4) + + if not attn.only_cross_attention: + key = attn.to_k(hidden_states, *args) + value = attn.to_v(hidden_states, *args) + key = attn.head_to_batch_dim(key, out_dim=4) + value = attn.head_to_batch_dim(value, out_dim=4) + key = torch.cat([encoder_hidden_states_key_proj, key], dim=2) + value = torch.cat([encoder_hidden_states_value_proj, value], dim=2) + else: + key = encoder_hidden_states_key_proj + value = encoder_hidden_states_value_proj + + # the output of sdp = (batch, num_heads, seq_len, head_dim) + # TODO: add support for attn.scale when we move to Torch 2.1 + hidden_states = F.scaled_dot_product_attention( + query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False + ) + hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, residual.shape[1]) + + # linear proj + hidden_states = attn.to_out[0](hidden_states, *args) + # dropout + hidden_states = attn.to_out[1](hidden_states) + + hidden_states = hidden_states.transpose(-1, -2).reshape(residual.shape) + hidden_states = hidden_states + residual + + return hidden_states + + +class AttnProcessor2_0: + r""" + Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0). + """ + + def __init__(self): + if not hasattr(F, "scaled_dot_product_attention"): + raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.") + + def __call__( + self, + attn: Attention, + hidden_states: torch.FloatTensor, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + temb: Optional[torch.FloatTensor] = None, + scale: float = 1.0, + ) -> torch.FloatTensor: + residual = hidden_states + if attn.spatial_norm is not None: + hidden_states = attn.spatial_norm(hidden_states, temb) + + input_ndim = hidden_states.ndim + + if input_ndim == 4: + batch_size, channel, height, width = hidden_states.shape + hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2) + + batch_size, sequence_length, _ = ( + hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape + ) + + if attention_mask is not None: + attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) + # scaled_dot_product_attention expects attention_mask shape to be + # (batch, heads, source_length, target_length) + attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1]) + + if attn.group_norm is not None: + hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2) + + args = () if USE_PEFT_BACKEND else (scale,) + query = attn.to_q(hidden_states, *args) + + if encoder_hidden_states is None: + encoder_hidden_states = hidden_states + elif attn.norm_cross: + encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states) + + key = attn.to_k(encoder_hidden_states, *args) + value = attn.to_v(encoder_hidden_states, *args) + + inner_dim = key.shape[-1] + head_dim = inner_dim // attn.heads + + query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) + + key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) + value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) + + hidden_states = torch_npu.npu_fusion_attention( + query, key, value, + atten_mask=attention_mask, + input_layout='BNSD', + scale=head_dim**-0.5, + head_num=attn.heads, + )[0] + # hidden_states = F.scaled_dot_product_attention( + # query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False + # ) + + hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim) + hidden_states = hidden_states.to(query.dtype) + + # linear proj + hidden_states = attn.to_out[0](hidden_states, *args) + # dropout + hidden_states = attn.to_out[1](hidden_states) + + if input_ndim == 4: + hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width) + + if attn.residual_connection: + hidden_states = hidden_states + residual + + hidden_states = hidden_states / attn.rescale_output_factor + + return hidden_states + + + +ADDED_KV_ATTENTION_PROCESSORS = ( + AttnAddedKVProcessor, + AttnAddedKVProcessor2_0, +) + +CROSS_ATTENTION_PROCESSORS = ( + AttnProcessor, + AttnProcessor2_0, +) + +AttentionProcessor = Union[ + AttnProcessor, + AttnProcessor2_0, + AttnAddedKVProcessor, + AttnAddedKVProcessor2_0, +] diff --git a/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/layers/embeddings.py b/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/layers/embeddings.py new file mode 100644 index 0000000000000000000000000000000000000000..9f46aa5f3b03048bfb3affeafb20894982775c5d --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/layers/embeddings.py @@ -0,0 +1,747 @@ +# Copyright 2023 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +import math +from typing import List, Optional, Tuple, Union + +import numpy as np +import torch +from torch import nn + +from diffusers.utils import USE_PEFT_BACKEND, deprecate +from diffusers.models.lora import LoRACompatibleLinear + +from .activations import get_activation +from diffusers.models.embeddings import PatchEmbed, ImagePositionalEmbeddings, PixArtAlphaCombinedTimestepSizeEmbeddings + +def get_timestep_embedding( + timesteps: torch.Tensor, + embedding_dim: int, + flip_sin_to_cos: bool = False, + downscale_freq_shift: float = 1, + scale: float = 1, + max_period: int = 10000, +): + """ + This matches the implementation in Denoising Diffusion Probabilistic Models: Create sinusoidal timestep embeddings. + + :param timesteps: a 1-D Tensor of N indices, one per batch element. + These may be fractional. + :param embedding_dim: the dimension of the output. :param max_period: controls the minimum frequency of the + embeddings. :return: an [N x dim] Tensor of positional embeddings. + """ + assert len(timesteps.shape) == 1, "Timesteps should be a 1d-array" + + half_dim = embedding_dim // 2 + exponent = -math.log(max_period) * torch.arange( + start=0, end=half_dim, dtype=torch.float32, device=timesteps.device + ) + exponent = exponent / (half_dim - downscale_freq_shift) + + emb = torch.exp(exponent) + emb = timesteps[:, None].float() * emb[None, :] + + # scale embeddings + emb = scale * emb + + # concat sine and cosine embeddings + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=-1) + + # flip sine and cosine embeddings + if flip_sin_to_cos: + emb = torch.cat([emb[:, half_dim:], emb[:, :half_dim]], dim=-1) + + # zero pad + if embedding_dim % 2 == 1: + emb = torch.nn.functional.pad(emb, (0, 1, 0, 0)) + return emb + + +def get_2d_sincos_pos_embed( + embed_dim, grid_size, cls_token=False, extra_tokens=0, interpolation_scale=1.0, base_size=16 +): + """ + grid_size: int of the grid height and width return: pos_embed: [grid_size*grid_size, embed_dim] or + [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token) + """ + if isinstance(grid_size, int): + grid_size = (grid_size, grid_size) + + grid_h = np.arange(grid_size[0], dtype=np.float32) / (grid_size[0] / base_size) / interpolation_scale + grid_w = np.arange(grid_size[1], dtype=np.float32) / (grid_size[1] / base_size) / interpolation_scale + grid = np.meshgrid(grid_w, grid_h) # here w goes first + grid = np.stack(grid, axis=0) + + grid = grid.reshape([2, 1, grid_size[1], grid_size[0]]) + pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid) + if cls_token and extra_tokens > 0: + pos_embed = np.concatenate([np.zeros([extra_tokens, embed_dim]), pos_embed], axis=0) + return pos_embed + + +def get_2d_sincos_pos_embed_from_grid(embed_dim, grid): + if embed_dim % 2 != 0: + raise ValueError("embed_dim must be divisible by 2") + + # use half of dimensions to encode grid_h + emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0]) # (H*W, D/2) + emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1]) # (H*W, D/2) + + emb = np.concatenate([emb_h, emb_w], axis=1) # (H*W, D) + return emb + + +def get_1d_sincos_pos_embed_from_grid(embed_dim, pos): + """ + embed_dim: output dimension for each position pos: a list of positions to be encoded: size (M,) out: (M, D) + """ + if embed_dim % 2 != 0: + raise ValueError("embed_dim must be divisible by 2") + + omega = np.arange(embed_dim // 2, dtype=np.float64) + omega /= embed_dim / 2.0 + omega = 1.0 / 10000**omega # (D/2,) + + pos = pos.reshape(-1) # (M,) + out = np.einsum("m,d->md", pos, omega) # (M, D/2), outer product + + emb_sin = np.sin(out) # (M, D/2) + emb_cos = np.cos(out) # (M, D/2) + + emb = np.concatenate([emb_sin, emb_cos], axis=1) # (M, D) + return emb + + +class TimestepEmbedding(nn.Module): + def __init__( + self, + in_channels: int, + time_embed_dim: int, + act_fn: str = "silu", + out_dim: int = None, + post_act_fn: Optional[str] = None, + cond_proj_dim=None, + sample_proj_bias=True, + ): + super().__init__() + linear_cls = nn.Linear if USE_PEFT_BACKEND else LoRACompatibleLinear + + self.linear_1 = linear_cls(in_channels, time_embed_dim, sample_proj_bias) + + if cond_proj_dim is not None: + self.cond_proj = nn.Linear(cond_proj_dim, in_channels, bias=False) + else: + self.cond_proj = None + + self.act = get_activation(act_fn) + + if out_dim is not None: + time_embed_dim_out = out_dim + else: + time_embed_dim_out = time_embed_dim + self.linear_2 = linear_cls(time_embed_dim, time_embed_dim_out, sample_proj_bias) + + if post_act_fn is None: + self.post_act = None + else: + self.post_act = get_activation(post_act_fn) + + def forward(self, sample, condition=None): + if condition is not None: + sample = sample + self.cond_proj(condition) + sample = self.linear_1(sample) + + if self.act is not None: + sample = self.act(sample) + + sample = self.linear_2(sample) + + if self.post_act is not None: + sample = self.post_act(sample) + return sample + + +class Timesteps(nn.Module): + def __init__(self, num_channels: int, flip_sin_to_cos: bool, downscale_freq_shift: float): + super().__init__() + self.num_channels = num_channels + self.flip_sin_to_cos = flip_sin_to_cos + self.downscale_freq_shift = downscale_freq_shift + + def forward(self, timesteps): + t_emb = get_timestep_embedding( + timesteps, + self.num_channels, + flip_sin_to_cos=self.flip_sin_to_cos, + downscale_freq_shift=self.downscale_freq_shift, + ) + return t_emb + + +class GaussianFourierProjection(nn.Module): + """Gaussian Fourier embeddings for noise levels.""" + + def __init__( + self, embedding_size: int = 256, scale: float = 1.0, set_W_to_weight=True, log=True, flip_sin_to_cos=False + ): + super().__init__() + self.weight = nn.Parameter(torch.randn(embedding_size) * scale, requires_grad=False) + self.log = log + self.flip_sin_to_cos = flip_sin_to_cos + + if set_W_to_weight: + # to delete later + self.W = nn.Parameter(torch.randn(embedding_size) * scale, requires_grad=False) + + self.weight = self.W + + def forward(self, x): + if self.log: + x = torch.log(x) + + x_proj = x[:, None] * self.weight[None, :] * 2 * np.pi + + if self.flip_sin_to_cos: + out = torch.cat([torch.cos(x_proj), torch.sin(x_proj)], dim=-1) + else: + out = torch.cat([torch.sin(x_proj), torch.cos(x_proj)], dim=-1) + return out + + +class SinusoidalPositionalEmbedding(nn.Module): + """Apply positional information to a sequence of embeddings. + + Takes in a sequence of embeddings with shape (batch_size, seq_length, embed_dim) and adds positional embeddings to + them + + Args: + embed_dim: (int): Dimension of the positional embedding. + max_seq_length: Maximum sequence length to apply positional embeddings + + """ + + def __init__(self, embed_dim: int, max_seq_length: int = 32): + super().__init__() + position = torch.arange(max_seq_length).unsqueeze(1) + div_term = torch.exp(torch.arange(0, embed_dim, 2) * (-math.log(10000.0) / embed_dim)) + pe = torch.zeros(1, max_seq_length, embed_dim) + pe[0, :, 0::2] = torch.sin(position * div_term) + pe[0, :, 1::2] = torch.cos(position * div_term) + self.register_buffer("pe", pe) + + def forward(self, x): + _, seq_length, _ = x.shape + x = x + self.pe[:, :seq_length] + return x + + +class LabelEmbedding(nn.Module): + """ + Embeds class labels into vector representations. Also handles label dropout for classifier-free guidance. + + Args: + num_classes (`int`): The number of classes. + hidden_size (`int`): The size of the vector embeddings. + dropout_prob (`float`): The probability of dropping a label. + """ + + def __init__(self, num_classes, hidden_size, dropout_prob): + super().__init__() + use_cfg_embedding = dropout_prob > 0 + self.embedding_table = nn.Embedding(num_classes + use_cfg_embedding, hidden_size) + self.num_classes = num_classes + self.dropout_prob = dropout_prob + + def token_drop(self, labels, force_drop_ids=None): + """ + Drops labels to enable classifier-free guidance. + """ + if force_drop_ids is None: + drop_ids = torch.rand(labels.shape[0], device=labels.device) < self.dropout_prob + else: + drop_ids = torch.tensor(force_drop_ids == 1) + labels = torch.where(drop_ids, self.num_classes, labels) + return labels + + def forward(self, labels: torch.LongTensor, force_drop_ids=None): + use_dropout = self.dropout_prob > 0 + if (self.training and use_dropout) or (force_drop_ids is not None): + labels = self.token_drop(labels, force_drop_ids) + embeddings = self.embedding_table(labels) + return embeddings + + +class TextImageProjection(nn.Module): + def __init__( + self, + text_embed_dim: int = 1024, + image_embed_dim: int = 768, + cross_attention_dim: int = 768, + num_image_text_embeds: int = 10, + ): + super().__init__() + + self.num_image_text_embeds = num_image_text_embeds + self.image_embeds = nn.Linear(image_embed_dim, self.num_image_text_embeds * cross_attention_dim) + self.text_proj = nn.Linear(text_embed_dim, cross_attention_dim) + + def forward(self, text_embeds: torch.FloatTensor, image_embeds: torch.FloatTensor): + batch_size = text_embeds.shape[0] + + # image + image_text_embeds = self.image_embeds(image_embeds) + image_text_embeds = image_text_embeds.reshape(batch_size, self.num_image_text_embeds, -1) + + # text + text_embeds = self.text_proj(text_embeds) + + return torch.cat([image_text_embeds, text_embeds], dim=1) + + +class ImageProjection(nn.Module): + def __init__( + self, + image_embed_dim: int = 768, + cross_attention_dim: int = 768, + num_image_text_embeds: int = 32, + ): + super().__init__() + + self.num_image_text_embeds = num_image_text_embeds + self.image_embeds = nn.Linear(image_embed_dim, self.num_image_text_embeds * cross_attention_dim) + self.norm = nn.LayerNorm(cross_attention_dim) + + def forward(self, image_embeds: torch.FloatTensor): + batch_size = image_embeds.shape[0] + + # image + image_embeds = self.image_embeds(image_embeds) + image_embeds = image_embeds.reshape(batch_size, self.num_image_text_embeds, -1) + image_embeds = self.norm(image_embeds) + return image_embeds + + +class IPAdapterFullImageProjection(nn.Module): + def __init__(self, image_embed_dim=1024, cross_attention_dim=1024): + super().__init__() + from .attention import FeedForward + + self.ff = FeedForward(image_embed_dim, cross_attention_dim, mult=1, activation_fn="gelu") + self.norm = nn.LayerNorm(cross_attention_dim) + + def forward(self, image_embeds: torch.FloatTensor): + return self.norm(self.ff(image_embeds)) + + +class CombinedTimestepLabelEmbeddings(nn.Module): + def __init__(self, num_classes, embedding_dim, class_dropout_prob=0.1): + super().__init__() + + self.time_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=1) + self.timestep_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=embedding_dim) + self.class_embedder = LabelEmbedding(num_classes, embedding_dim, class_dropout_prob) + + def forward(self, timestep, class_labels, hidden_dtype=None): + timesteps_proj = self.time_proj(timestep) + timesteps_emb = self.timestep_embedder(timesteps_proj.to(dtype=hidden_dtype)) # (N, D) + + class_labels = self.class_embedder(class_labels) # (N, D) + + conditioning = timesteps_emb + class_labels # (N, D) + + return conditioning + + +class TextTimeEmbedding(nn.Module): + def __init__(self, encoder_dim: int, time_embed_dim: int, num_heads: int = 64): + super().__init__() + self.norm1 = nn.LayerNorm(encoder_dim) + self.pool = AttentionPooling(num_heads, encoder_dim) + self.proj = nn.Linear(encoder_dim, time_embed_dim) + self.norm2 = nn.LayerNorm(time_embed_dim) + + def forward(self, hidden_states): + hidden_states = self.norm1(hidden_states) + hidden_states = self.pool(hidden_states) + hidden_states = self.proj(hidden_states) + hidden_states = self.norm2(hidden_states) + return hidden_states + + +class TextImageTimeEmbedding(nn.Module): + def __init__(self, text_embed_dim: int = 768, image_embed_dim: int = 768, time_embed_dim: int = 1536): + super().__init__() + self.text_proj = nn.Linear(text_embed_dim, time_embed_dim) + self.text_norm = nn.LayerNorm(time_embed_dim) + self.image_proj = nn.Linear(image_embed_dim, time_embed_dim) + + def forward(self, text_embeds: torch.FloatTensor, image_embeds: torch.FloatTensor): + # text + time_text_embeds = self.text_proj(text_embeds) + time_text_embeds = self.text_norm(time_text_embeds) + + # image + time_image_embeds = self.image_proj(image_embeds) + + return time_image_embeds + time_text_embeds + + +class ImageTimeEmbedding(nn.Module): + def __init__(self, image_embed_dim: int = 768, time_embed_dim: int = 1536): + super().__init__() + self.image_proj = nn.Linear(image_embed_dim, time_embed_dim) + self.image_norm = nn.LayerNorm(time_embed_dim) + + def forward(self, image_embeds: torch.FloatTensor): + # image + time_image_embeds = self.image_proj(image_embeds) + time_image_embeds = self.image_norm(time_image_embeds) + return time_image_embeds + + +class ImageHintTimeEmbedding(nn.Module): + def __init__(self, image_embed_dim: int = 768, time_embed_dim: int = 1536): + super().__init__() + self.image_proj = nn.Linear(image_embed_dim, time_embed_dim) + self.image_norm = nn.LayerNorm(time_embed_dim) + self.input_hint_block = nn.Sequential( + nn.Conv2d(3, 16, 3, padding=1), + nn.SiLU(), + nn.Conv2d(16, 16, 3, padding=1), + nn.SiLU(), + nn.Conv2d(16, 32, 3, padding=1, stride=2), + nn.SiLU(), + nn.Conv2d(32, 32, 3, padding=1), + nn.SiLU(), + nn.Conv2d(32, 96, 3, padding=1, stride=2), + nn.SiLU(), + nn.Conv2d(96, 96, 3, padding=1), + nn.SiLU(), + nn.Conv2d(96, 256, 3, padding=1, stride=2), + nn.SiLU(), + nn.Conv2d(256, 4, 3, padding=1), + ) + + def forward(self, image_embeds: torch.FloatTensor, hint: torch.FloatTensor): + # image + time_image_embeds = self.image_proj(image_embeds) + time_image_embeds = self.image_norm(time_image_embeds) + hint = self.input_hint_block(hint) + return time_image_embeds, hint + + +class AttentionPooling(nn.Module): + # Copied from https://github.com/deep-floyd/IF/blob/2f91391f27dd3c468bf174be5805b4cc92980c0b/deepfloyd_if/model/nn.py#L54 + + def __init__(self, num_heads, embed_dim, dtype=None): + super().__init__() + self.dtype = dtype + self.positional_embedding = nn.Parameter(torch.randn(1, embed_dim) / embed_dim**0.5) + self.k_proj = nn.Linear(embed_dim, embed_dim, dtype=self.dtype) + self.q_proj = nn.Linear(embed_dim, embed_dim, dtype=self.dtype) + self.v_proj = nn.Linear(embed_dim, embed_dim, dtype=self.dtype) + self.num_heads = num_heads + self.dim_per_head = embed_dim // self.num_heads + + def forward(self, x): + bs, length, width = x.size() + + def shape(x): + # (bs, length, width) --> (bs, length, n_heads, dim_per_head) + x = x.view(bs, -1, self.num_heads, self.dim_per_head) + # (bs, length, n_heads, dim_per_head) --> (bs, n_heads, length, dim_per_head) + x = x.transpose(1, 2) + # (bs, n_heads, length, dim_per_head) --> (bs*n_heads, length, dim_per_head) + x = x.reshape(bs * self.num_heads, -1, self.dim_per_head) + # (bs*n_heads, length, dim_per_head) --> (bs*n_heads, dim_per_head, length) + x = x.transpose(1, 2) + return x + + class_token = x.mean(dim=1, keepdim=True) + self.positional_embedding.to(x.dtype) + x = torch.cat([class_token, x], dim=1) # (bs, length+1, width) + + # (bs*n_heads, class_token_length, dim_per_head) + q = shape(self.q_proj(class_token)) + # (bs*n_heads, length+class_token_length, dim_per_head) + k = shape(self.k_proj(x)) + v = shape(self.v_proj(x)) + + # (bs*n_heads, class_token_length, length+class_token_length): + scale = 1 / math.sqrt(math.sqrt(self.dim_per_head)) + weight = torch.einsum("bct,bcs->bts", q * scale, k * scale) # More stable with f16 than dividing afterwards + weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype) + + # (bs*n_heads, dim_per_head, class_token_length) + a = torch.einsum("bts,bcs->bct", weight, v) + + # (bs, length+1, width) + a = a.reshape(bs, -1, 1).transpose(1, 2) + + return a[:, 0, :] # cls_token + + +def get_fourier_embeds_from_boundingbox(embed_dim, box): + """ + Args: + embed_dim: int + box: a 3-D tensor [B x N x 4] representing the bounding boxes for GLIGEN pipeline + Returns: + [B x N x embed_dim] tensor of positional embeddings + """ + + batch_size, num_boxes = box.shape[:2] + + emb = 100 ** (torch.arange(embed_dim) / embed_dim) + emb = emb[None, None, None].to(device=box.device, dtype=box.dtype) + emb = emb * box.unsqueeze(-1) + + emb = torch.stack((emb.sin(), emb.cos()), dim=-1) + emb = emb.permute(0, 1, 3, 4, 2).reshape(batch_size, num_boxes, embed_dim * 2 * 4) + + return emb + + +class GLIGENTextBoundingboxProjection(nn.Module): + def __init__(self, positive_len, out_dim, feature_type="text-only", fourier_freqs=8): + super().__init__() + self.positive_len = positive_len + self.out_dim = out_dim + + self.fourier_embedder_dim = fourier_freqs + self.position_dim = fourier_freqs * 2 * 4 # 2: sin/cos, 4: xyxy + + if isinstance(out_dim, tuple): + out_dim = out_dim[0] + + if feature_type == "text-only": + self.linears = nn.Sequential( + nn.Linear(self.positive_len + self.position_dim, 512), + nn.SiLU(), + nn.Linear(512, 512), + nn.SiLU(), + nn.Linear(512, out_dim), + ) + self.null_positive_feature = torch.nn.Parameter(torch.zeros([self.positive_len])) + + elif feature_type == "text-image": + self.linears_text = nn.Sequential( + nn.Linear(self.positive_len + self.position_dim, 512), + nn.SiLU(), + nn.Linear(512, 512), + nn.SiLU(), + nn.Linear(512, out_dim), + ) + self.linears_image = nn.Sequential( + nn.Linear(self.positive_len + self.position_dim, 512), + nn.SiLU(), + nn.Linear(512, 512), + nn.SiLU(), + nn.Linear(512, out_dim), + ) + self.null_text_feature = torch.nn.Parameter(torch.zeros([self.positive_len])) + self.null_image_feature = torch.nn.Parameter(torch.zeros([self.positive_len])) + + self.null_position_feature = torch.nn.Parameter(torch.zeros([self.position_dim])) + + def forward( + self, + boxes, + masks, + positive_embeddings=None, + phrases_masks=None, + image_masks=None, + phrases_embeddings=None, + image_embeddings=None, + ): + masks = masks.unsqueeze(-1) + + # embedding position (it may includes padding as placeholder) + xyxy_embedding = get_fourier_embeds_from_boundingbox(self.fourier_embedder_dim, boxes) # B*N*4 -> B*N*C + + # learnable null embedding + xyxy_null = self.null_position_feature.view(1, 1, -1) + + # replace padding with learnable null embedding + xyxy_embedding = xyxy_embedding * masks + (1 - masks) * xyxy_null + + # positionet with text only information + if positive_embeddings is not None: + # learnable null embedding + positive_null = self.null_positive_feature.view(1, 1, -1) + + # replace padding with learnable null embedding + positive_embeddings = positive_embeddings * masks + (1 - masks) * positive_null + + objs = self.linears(torch.cat([positive_embeddings, xyxy_embedding], dim=-1)) + + # positionet with text and image infomation + else: + phrases_masks = phrases_masks.unsqueeze(-1) + image_masks = image_masks.unsqueeze(-1) + + # learnable null embedding + text_null = self.null_text_feature.view(1, 1, -1) + image_null = self.null_image_feature.view(1, 1, -1) + + # replace padding with learnable null embedding + phrases_embeddings = phrases_embeddings * phrases_masks + (1 - phrases_masks) * text_null + image_embeddings = image_embeddings * image_masks + (1 - image_masks) * image_null + + objs_text = self.linears_text(torch.cat([phrases_embeddings, xyxy_embedding], dim=-1)) + objs_image = self.linears_image(torch.cat([image_embeddings, xyxy_embedding], dim=-1)) + objs = torch.cat([objs_text, objs_image], dim=1) + + return objs + + +class PixArtAlphaTextProjection(nn.Module): + """ + Projects caption embeddings. Also handles dropout for classifier-free guidance. + + Adapted from https://github.com/PixArt-alpha/PixArt-alpha/blob/master/diffusion/model/nets/PixArt_blocks.py + """ + + def __init__(self, in_features, hidden_size, num_tokens=120): + super().__init__() + self.linear_1 = nn.Linear(in_features=in_features, out_features=hidden_size, bias=True) + self.act_1 = nn.GELU(approximate="tanh") + self.linear_2 = nn.Linear(in_features=hidden_size, out_features=hidden_size, bias=True) + + def forward(self, caption): + hidden_states = self.linear_1(caption) + hidden_states = self.act_1(hidden_states) + hidden_states = self.linear_2(hidden_states) + return hidden_states + + +class IPAdapterPlusImageProjection(nn.Module): + """Resampler of IP-Adapter Plus. + + Args: + ---- + embed_dims (int): The feature dimension. Defaults to 768. + output_dims (int): The number of output channels, that is the same + number of the channels in the + `unet.config.cross_attention_dim`. Defaults to 1024. + hidden_dims (int): The number of hidden channels. Defaults to 1280. + depth (int): The number of blocks. Defaults to 8. + dim_head (int): The number of head channels. Defaults to 64. + heads (int): Parallel attention heads. Defaults to 16. + num_queries (int): The number of queries. Defaults to 8. + ffn_ratio (float): The expansion ratio of feedforward network hidden + layer channels. Defaults to 4. + """ + + def __init__( + self, + embed_dims: int = 768, + output_dims: int = 1024, + hidden_dims: int = 1280, + depth: int = 4, + dim_head: int = 64, + heads: int = 16, + num_queries: int = 8, + ffn_ratio: float = 4, + ) -> None: + super().__init__() + from .attention import FeedForward # Lazy import to avoid circular import + + self.latents = nn.Parameter(torch.randn(1, num_queries, hidden_dims) / hidden_dims**0.5) + + self.proj_in = nn.Linear(embed_dims, hidden_dims) + + self.proj_out = nn.Linear(hidden_dims, output_dims) + self.norm_out = nn.LayerNorm(output_dims) + + self.layers = nn.ModuleList([]) + for _ in range(depth): + self.layers.append( + nn.ModuleList( + [ + nn.LayerNorm(hidden_dims), + nn.LayerNorm(hidden_dims), + Attention( + query_dim=hidden_dims, + dim_head=dim_head, + heads=heads, + out_bias=False, + ), + nn.Sequential( + nn.LayerNorm(hidden_dims), + FeedForward(hidden_dims, hidden_dims, activation_fn="gelu", mult=ffn_ratio, bias=False), + ), + ] + ) + ) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + """Forward pass. + + Args: + ---- + x (torch.Tensor): Input Tensor. + + Returns: + ------- + torch.Tensor: Output Tensor. + """ + latents = self.latents.repeat(x.size(0), 1, 1) + + x = self.proj_in(x) + + for ln0, ln1, attn, ff in self.layers: + residual = latents + + encoder_hidden_states = ln0(x) + latents = ln1(latents) + encoder_hidden_states = torch.cat([encoder_hidden_states, latents], dim=-2) + latents = attn(latents, encoder_hidden_states) + residual + latents = ff(latents) + latents + + latents = self.proj_out(latents) + return self.norm_out(latents) + + +class MultiIPAdapterImageProjection(nn.Module): + def __init__(self, IPAdapterImageProjectionLayers: Union[List[nn.Module], Tuple[nn.Module]]): + super().__init__() + self.image_projection_layers = nn.ModuleList(IPAdapterImageProjectionLayers) + + def forward(self, image_embeds: List[torch.FloatTensor]): + projected_image_embeds = [] + + # currently, we accept `image_embeds` as + # 1. a tensor (deprecated) with shape [batch_size, embed_dim] or [batch_size, sequence_length, embed_dim] + # 2. list of `n` tensors where `n` is number of ip-adapters, each tensor can hae shape [batch_size, num_images, embed_dim] or [batch_size, num_images, sequence_length, embed_dim] + if not isinstance(image_embeds, list): + deprecation_message = ( + "You have passed a tensor as `image_embeds`.This is deprecated and will be removed in a future release." + " Please make sure to update your script to pass `image_embeds` as a list of tensors to supress this warning." + ) + deprecate("image_embeds not a list", "1.0.0", deprecation_message, standard_warn=False) + image_embeds = [image_embeds.unsqueeze(1)] + + if len(image_embeds) != len(self.image_projection_layers): + raise ValueError( + f"image_embeds must have the same length as image_projection_layers, got {len(image_embeds)} and {len(self.image_projection_layers)}" + ) + + for image_embed, image_projection_layer in zip(image_embeds, self.image_projection_layers): + batch_size, num_images = image_embed.shape[0], image_embed.shape[1] + image_embed = image_embed.reshape((batch_size * num_images,) + image_embed.shape[2:]) + image_embed = image_projection_layer(image_embed) + image_embed = image_embed.reshape((batch_size, num_images) + image_embed.shape[1:]) + + projected_image_embeds.append(image_embed) + + return projected_image_embeds diff --git a/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/layers/normalization.py b/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/layers/normalization.py new file mode 100644 index 0000000000000000000000000000000000000000..5acced9ea6a6c10c800e2560c6390e3134bc82a5 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/layers/normalization.py @@ -0,0 +1,254 @@ +# coding=utf-8 +# Copyright 2023 HuggingFace Inc. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import numbers +from typing import Dict, Optional, Tuple + +import torch +import torch.nn as nn +import torch.nn.functional as F + +from diffusers.utils import is_torch_version +from .activations import get_activation +from .embeddings import CombinedTimestepLabelEmbeddings, PixArtAlphaCombinedTimestepSizeEmbeddings + + +class AdaLayerNorm(nn.Module): + r""" + Norm layer modified to incorporate timestep embeddings. + + Parameters: + embedding_dim (`int`): The size of each embedding vector. + num_embeddings (`int`): The size of the embeddings dictionary. + """ + + def __init__(self, embedding_dim: int, num_embeddings: int): + super().__init__() + self.emb = nn.Embedding(num_embeddings, embedding_dim) + self.silu = nn.SiLU() + self.linear = nn.Linear(embedding_dim, embedding_dim * 2) + self.norm = nn.LayerNorm(embedding_dim, elementwise_affine=False) + + def forward(self, x: torch.Tensor, timestep: torch.Tensor) -> torch.Tensor: + emb = self.linear(self.silu(self.emb(timestep))) + scale, shift = torch.chunk(emb, 2) + x = self.norm(x) * (1 + scale) + shift + return x + + +class AdaLayerNormZero(nn.Module): + r""" + Norm layer adaptive layer norm zero (adaLN-Zero). + + Parameters: + embedding_dim (`int`): The size of each embedding vector. + num_embeddings (`int`): The size of the embeddings dictionary. + """ + + def __init__(self, embedding_dim: int, num_embeddings: int): + super().__init__() + + self.emb = CombinedTimestepLabelEmbeddings(num_embeddings, embedding_dim) + + self.silu = nn.SiLU() + self.linear = nn.Linear(embedding_dim, 6 * embedding_dim, bias=True) + self.norm = nn.LayerNorm(embedding_dim, elementwise_affine=False, eps=1e-6) + + def forward( + self, + x: torch.Tensor, + timestep: torch.Tensor, + class_labels: torch.LongTensor, + hidden_dtype: Optional[torch.dtype] = None, + ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]: + emb = self.linear(self.silu(self.emb(timestep, class_labels, hidden_dtype=hidden_dtype))) + shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = emb.chunk(6, dim=1) + x = self.norm(x) * (1 + scale_msa[:, None]) + shift_msa[:, None] + return x, gate_msa, shift_mlp, scale_mlp, gate_mlp + + +class AdaLayerNormSingle(nn.Module): + r""" + Norm layer adaptive layer norm single (adaLN-single). + + As proposed in PixArt-Alpha (see: https://arxiv.org/abs/2310.00426; Section 2.3). + + Parameters: + embedding_dim (`int`): The size of each embedding vector. + use_additional_conditions (`bool`): To use additional conditions for normalization or not. + """ + + def __init__(self, embedding_dim: int, use_additional_conditions: bool = False): + super().__init__() + + self.emb = PixArtAlphaCombinedTimestepSizeEmbeddings( + embedding_dim, size_emb_dim=embedding_dim // 3, use_additional_conditions=use_additional_conditions + ) + + self.silu = nn.SiLU() + self.linear = nn.Linear(embedding_dim, 6 * embedding_dim, bias=True) + + def forward( + self, + timestep: torch.Tensor, + added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None, + batch_size: Optional[int] = None, + hidden_dtype: Optional[torch.dtype] = None, + ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]: + # No modulation happening here. + embedded_timestep = self.emb(timestep, **added_cond_kwargs, batch_size=batch_size, hidden_dtype=hidden_dtype) + return self.linear(self.silu(embedded_timestep)), embedded_timestep + + +class AdaGroupNorm(nn.Module): + r""" + GroupNorm layer modified to incorporate timestep embeddings. + + Parameters: + embedding_dim (`int`): The size of each embedding vector. + num_embeddings (`int`): The size of the embeddings dictionary. + num_groups (`int`): The number of groups to separate the channels into. + act_fn (`str`, *optional*, defaults to `None`): The activation function to use. + eps (`float`, *optional*, defaults to `1e-5`): The epsilon value to use for numerical stability. + """ + + def __init__( + self, embedding_dim: int, out_dim: int, num_groups: int, act_fn: Optional[str] = None, eps: float = 1e-5 + ): + super().__init__() + self.num_groups = num_groups + self.eps = eps + + if act_fn is None: + self.act = None + else: + self.act = get_activation(act_fn) + + self.linear = nn.Linear(embedding_dim, out_dim * 2) + + def forward(self, x: torch.Tensor, emb: torch.Tensor) -> torch.Tensor: + if self.act: + emb = self.act(emb) + emb = self.linear(emb) + emb = emb[:, :, None, None] + scale, shift = emb.chunk(2, dim=1) + + x = F.group_norm(x, self.num_groups, eps=self.eps) + x = x * (1 + scale) + shift + return x + + +class AdaLayerNormContinuous(nn.Module): + def __init__( + self, + embedding_dim: int, + conditioning_embedding_dim: int, + # NOTE: It is a bit weird that the norm layer can be configured to have scale and shift parameters + # because the output is immediately scaled and shifted by the projected conditioning embeddings. + # Note that AdaLayerNorm does not let the norm layer have scale and shift parameters. + # However, this is how it was implemented in the original code, and it's rather likely you should + # set `elementwise_affine` to False. + elementwise_affine=True, + eps=1e-5, + bias=True, + norm_type="layer_norm", + ): + super().__init__() + self.silu = nn.SiLU() + self.linear = nn.Linear(conditioning_embedding_dim, embedding_dim * 2, bias=bias) + if norm_type == "layer_norm": + self.norm = LayerNorm(embedding_dim, eps, elementwise_affine, bias) + elif norm_type == "rms_norm": + self.norm = RMSNorm(embedding_dim, eps, elementwise_affine) + else: + raise ValueError(f"unknown norm_type {norm_type}") + + def forward(self, x: torch.Tensor, conditioning_embedding: torch.Tensor) -> torch.Tensor: + emb = self.linear(self.silu(conditioning_embedding)) + scale, shift = torch.chunk(emb, 2, dim=1) + x = self.norm(x) * (1 + scale)[:, None, :] + shift[:, None, :] + return x + + +if is_torch_version(">=", "2.1.0"): + LayerNorm = nn.LayerNorm +else: + # Has optional bias parameter compared to torch layer norm + # TODO: replace with torch layernorm once min required torch version >= 2.1 + class LayerNorm(nn.Module): + def __init__(self, dim, eps: float = 1e-5, elementwise_affine: bool = True, bias: bool = True): + super().__init__() + + self.eps = eps + + if isinstance(dim, numbers.Integral): + dim = (dim,) + + self.dim = torch.Size(dim) + + if elementwise_affine: + self.weight = nn.Parameter(torch.ones(dim)) + self.bias = nn.Parameter(torch.zeros(dim)) if bias else None + else: + self.weight = None + self.bias = None + + def forward(self, input): + return F.layer_norm(input, self.dim, self.weight, self.bias, self.eps) + + +class RMSNorm(nn.Module): + def __init__(self, dim, eps: float, elementwise_affine: bool = True): + super().__init__() + + self.eps = eps + + if isinstance(dim, numbers.Integral): + dim = (dim,) + + self.dim = torch.Size(dim) + + if elementwise_affine: + self.weight = nn.Parameter(torch.ones(dim)) + else: + self.weight = None + + def forward(self, hidden_states): + input_dtype = hidden_states.dtype + variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True) + hidden_states = hidden_states * torch.rsqrt(variance + self.eps) + + if self.weight is not None: + # convert into half-precision if necessary + if self.weight.dtype in [torch.float16, torch.bfloat16]: + hidden_states = hidden_states.to(self.weight.dtype) + hidden_states = hidden_states * self.weight + else: + hidden_states = hidden_states.to(input_dtype) + + return hidden_states + + +class GlobalResponseNorm(nn.Module): + # Taken from https://github.com/facebookresearch/ConvNeXt-V2/blob/3608f67cc1dae164790c5d0aead7bf2d73d9719b/models/utils.py#L105 + def __init__(self, dim): + super().__init__() + self.gamma = nn.Parameter(torch.zeros(1, 1, 1, dim)) + self.beta = nn.Parameter(torch.zeros(1, 1, 1, dim)) + + def forward(self, x): + gx = torch.norm(x, p=2, dim=(1, 2), keepdim=True) + nx = gx / (gx.mean(dim=-1, keepdim=True) + 1e-6) + return self.gamma * (x * nx) + self.beta + x diff --git a/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/models/__init__.py b/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/models/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..5af450275b73656a52deb912765cab1461ffd090 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/models/__init__.py @@ -0,0 +1 @@ +from .unets.unet_2d_condition import UNet2DConditionModel \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/models/model_utils.py b/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/models/model_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..0c9d557c2e7ca0446054d8ec36a71ee1f9de7d80 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/models/model_utils.py @@ -0,0 +1 @@ +from diffusers.models.modeling_utils import ModelMixin \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/models/transformers/__init__.py b/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/models/transformers/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..f5429a94ed6f940af278e7a7b7050fadf912f627 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/models/transformers/__init__.py @@ -0,0 +1,2 @@ +from .dual_transformer_2d import DualTransformer2DModel +from .transformer_2d import Transformer2DModel \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/models/transformers/dual_transformer_2d.py b/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/models/transformers/dual_transformer_2d.py new file mode 100644 index 0000000000000000000000000000000000000000..21b135c2eb86a6cc1ec2799c61c094e814893267 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/models/transformers/dual_transformer_2d.py @@ -0,0 +1,155 @@ +# Copyright 2023 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from typing import Optional + +from torch import nn + +from .transformer_2d import Transformer2DModel, Transformer2DModelOutput + + +class DualTransformer2DModel(nn.Module): + """ + Dual transformer wrapper that combines two `Transformer2DModel`s for mixed inference. + + Parameters: + num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention. + attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head. + in_channels (`int`, *optional*): + Pass if the input is continuous. The number of channels in the input and output. + num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use. + dropout (`float`, *optional*, defaults to 0.1): The dropout probability to use. + cross_attention_dim (`int`, *optional*): The number of encoder_hidden_states dimensions to use. + sample_size (`int`, *optional*): Pass if the input is discrete. The width of the latent images. + Note that this is fixed at training time as it is used for learning a number of position embeddings. See + `ImagePositionalEmbeddings`. + num_vector_embeds (`int`, *optional*): + Pass if the input is discrete. The number of classes of the vector embeddings of the latent pixels. + Includes the class for the masked latent pixel. + activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward. + num_embeds_ada_norm ( `int`, *optional*): Pass if at least one of the norm_layers is `AdaLayerNorm`. + The number of diffusion steps used during training. Note that this is fixed at training time as it is used + to learn a number of embeddings that are added to the hidden states. During inference, you can denoise for + up to but not more than steps than `num_embeds_ada_norm`. + attention_bias (`bool`, *optional*): + Configure if the TransformerBlocks' attention should contain a bias parameter. + """ + + def __init__( + self, + num_attention_heads: int = 16, + attention_head_dim: int = 88, + in_channels: Optional[int] = None, + num_layers: int = 1, + dropout: float = 0.0, + norm_num_groups: int = 32, + cross_attention_dim: Optional[int] = None, + attention_bias: bool = False, + sample_size: Optional[int] = None, + num_vector_embeds: Optional[int] = None, + activation_fn: str = "geglu", + num_embeds_ada_norm: Optional[int] = None, + ): + super().__init__() + self.transformers = nn.ModuleList( + [ + Transformer2DModel( + num_attention_heads=num_attention_heads, + attention_head_dim=attention_head_dim, + in_channels=in_channels, + num_layers=num_layers, + dropout=dropout, + norm_num_groups=norm_num_groups, + cross_attention_dim=cross_attention_dim, + attention_bias=attention_bias, + sample_size=sample_size, + num_vector_embeds=num_vector_embeds, + activation_fn=activation_fn, + num_embeds_ada_norm=num_embeds_ada_norm, + ) + for _ in range(2) + ] + ) + + # Variables that can be set by a pipeline: + + # The ratio of transformer1 to transformer2's output states to be combined during inference + self.mix_ratio = 0.5 + + # The shape of `encoder_hidden_states` is expected to be + # `(batch_size, condition_lengths[0]+condition_lengths[1], num_features)` + self.condition_lengths = [77, 257] + + # Which transformer to use to encode which condition. + # E.g. `(1, 0)` means that we'll use `transformers[1](conditions[0])` and `transformers[0](conditions[1])` + self.transformer_index_for_condition = [1, 0] + + def forward( + self, + hidden_states, + encoder_hidden_states, + timestep=None, + attention_mask=None, + cross_attention_kwargs=None, + return_dict: bool = True, + ): + """ + Args: + hidden_states ( When discrete, `torch.LongTensor` of shape `(batch size, num latent pixels)`. + When continuous, `torch.FloatTensor` of shape `(batch size, channel, height, width)`): Input + hidden_states. + encoder_hidden_states ( `torch.LongTensor` of shape `(batch size, encoder_hidden_states dim)`, *optional*): + Conditional embeddings for cross attention layer. If not given, cross-attention defaults to + self-attention. + timestep ( `torch.long`, *optional*): + Optional timestep to be applied as an embedding in AdaLayerNorm's. Used to indicate denoising step. + attention_mask (`torch.FloatTensor`, *optional*): + Optional attention mask to be applied in Attention. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`models.unets.unet_2d_condition.UNet2DConditionOutput`] instead of a plain tuple. + + Returns: + [`~models.transformer_2d.Transformer2DModelOutput`] or `tuple`: + [`~models.transformer_2d.Transformer2DModelOutput`] if `return_dict` is True, otherwise a `tuple`. When + returning a tuple, the first element is the sample tensor. + """ + input_states = hidden_states + + encoded_states = [] + tokens_start = 0 + # attention_mask is not used yet + for i in range(2): + # for each of the two transformers, pass the corresponding condition tokens + condition_state = encoder_hidden_states[:, tokens_start : tokens_start + self.condition_lengths[i]] + transformer_index = self.transformer_index_for_condition[i] + encoded_state = self.transformers[transformer_index]( + input_states, + encoder_hidden_states=condition_state, + timestep=timestep, + cross_attention_kwargs=cross_attention_kwargs, + return_dict=False, + )[0] + encoded_states.append(encoded_state - input_states) + tokens_start += self.condition_lengths[i] + + output_states = encoded_states[0] * self.mix_ratio + encoded_states[1] * (1 - self.mix_ratio) + output_states = output_states + input_states + + if not return_dict: + return (output_states,) + + return Transformer2DModelOutput(sample=output_states) diff --git a/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/models/transformers/transformer_2d.py b/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/models/transformers/transformer_2d.py new file mode 100644 index 0000000000000000000000000000000000000000..e33621f2497a737e92cd1e6e5d7386ea1f83b76c --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/models/transformers/transformer_2d.py @@ -0,0 +1,459 @@ +# Copyright 2023 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from dataclasses import dataclass +from typing import Any, Dict, Optional + +import torch +import torch.nn.functional as F +from torch import nn + +from diffusers.configuration_utils import ConfigMixin, register_to_config +from diffusers.utils import USE_PEFT_BACKEND, BaseOutput, deprecate, is_torch_version +from diffusers.models.lora import LoRACompatibleConv, LoRACompatibleLinear + +from stablediffusion.layers.attention import BasicTransformerBlock +from stablediffusion.layers.embeddings import ImagePositionalEmbeddings, PatchEmbed, PixArtAlphaTextProjection +from stablediffusion.layers.normalization import AdaLayerNormSingle +from stablediffusion.models.model_utils import ModelMixin + + +@dataclass +class Transformer2DModelOutput(BaseOutput): + """ + The output of [`Transformer2DModel`]. + + Args: + sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` or `(batch size, num_vector_embeds - 1, num_latent_pixels)` if [`Transformer2DModel`] is discrete): + The hidden states output conditioned on the `encoder_hidden_states` input. If discrete, returns probability + distributions for the unnoised latent pixels. + """ + + sample: torch.FloatTensor + + +class Transformer2DModel(ModelMixin, ConfigMixin): + """ + A 2D Transformer model for image-like data. + + Parameters: + num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention. + attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head. + in_channels (`int`, *optional*): + The number of channels in the input and output (specify if the input is **continuous**). + num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use. + dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. + cross_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use. + sample_size (`int`, *optional*): The width of the latent images (specify if the input is **discrete**). + This is fixed during training since it is used to learn a number of position embeddings. + num_vector_embeds (`int`, *optional*): + The number of classes of the vector embeddings of the latent pixels (specify if the input is **discrete**). + Includes the class for the masked latent pixel. + activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to use in feed-forward. + num_embeds_ada_norm ( `int`, *optional*): + The number of diffusion steps used during training. Pass if at least one of the norm_layers is + `AdaLayerNorm`. This is fixed during training since it is used to learn a number of embeddings that are + added to the hidden states. + + During inference, you can denoise for up to but not more steps than `num_embeds_ada_norm`. + attention_bias (`bool`, *optional*): + Configure if the `TransformerBlocks` attention should contain a bias parameter. + """ + + _supports_gradient_checkpointing = True + + @register_to_config + def __init__( + self, + num_attention_heads: int = 16, + attention_head_dim: int = 88, + in_channels: Optional[int] = None, + out_channels: Optional[int] = None, + num_layers: int = 1, + dropout: float = 0.0, + norm_num_groups: int = 32, + cross_attention_dim: Optional[int] = None, + attention_bias: bool = False, + sample_size: Optional[int] = None, + num_vector_embeds: Optional[int] = None, + patch_size: Optional[int] = None, + activation_fn: str = "geglu", + num_embeds_ada_norm: Optional[int] = None, + use_linear_projection: bool = False, + only_cross_attention: bool = False, + double_self_attention: bool = False, + upcast_attention: bool = False, + norm_type: str = "layer_norm", + norm_elementwise_affine: bool = True, + norm_eps: float = 1e-5, + attention_type: str = "default", + caption_channels: int = None, + ): + super().__init__() + self.use_linear_projection = use_linear_projection + self.num_attention_heads = num_attention_heads + self.attention_head_dim = attention_head_dim + inner_dim = num_attention_heads * attention_head_dim + + conv_cls = nn.Conv2d if USE_PEFT_BACKEND else LoRACompatibleConv + linear_cls = nn.Linear if USE_PEFT_BACKEND else LoRACompatibleLinear + + # 1. Transformer2DModel can process both standard continuous images of shape `(batch_size, num_channels, width, height)` as well as quantized image embeddings of shape `(batch_size, num_image_vectors)` + # Define whether input is continuous or discrete depending on configuration + self.is_input_continuous = (in_channels is not None) and (patch_size is None) + self.is_input_vectorized = num_vector_embeds is not None + self.is_input_patches = in_channels is not None and patch_size is not None + + if norm_type == "layer_norm" and num_embeds_ada_norm is not None: + deprecation_message = ( + f"The configuration file of this model: {self.__class__} is outdated. `norm_type` is either not set or" + " incorrectly set to `'layer_norm'`.Make sure to set `norm_type` to `'ada_norm'` in the config." + " Please make sure to update the config accordingly as leaving `norm_type` might led to incorrect" + " results in future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it" + " would be very nice if you could open a Pull request for the `transformer/config.json` file" + ) + deprecate("norm_type!=num_embeds_ada_norm", "1.0.0", deprecation_message, standard_warn=False) + norm_type = "ada_norm" + + if self.is_input_continuous and self.is_input_vectorized: + raise ValueError( + f"Cannot define both `in_channels`: {in_channels} and `num_vector_embeds`: {num_vector_embeds}. Make" + " sure that either `in_channels` or `num_vector_embeds` is None." + ) + elif self.is_input_vectorized and self.is_input_patches: + raise ValueError( + f"Cannot define both `num_vector_embeds`: {num_vector_embeds} and `patch_size`: {patch_size}. Make" + " sure that either `num_vector_embeds` or `num_patches` is None." + ) + elif not self.is_input_continuous and not self.is_input_vectorized and not self.is_input_patches: + raise ValueError( + f"Has to define `in_channels`: {in_channels}, `num_vector_embeds`: {num_vector_embeds}, or patch_size:" + f" {patch_size}. Make sure that `in_channels`, `num_vector_embeds` or `num_patches` is not None." + ) + + # 2. Define input layers + if self.is_input_continuous: + self.in_channels = in_channels + + self.norm = torch.nn.GroupNorm(num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True) + if use_linear_projection: + self.proj_in = linear_cls(in_channels, inner_dim) + else: + self.proj_in = conv_cls(in_channels, inner_dim, kernel_size=1, stride=1, padding=0) + elif self.is_input_vectorized: + assert sample_size is not None, "Transformer2DModel over discrete input must provide sample_size" + assert num_vector_embeds is not None, "Transformer2DModel over discrete input must provide num_embed" + + self.height = sample_size + self.width = sample_size + self.num_vector_embeds = num_vector_embeds + self.num_latent_pixels = self.height * self.width + + self.latent_image_embedding = ImagePositionalEmbeddings( + num_embed=num_vector_embeds, embed_dim=inner_dim, height=self.height, width=self.width + ) + elif self.is_input_patches: + assert sample_size is not None, "Transformer2DModel over patched input must provide sample_size" + + self.height = sample_size + self.width = sample_size + + self.patch_size = patch_size + interpolation_scale = self.config.sample_size // 64 # => 64 (= 512 pixart) has interpolation scale 1 + interpolation_scale = max(interpolation_scale, 1) + self.pos_embed = PatchEmbed( + height=sample_size, + width=sample_size, + patch_size=patch_size, + in_channels=in_channels, + embed_dim=inner_dim, + interpolation_scale=interpolation_scale, + ) + + # 3. Define transformers blocks + self.transformer_blocks = nn.ModuleList( + [ + BasicTransformerBlock( + inner_dim, + num_attention_heads, + attention_head_dim, + dropout=dropout, + cross_attention_dim=cross_attention_dim, + activation_fn=activation_fn, + num_embeds_ada_norm=num_embeds_ada_norm, + attention_bias=attention_bias, + only_cross_attention=only_cross_attention, + double_self_attention=double_self_attention, + upcast_attention=upcast_attention, + norm_type=norm_type, + norm_elementwise_affine=norm_elementwise_affine, + norm_eps=norm_eps, + attention_type=attention_type, + ) + for d in range(num_layers) + ] + ) + + # 4. Define output layers + self.out_channels = in_channels if out_channels is None else out_channels + if self.is_input_continuous: + # TODO: should use out_channels for continuous projections + if use_linear_projection: + self.proj_out = linear_cls(inner_dim, in_channels) + else: + self.proj_out = conv_cls(inner_dim, in_channels, kernel_size=1, stride=1, padding=0) + elif self.is_input_vectorized: + self.norm_out = nn.LayerNorm(inner_dim) + self.out = nn.Linear(inner_dim, self.num_vector_embeds - 1) + elif self.is_input_patches and norm_type != "ada_norm_single": + self.norm_out = nn.LayerNorm(inner_dim, elementwise_affine=False, eps=1e-6) + self.proj_out_1 = nn.Linear(inner_dim, 2 * inner_dim) + self.proj_out_2 = nn.Linear(inner_dim, patch_size * patch_size * self.out_channels) + elif self.is_input_patches and norm_type == "ada_norm_single": + self.norm_out = nn.LayerNorm(inner_dim, elementwise_affine=False, eps=1e-6) + self.scale_shift_table = nn.Parameter(torch.randn(2, inner_dim) / inner_dim**0.5) + self.proj_out = nn.Linear(inner_dim, patch_size * patch_size * self.out_channels) + + # 5. PixArt-Alpha blocks. + self.adaln_single = None + self.use_additional_conditions = False + if norm_type == "ada_norm_single": + self.use_additional_conditions = self.config.sample_size == 128 + # TODO(Sayak, PVP) clean this, for now we use sample size to determine whether to use + # additional conditions until we find better name + self.adaln_single = AdaLayerNormSingle(inner_dim, use_additional_conditions=self.use_additional_conditions) + + self.caption_projection = None + if caption_channels is not None: + self.caption_projection = PixArtAlphaTextProjection(in_features=caption_channels, hidden_size=inner_dim) + + self.gradient_checkpointing = False + + def _set_gradient_checkpointing(self, module, value=False): + if hasattr(module, "gradient_checkpointing"): + module.gradient_checkpointing = value + + def forward( + self, + hidden_states: torch.Tensor, + encoder_hidden_states: Optional[torch.Tensor] = None, + timestep: Optional[torch.LongTensor] = None, + added_cond_kwargs: Dict[str, torch.Tensor] = None, + class_labels: Optional[torch.LongTensor] = None, + cross_attention_kwargs: Dict[str, Any] = None, + attention_mask: Optional[torch.Tensor] = None, + encoder_attention_mask: Optional[torch.Tensor] = None, + return_dict: bool = True, + ): + """ + The [`Transformer2DModel`] forward method. + + Args: + hidden_states (`torch.LongTensor` of shape `(batch size, num latent pixels)` if discrete, `torch.FloatTensor` of shape `(batch size, channel, height, width)` if continuous): + Input `hidden_states`. + encoder_hidden_states ( `torch.FloatTensor` of shape `(batch size, sequence len, embed dims)`, *optional*): + Conditional embeddings for cross attention layer. If not given, cross-attention defaults to + self-attention. + timestep ( `torch.LongTensor`, *optional*): + Used to indicate denoising step. Optional timestep to be applied as an embedding in `AdaLayerNorm`. + class_labels ( `torch.LongTensor` of shape `(batch size, num classes)`, *optional*): + Used to indicate class labels conditioning. Optional class labels to be applied as an embedding in + `AdaLayerZeroNorm`. + cross_attention_kwargs ( `Dict[str, Any]`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + attention_mask ( `torch.Tensor`, *optional*): + An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask + is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large + negative values to the attention scores corresponding to "discard" tokens. + encoder_attention_mask ( `torch.Tensor`, *optional*): + Cross-attention mask applied to `encoder_hidden_states`. Two formats supported: + + * Mask `(batch, sequence_length)` True = keep, False = discard. + * Bias `(batch, 1, sequence_length)` 0 = keep, -10000 = discard. + + If `ndim == 2`: will be interpreted as a mask, then converted into a bias consistent with the format + above. This bias will be added to the cross-attention scores. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] instead of a plain + tuple. + + Returns: + If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a + `tuple` where the first element is the sample tensor. + """ + # ensure attention_mask is a bias, and give it a singleton query_tokens dimension. + # we may have done this conversion already, e.g. if we came here via UNet2DConditionModel#forward. + # we can tell by counting dims; if ndim == 2: it's a mask rather than a bias. + # expects mask of shape: + # [batch, key_tokens] + # adds singleton query_tokens dimension: + # [batch, 1, key_tokens] + # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes: + # [batch, heads, query_tokens, key_tokens] (e.g. torch sdp attn) + # [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn) + if attention_mask is not None and attention_mask.ndim == 2: + # assume that mask is expressed as: + # (1 = keep, 0 = discard) + # convert mask into a bias that can be added to attention scores: + # (keep = +0, discard = -10000.0) + attention_mask = (1 - attention_mask.to(hidden_states.dtype)) * -10000.0 + attention_mask = attention_mask.unsqueeze(1) + + # convert encoder_attention_mask to a bias the same way we do for attention_mask + if encoder_attention_mask is not None and encoder_attention_mask.ndim == 2: + encoder_attention_mask = (1 - encoder_attention_mask.to(hidden_states.dtype)) * -10000.0 + encoder_attention_mask = encoder_attention_mask.unsqueeze(1) + + # Retrieve lora scale. + lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0 + + # 1. Input + if self.is_input_continuous: + batch, _, height, width = hidden_states.shape + residual = hidden_states + + hidden_states = self.norm(hidden_states) + if not self.use_linear_projection: + hidden_states = ( + self.proj_in(hidden_states, scale=lora_scale) + if not USE_PEFT_BACKEND + else self.proj_in(hidden_states) + ) + inner_dim = hidden_states.shape[1] + hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * width, inner_dim) + else: + inner_dim = hidden_states.shape[1] + hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * width, inner_dim) + hidden_states = ( + self.proj_in(hidden_states, scale=lora_scale) + if not USE_PEFT_BACKEND + else self.proj_in(hidden_states) + ) + + elif self.is_input_vectorized: + hidden_states = self.latent_image_embedding(hidden_states) + elif self.is_input_patches: + height, width = hidden_states.shape[-2] // self.patch_size, hidden_states.shape[-1] // self.patch_size + hidden_states = self.pos_embed(hidden_states) + + if self.adaln_single is not None: + if self.use_additional_conditions and added_cond_kwargs is None: + raise ValueError( + "`added_cond_kwargs` cannot be None when using additional conditions for `adaln_single`." + ) + batch_size = hidden_states.shape[0] + timestep, embedded_timestep = self.adaln_single( + timestep, added_cond_kwargs, batch_size=batch_size, hidden_dtype=hidden_states.dtype + ) + + # 2. Blocks + if self.caption_projection is not None: + batch_size = hidden_states.shape[0] + encoder_hidden_states = self.caption_projection(encoder_hidden_states) + encoder_hidden_states = encoder_hidden_states.view(batch_size, -1, hidden_states.shape[-1]) + + for block in self.transformer_blocks: + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {} + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(block), + hidden_states, + attention_mask, + encoder_hidden_states, + encoder_attention_mask, + timestep, + cross_attention_kwargs, + class_labels, + **ckpt_kwargs, + ) + else: + hidden_states = block( + hidden_states, + attention_mask=attention_mask, + encoder_hidden_states=encoder_hidden_states, + encoder_attention_mask=encoder_attention_mask, + timestep=timestep, + cross_attention_kwargs=cross_attention_kwargs, + class_labels=class_labels, + ) + + # 3. Output + if self.is_input_continuous: + if not self.use_linear_projection: + hidden_states = hidden_states.reshape(batch, height, width, inner_dim).permute(0, 3, 1, 2).contiguous() + hidden_states = ( + self.proj_out(hidden_states, scale=lora_scale) + if not USE_PEFT_BACKEND + else self.proj_out(hidden_states) + ) + else: + hidden_states = ( + self.proj_out(hidden_states, scale=lora_scale) + if not USE_PEFT_BACKEND + else self.proj_out(hidden_states) + ) + hidden_states = hidden_states.reshape(batch, height, width, inner_dim).permute(0, 3, 1, 2).contiguous() + + output = hidden_states + residual + elif self.is_input_vectorized: + hidden_states = self.norm_out(hidden_states) + logits = self.out(hidden_states) + # (batch, self.num_vector_embeds - 1, self.num_latent_pixels) + logits = logits.permute(0, 2, 1) + + # log(p(x_0)) + output = F.log_softmax(logits.double(), dim=1).float() + + if self.is_input_patches: + if self.config.norm_type != "ada_norm_single": + conditioning = self.transformer_blocks[0].norm1.emb( + timestep, class_labels, hidden_dtype=hidden_states.dtype + ) + shift, scale = self.proj_out_1(F.silu(conditioning)).chunk(2, dim=1) + hidden_states = self.norm_out(hidden_states) * (1 + scale[:, None]) + shift[:, None] + hidden_states = self.proj_out_2(hidden_states) + elif self.config.norm_type == "ada_norm_single": + shift, scale = (self.scale_shift_table[None] + embedded_timestep[:, None]).chunk(2, dim=1) + hidden_states = self.norm_out(hidden_states) + # Modulation + hidden_states = hidden_states * (1 + scale) + shift + hidden_states = self.proj_out(hidden_states) + hidden_states = hidden_states.squeeze(1) + + # unpatchify + if self.adaln_single is None: + height = width = int(hidden_states.shape[1] ** 0.5) + hidden_states = hidden_states.reshape( + shape=(-1, height, width, self.patch_size, self.patch_size, self.out_channels) + ) + hidden_states = torch.einsum("nhwpqc->nchpwq", hidden_states) + output = hidden_states.reshape( + shape=(-1, self.out_channels, height * self.patch_size, width * self.patch_size) + ) + + if not return_dict: + return (output,) + + return Transformer2DModelOutput(sample=output) diff --git a/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/models/unets/__init__.py b/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/models/unets/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..1d3437072dac140d915d4e7922e1edbf31ed34de --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/models/unets/__init__.py @@ -0,0 +1 @@ +from .unet_2d_condition import UNet2DConditionModel \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/models/unets/unet_2d_blocks.py b/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/models/unets/unet_2d_blocks.py new file mode 100644 index 0000000000000000000000000000000000000000..b7d60b8bf0ce78a88f3f6cf29093e54812cc6cb4 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/models/unets/unet_2d_blocks.py @@ -0,0 +1,3592 @@ +# Copyright 2023 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from typing import Any, Dict, Optional, Tuple, Union + +import numpy as np +import torch +import torch.nn.functional as F +from torch import nn + +from diffusers.utils import is_torch_version, logging +from diffusers.utils.torch_utils import apply_freeu +from diffusers.models.resnet import ( + Downsample2D, + FirDownsample2D, + FirUpsample2D, + KDownsample2D, + KUpsample2D, + ResnetBlock2D, + ResnetBlockCondNorm2D, + Upsample2D, +) + +from stablediffusion.layers.activations import get_activation +from stablediffusion.layers.attention_processor import Attention, AttnAddedKVProcessor, AttnAddedKVProcessor2_0 +from stablediffusion.layers.normalization import AdaGroupNorm +from ..transformers.dual_transformer_2d import DualTransformer2DModel +from ..transformers.transformer_2d import Transformer2DModel + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +def get_down_block( + down_block_type: str, + num_layers: int, + in_channels: int, + out_channels: int, + temb_channels: int, + add_downsample: bool, + resnet_eps: float, + resnet_act_fn: str, + transformer_layers_per_block: int = 1, + num_attention_heads: Optional[int] = None, + resnet_groups: Optional[int] = None, + cross_attention_dim: Optional[int] = None, + downsample_padding: Optional[int] = None, + dual_cross_attention: bool = False, + use_linear_projection: bool = False, + only_cross_attention: bool = False, + upcast_attention: bool = False, + resnet_time_scale_shift: str = "default", + attention_type: str = "default", + resnet_skip_time_act: bool = False, + resnet_out_scale_factor: float = 1.0, + cross_attention_norm: Optional[str] = None, + attention_head_dim: Optional[int] = None, + downsample_type: Optional[str] = None, + dropout: float = 0.0, +): + # If attn head dim is not defined, we default it to the number of heads + if attention_head_dim is None: + logger.warn( + f"It is recommended to provide `attention_head_dim` when calling `get_down_block`. Defaulting `attention_head_dim` to {num_attention_heads}." + ) + attention_head_dim = num_attention_heads + + down_block_type = down_block_type[7:] if down_block_type.startswith("UNetRes") else down_block_type + if down_block_type == "DownBlock2D": + return DownBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + downsample_padding=downsample_padding, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + elif down_block_type == "ResnetDownsampleBlock2D": + return ResnetDownsampleBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + resnet_time_scale_shift=resnet_time_scale_shift, + skip_time_act=resnet_skip_time_act, + output_scale_factor=resnet_out_scale_factor, + ) + elif down_block_type == "AttnDownBlock2D": + if add_downsample is False: + downsample_type = None + else: + downsample_type = downsample_type or "conv" # default to 'conv' + return AttnDownBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + downsample_padding=downsample_padding, + attention_head_dim=attention_head_dim, + resnet_time_scale_shift=resnet_time_scale_shift, + downsample_type=downsample_type, + ) + elif down_block_type == "CrossAttnDownBlock2D": + if cross_attention_dim is None: + raise ValueError("cross_attention_dim must be specified for CrossAttnDownBlock2D") + return CrossAttnDownBlock2D( + num_layers=num_layers, + transformer_layers_per_block=transformer_layers_per_block, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + downsample_padding=downsample_padding, + cross_attention_dim=cross_attention_dim, + num_attention_heads=num_attention_heads, + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention, + upcast_attention=upcast_attention, + resnet_time_scale_shift=resnet_time_scale_shift, + attention_type=attention_type, + ) + elif down_block_type == "SimpleCrossAttnDownBlock2D": + if cross_attention_dim is None: + raise ValueError("cross_attention_dim must be specified for SimpleCrossAttnDownBlock2D") + return SimpleCrossAttnDownBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + cross_attention_dim=cross_attention_dim, + attention_head_dim=attention_head_dim, + resnet_time_scale_shift=resnet_time_scale_shift, + skip_time_act=resnet_skip_time_act, + output_scale_factor=resnet_out_scale_factor, + only_cross_attention=only_cross_attention, + cross_attention_norm=cross_attention_norm, + ) + elif down_block_type == "SkipDownBlock2D": + return SkipDownBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + downsample_padding=downsample_padding, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + elif down_block_type == "AttnSkipDownBlock2D": + return AttnSkipDownBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + attention_head_dim=attention_head_dim, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + elif down_block_type == "DownEncoderBlock2D": + return DownEncoderBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + downsample_padding=downsample_padding, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + elif down_block_type == "AttnDownEncoderBlock2D": + return AttnDownEncoderBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + downsample_padding=downsample_padding, + attention_head_dim=attention_head_dim, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + elif down_block_type == "KDownBlock2D": + return KDownBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + ) + elif down_block_type == "KCrossAttnDownBlock2D": + return KCrossAttnDownBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + cross_attention_dim=cross_attention_dim, + attention_head_dim=attention_head_dim, + add_self_attention=True if not add_downsample else False, + ) + raise ValueError(f"{down_block_type} does not exist.") + + +def get_up_block( + up_block_type: str, + num_layers: int, + in_channels: int, + out_channels: int, + prev_output_channel: int, + temb_channels: int, + add_upsample: bool, + resnet_eps: float, + resnet_act_fn: str, + resolution_idx: Optional[int] = None, + transformer_layers_per_block: int = 1, + num_attention_heads: Optional[int] = None, + resnet_groups: Optional[int] = None, + cross_attention_dim: Optional[int] = None, + dual_cross_attention: bool = False, + use_linear_projection: bool = False, + only_cross_attention: bool = False, + upcast_attention: bool = False, + resnet_time_scale_shift: str = "default", + attention_type: str = "default", + resnet_skip_time_act: bool = False, + resnet_out_scale_factor: float = 1.0, + cross_attention_norm: Optional[str] = None, + attention_head_dim: Optional[int] = None, + upsample_type: Optional[str] = None, + dropout: float = 0.0, +) -> nn.Module: + # If attn head dim is not defined, we default it to the number of heads + if attention_head_dim is None: + logger.warn( + f"It is recommended to provide `attention_head_dim` when calling `get_up_block`. Defaulting `attention_head_dim` to {num_attention_heads}." + ) + attention_head_dim = num_attention_heads + + up_block_type = up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type + if up_block_type == "UpBlock2D": + return UpBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + prev_output_channel=prev_output_channel, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + elif up_block_type == "ResnetUpsampleBlock2D": + return ResnetUpsampleBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + prev_output_channel=prev_output_channel, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + resnet_time_scale_shift=resnet_time_scale_shift, + skip_time_act=resnet_skip_time_act, + output_scale_factor=resnet_out_scale_factor, + ) + elif up_block_type == "CrossAttnUpBlock2D": + if cross_attention_dim is None: + raise ValueError("cross_attention_dim must be specified for CrossAttnUpBlock2D") + return CrossAttnUpBlock2D( + num_layers=num_layers, + transformer_layers_per_block=transformer_layers_per_block, + in_channels=in_channels, + out_channels=out_channels, + prev_output_channel=prev_output_channel, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + cross_attention_dim=cross_attention_dim, + num_attention_heads=num_attention_heads, + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention, + upcast_attention=upcast_attention, + resnet_time_scale_shift=resnet_time_scale_shift, + attention_type=attention_type, + ) + elif up_block_type == "SimpleCrossAttnUpBlock2D": + if cross_attention_dim is None: + raise ValueError("cross_attention_dim must be specified for SimpleCrossAttnUpBlock2D") + return SimpleCrossAttnUpBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + prev_output_channel=prev_output_channel, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + cross_attention_dim=cross_attention_dim, + attention_head_dim=attention_head_dim, + resnet_time_scale_shift=resnet_time_scale_shift, + skip_time_act=resnet_skip_time_act, + output_scale_factor=resnet_out_scale_factor, + only_cross_attention=only_cross_attention, + cross_attention_norm=cross_attention_norm, + ) + elif up_block_type == "AttnUpBlock2D": + if add_upsample is False: + upsample_type = None + else: + upsample_type = upsample_type or "conv" # default to 'conv' + + return AttnUpBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + prev_output_channel=prev_output_channel, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + attention_head_dim=attention_head_dim, + resnet_time_scale_shift=resnet_time_scale_shift, + upsample_type=upsample_type, + ) + elif up_block_type == "SkipUpBlock2D": + return SkipUpBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + prev_output_channel=prev_output_channel, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + elif up_block_type == "AttnSkipUpBlock2D": + return AttnSkipUpBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + prev_output_channel=prev_output_channel, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + attention_head_dim=attention_head_dim, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + elif up_block_type == "UpDecoderBlock2D": + return UpDecoderBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + resnet_time_scale_shift=resnet_time_scale_shift, + temb_channels=temb_channels, + ) + elif up_block_type == "AttnUpDecoderBlock2D": + return AttnUpDecoderBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + attention_head_dim=attention_head_dim, + resnet_time_scale_shift=resnet_time_scale_shift, + temb_channels=temb_channels, + ) + elif up_block_type == "KUpBlock2D": + return KUpBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + ) + elif up_block_type == "KCrossAttnUpBlock2D": + return KCrossAttnUpBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + cross_attention_dim=cross_attention_dim, + attention_head_dim=attention_head_dim, + ) + + raise ValueError(f"{up_block_type} does not exist.") + + +class AutoencoderTinyBlock(nn.Module): + """ + Tiny Autoencoder block used in [`AutoencoderTiny`]. It is a mini residual module consisting of plain conv + ReLU + blocks. + + Args: + in_channels (`int`): The number of input channels. + out_channels (`int`): The number of output channels. + act_fn (`str`): + ` The activation function to use. Supported values are `"swish"`, `"mish"`, `"gelu"`, and `"relu"`. + + Returns: + `torch.FloatTensor`: A tensor with the same shape as the input tensor, but with the number of channels equal to + `out_channels`. + """ + + def __init__(self, in_channels: int, out_channels: int, act_fn: str): + super().__init__() + act_fn = get_activation(act_fn) + self.conv = nn.Sequential( + nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1), + act_fn, + nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1), + act_fn, + nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1), + ) + self.skip = ( + nn.Conv2d(in_channels, out_channels, kernel_size=1, bias=False) + if in_channels != out_channels + else nn.Identity() + ) + self.fuse = nn.ReLU() + + def forward(self, x: torch.FloatTensor) -> torch.FloatTensor: + return self.fuse(self.conv(x) + self.skip(x)) + + +class UNetMidBlock2D(nn.Module): + """ + A 2D UNet mid-block [`UNetMidBlock2D`] with multiple residual blocks and optional attention blocks. + + Args: + in_channels (`int`): The number of input channels. + temb_channels (`int`): The number of temporal embedding channels. + dropout (`float`, *optional*, defaults to 0.0): The dropout rate. + num_layers (`int`, *optional*, defaults to 1): The number of residual blocks. + resnet_eps (`float`, *optional*, 1e-6 ): The epsilon value for the resnet blocks. + resnet_time_scale_shift (`str`, *optional*, defaults to `default`): + The type of normalization to apply to the time embeddings. This can help to improve the performance of the + model on tasks with long-range temporal dependencies. + resnet_act_fn (`str`, *optional*, defaults to `swish`): The activation function for the resnet blocks. + resnet_groups (`int`, *optional*, defaults to 32): + The number of groups to use in the group normalization layers of the resnet blocks. + attn_groups (`Optional[int]`, *optional*, defaults to None): The number of groups for the attention blocks. + resnet_pre_norm (`bool`, *optional*, defaults to `True`): + Whether to use pre-normalization for the resnet blocks. + add_attention (`bool`, *optional*, defaults to `True`): Whether to add attention blocks. + attention_head_dim (`int`, *optional*, defaults to 1): + Dimension of a single attention head. The number of attention heads is determined based on this value and + the number of input channels. + output_scale_factor (`float`, *optional*, defaults to 1.0): The output scale factor. + + Returns: + `torch.FloatTensor`: The output of the last residual block, which is a tensor of shape `(batch_size, + in_channels, height, width)`. + + """ + + def __init__( + self, + in_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", # default, spatial + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + attn_groups: Optional[int] = None, + resnet_pre_norm: bool = True, + add_attention: bool = True, + attention_head_dim: int = 1, + output_scale_factor: float = 1.0, + ): + super().__init__() + resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32) + self.add_attention = add_attention + + if attn_groups is None: + attn_groups = resnet_groups if resnet_time_scale_shift == "default" else None + + # there is always at least one resnet + if resnet_time_scale_shift == "spatial": + resnets = [ + ResnetBlockCondNorm2D( + in_channels=in_channels, + out_channels=in_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm="spatial", + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + ) + ] + else: + resnets = [ + ResnetBlock2D( + in_channels=in_channels, + out_channels=in_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ] + attentions = [] + + if attention_head_dim is None: + logger.warn( + f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {in_channels}." + ) + attention_head_dim = in_channels + + for _ in range(num_layers): + if self.add_attention: + attentions.append( + Attention( + in_channels, + heads=in_channels // attention_head_dim, + dim_head=attention_head_dim, + rescale_output_factor=output_scale_factor, + eps=resnet_eps, + norm_num_groups=attn_groups, + spatial_norm_dim=temb_channels if resnet_time_scale_shift == "spatial" else None, + residual_connection=True, + bias=True, + upcast_softmax=True, + _from_deprecated_attn_block=True, + ) + ) + else: + attentions.append(None) + + if resnet_time_scale_shift == "spatial": + resnets.append( + ResnetBlockCondNorm2D( + in_channels=in_channels, + out_channels=in_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm="spatial", + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + ) + ) + else: + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=in_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + def forward(self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None) -> torch.FloatTensor: + hidden_states = self.resnets[0](hidden_states, temb) + for attn, resnet in zip(self.attentions, self.resnets[1:]): + if attn is not None: + hidden_states = attn(hidden_states, temb=temb) + hidden_states = resnet(hidden_states, temb) + + return hidden_states + + +class UNetMidBlock2DCrossAttn(nn.Module): + def __init__( + self, + in_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + transformer_layers_per_block: Union[int, Tuple[int]] = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + num_attention_heads: int = 1, + output_scale_factor: float = 1.0, + cross_attention_dim: int = 1280, + dual_cross_attention: bool = False, + use_linear_projection: bool = False, + upcast_attention: bool = False, + attention_type: str = "default", + ): + super().__init__() + + self.has_cross_attention = True + self.num_attention_heads = num_attention_heads + resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32) + + # support for variable transformer layers per block + if isinstance(transformer_layers_per_block, int): + transformer_layers_per_block = [transformer_layers_per_block] * num_layers + + # there is always at least one resnet + resnets = [ + ResnetBlock2D( + in_channels=in_channels, + out_channels=in_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ] + attentions = [] + + for i in range(num_layers): + if not dual_cross_attention: + attentions.append( + Transformer2DModel( + num_attention_heads, + in_channels // num_attention_heads, + in_channels=in_channels, + num_layers=transformer_layers_per_block[i], + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + use_linear_projection=use_linear_projection, + upcast_attention=upcast_attention, + attention_type=attention_type, + ) + ) + else: + attentions.append( + DualTransformer2DModel( + num_attention_heads, + in_channels // num_attention_heads, + in_channels=in_channels, + num_layers=1, + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + ) + ) + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=in_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + self.gradient_checkpointing = False + + def forward( + self, + hidden_states: torch.FloatTensor, + temb: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + ) -> torch.FloatTensor: + lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0 + hidden_states = self.resnets[0](hidden_states, temb, scale=lora_scale) + for attn, resnet in zip(self.attentions, self.resnets[1:]): + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {} + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + return_dict=False, + )[0] + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), + hidden_states, + temb, + **ckpt_kwargs, + ) + else: + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + return_dict=False, + )[0] + hidden_states = resnet(hidden_states, temb, scale=lora_scale) + + return hidden_states + + +class UNetMidBlock2DSimpleCrossAttn(nn.Module): + def __init__( + self, + in_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + output_scale_factor: float = 1.0, + cross_attention_dim: int = 1280, + skip_time_act: bool = False, + only_cross_attention: bool = False, + cross_attention_norm: Optional[str] = None, + ): + super().__init__() + + self.has_cross_attention = True + + self.attention_head_dim = attention_head_dim + resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32) + + self.num_heads = in_channels // self.attention_head_dim + + # there is always at least one resnet + resnets = [ + ResnetBlock2D( + in_channels=in_channels, + out_channels=in_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + ) + ] + attentions = [] + + for _ in range(num_layers): + processor = ( + AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor() + ) + + attentions.append( + Attention( + query_dim=in_channels, + cross_attention_dim=in_channels, + heads=self.num_heads, + dim_head=self.attention_head_dim, + added_kv_proj_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + bias=True, + upcast_softmax=True, + only_cross_attention=only_cross_attention, + cross_attention_norm=cross_attention_norm, + processor=processor, + ) + ) + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=in_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + ) + ) + + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + def forward( + self, + hidden_states: torch.FloatTensor, + temb: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + ) -> torch.FloatTensor: + cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {} + lora_scale = cross_attention_kwargs.get("scale", 1.0) + + if attention_mask is None: + # if encoder_hidden_states is defined: we are doing cross-attn, so we should use cross-attn mask. + mask = None if encoder_hidden_states is None else encoder_attention_mask + else: + # when attention_mask is defined: we don't even check for encoder_attention_mask. + # this is to maintain compatibility with UnCLIP, which uses 'attention_mask' param for cross-attn masks. + # TODO: UnCLIP should express cross-attn mask via encoder_attention_mask param instead of via attention_mask. + # then we can simplify this whole if/else block to: + # mask = attention_mask if encoder_hidden_states is None else encoder_attention_mask + mask = attention_mask + + hidden_states = self.resnets[0](hidden_states, temb, scale=lora_scale) + for attn, resnet in zip(self.attentions, self.resnets[1:]): + # attn + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=mask, + **cross_attention_kwargs, + ) + + # resnet + hidden_states = resnet(hidden_states, temb, scale=lora_scale) + + return hidden_states + + +class AttnDownBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + output_scale_factor: float = 1.0, + downsample_padding: int = 1, + downsample_type: str = "conv", + ): + super().__init__() + resnets = [] + attentions = [] + self.downsample_type = downsample_type + + if attention_head_dim is None: + logger.warn( + f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {out_channels}." + ) + attention_head_dim = out_channels + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + attentions.append( + Attention( + out_channels, + heads=out_channels // attention_head_dim, + dim_head=attention_head_dim, + rescale_output_factor=output_scale_factor, + eps=resnet_eps, + norm_num_groups=resnet_groups, + residual_connection=True, + bias=True, + upcast_softmax=True, + _from_deprecated_attn_block=True, + ) + ) + + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if downsample_type == "conv": + self.downsamplers = nn.ModuleList( + [ + Downsample2D( + out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op" + ) + ] + ) + elif downsample_type == "resnet": + self.downsamplers = nn.ModuleList( + [ + ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + down=True, + ) + ] + ) + else: + self.downsamplers = None + + def forward( + self, + hidden_states: torch.FloatTensor, + temb: Optional[torch.FloatTensor] = None, + upsample_size: Optional[int] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]: + cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {} + + lora_scale = cross_attention_kwargs.get("scale", 1.0) + + output_states = () + + for resnet, attn in zip(self.resnets, self.attentions): + cross_attention_kwargs.update({"scale": lora_scale}) + hidden_states = resnet(hidden_states, temb, scale=lora_scale) + hidden_states = attn(hidden_states, **cross_attention_kwargs) + output_states = output_states + (hidden_states,) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + if self.downsample_type == "resnet": + hidden_states = downsampler(hidden_states, temb=temb, scale=lora_scale) + else: + hidden_states = downsampler(hidden_states, scale=lora_scale) + + output_states += (hidden_states,) + + return hidden_states, output_states + + +class CrossAttnDownBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + transformer_layers_per_block: Union[int, Tuple[int]] = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + num_attention_heads: int = 1, + cross_attention_dim: int = 1280, + output_scale_factor: float = 1.0, + downsample_padding: int = 1, + add_downsample: bool = True, + dual_cross_attention: bool = False, + use_linear_projection: bool = False, + only_cross_attention: bool = False, + upcast_attention: bool = False, + attention_type: str = "default", + ): + super().__init__() + resnets = [] + attentions = [] + + self.has_cross_attention = True + self.num_attention_heads = num_attention_heads + if isinstance(transformer_layers_per_block, int): + transformer_layers_per_block = [transformer_layers_per_block] * num_layers + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + if not dual_cross_attention: + attentions.append( + Transformer2DModel( + num_attention_heads, + out_channels // num_attention_heads, + in_channels=out_channels, + num_layers=transformer_layers_per_block[i], + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention, + upcast_attention=upcast_attention, + attention_type=attention_type, + ) + ) + else: + attentions.append( + DualTransformer2DModel( + num_attention_heads, + out_channels // num_attention_heads, + in_channels=out_channels, + num_layers=1, + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + ) + ) + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if add_downsample: + self.downsamplers = nn.ModuleList( + [ + Downsample2D( + out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op" + ) + ] + ) + else: + self.downsamplers = None + + self.gradient_checkpointing = False + + def forward( + self, + hidden_states: torch.FloatTensor, + temb: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + additional_residuals: Optional[torch.FloatTensor] = None, + ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]: + output_states = () + + lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0 + + blocks = list(zip(self.resnets, self.attentions)) + + for i, (resnet, attn) in enumerate(blocks): + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {} + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), + hidden_states, + temb, + **ckpt_kwargs, + ) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + return_dict=False, + )[0] + else: + hidden_states = resnet(hidden_states, temb, scale=lora_scale) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + return_dict=False, + )[0] + + # apply additional residuals to the output of the last pair of resnet and attention blocks + if i == len(blocks) - 1 and additional_residuals is not None: + hidden_states = hidden_states + additional_residuals + + output_states = output_states + (hidden_states,) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states, scale=lora_scale) + + output_states = output_states + (hidden_states,) + + return hidden_states, output_states + + +class DownBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + output_scale_factor: float = 1.0, + add_downsample: bool = True, + downsample_padding: int = 1, + ): + super().__init__() + resnets = [] + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + self.resnets = nn.ModuleList(resnets) + + if add_downsample: + self.downsamplers = nn.ModuleList( + [ + Downsample2D( + out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op" + ) + ] + ) + else: + self.downsamplers = None + + self.gradient_checkpointing = False + + def forward( + self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0 + ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]: + output_states = () + + for resnet in self.resnets: + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module): + def custom_forward(*inputs): + return module(*inputs) + + return custom_forward + + if is_torch_version(">=", "1.11.0"): + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb, use_reentrant=False + ) + else: + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb + ) + else: + hidden_states = resnet(hidden_states, temb, scale=scale) + + output_states = output_states + (hidden_states,) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states, scale=scale) + + output_states = output_states + (hidden_states,) + + return hidden_states, output_states + + +class DownEncoderBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + output_scale_factor: float = 1.0, + add_downsample: bool = True, + downsample_padding: int = 1, + ): + super().__init__() + resnets = [] + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + if resnet_time_scale_shift == "spatial": + resnets.append( + ResnetBlockCondNorm2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=None, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm="spatial", + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + ) + ) + else: + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=None, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + self.resnets = nn.ModuleList(resnets) + + if add_downsample: + self.downsamplers = nn.ModuleList( + [ + Downsample2D( + out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op" + ) + ] + ) + else: + self.downsamplers = None + + def forward(self, hidden_states: torch.FloatTensor, scale: float = 1.0) -> torch.FloatTensor: + for resnet in self.resnets: + hidden_states = resnet(hidden_states, temb=None, scale=scale) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states, scale) + + return hidden_states + + +class AttnDownEncoderBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + output_scale_factor: float = 1.0, + add_downsample: bool = True, + downsample_padding: int = 1, + ): + super().__init__() + resnets = [] + attentions = [] + + if attention_head_dim is None: + logger.warn( + f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {out_channels}." + ) + attention_head_dim = out_channels + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + if resnet_time_scale_shift == "spatial": + resnets.append( + ResnetBlockCondNorm2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=None, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm="spatial", + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + ) + ) + else: + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=None, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + attentions.append( + Attention( + out_channels, + heads=out_channels // attention_head_dim, + dim_head=attention_head_dim, + rescale_output_factor=output_scale_factor, + eps=resnet_eps, + norm_num_groups=resnet_groups, + residual_connection=True, + bias=True, + upcast_softmax=True, + _from_deprecated_attn_block=True, + ) + ) + + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if add_downsample: + self.downsamplers = nn.ModuleList( + [ + Downsample2D( + out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op" + ) + ] + ) + else: + self.downsamplers = None + + def forward(self, hidden_states: torch.FloatTensor, scale: float = 1.0) -> torch.FloatTensor: + for resnet, attn in zip(self.resnets, self.attentions): + hidden_states = resnet(hidden_states, temb=None, scale=scale) + cross_attention_kwargs = {"scale": scale} + hidden_states = attn(hidden_states, **cross_attention_kwargs) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states, scale) + + return hidden_states + + +class AttnSkipDownBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + output_scale_factor: float = np.sqrt(2.0), + add_downsample: bool = True, + ): + super().__init__() + self.attentions = nn.ModuleList([]) + self.resnets = nn.ModuleList([]) + + if attention_head_dim is None: + logger.warn( + f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {out_channels}." + ) + attention_head_dim = out_channels + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + self.resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=min(in_channels // 4, 32), + groups_out=min(out_channels // 4, 32), + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + self.attentions.append( + Attention( + out_channels, + heads=out_channels // attention_head_dim, + dim_head=attention_head_dim, + rescale_output_factor=output_scale_factor, + eps=resnet_eps, + norm_num_groups=32, + residual_connection=True, + bias=True, + upcast_softmax=True, + _from_deprecated_attn_block=True, + ) + ) + + if add_downsample: + self.resnet_down = ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=min(out_channels // 4, 32), + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + use_in_shortcut=True, + down=True, + kernel="fir", + ) + self.downsamplers = nn.ModuleList([FirDownsample2D(out_channels, out_channels=out_channels)]) + self.skip_conv = nn.Conv2d(3, out_channels, kernel_size=(1, 1), stride=(1, 1)) + else: + self.resnet_down = None + self.downsamplers = None + self.skip_conv = None + + def forward( + self, + hidden_states: torch.FloatTensor, + temb: Optional[torch.FloatTensor] = None, + skip_sample: Optional[torch.FloatTensor] = None, + scale: float = 1.0, + ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...], torch.FloatTensor]: + output_states = () + + for resnet, attn in zip(self.resnets, self.attentions): + hidden_states = resnet(hidden_states, temb, scale=scale) + cross_attention_kwargs = {"scale": scale} + hidden_states = attn(hidden_states, **cross_attention_kwargs) + output_states += (hidden_states,) + + if self.downsamplers is not None: + hidden_states = self.resnet_down(hidden_states, temb, scale=scale) + for downsampler in self.downsamplers: + skip_sample = downsampler(skip_sample) + + hidden_states = self.skip_conv(skip_sample) + hidden_states + + output_states += (hidden_states,) + + return hidden_states, output_states, skip_sample + + +class SkipDownBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_pre_norm: bool = True, + output_scale_factor: float = np.sqrt(2.0), + add_downsample: bool = True, + downsample_padding: int = 1, + ): + super().__init__() + self.resnets = nn.ModuleList([]) + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + self.resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=min(in_channels // 4, 32), + groups_out=min(out_channels // 4, 32), + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + if add_downsample: + self.resnet_down = ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=min(out_channels // 4, 32), + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + use_in_shortcut=True, + down=True, + kernel="fir", + ) + self.downsamplers = nn.ModuleList([FirDownsample2D(out_channels, out_channels=out_channels)]) + self.skip_conv = nn.Conv2d(3, out_channels, kernel_size=(1, 1), stride=(1, 1)) + else: + self.resnet_down = None + self.downsamplers = None + self.skip_conv = None + + def forward( + self, + hidden_states: torch.FloatTensor, + temb: Optional[torch.FloatTensor] = None, + skip_sample: Optional[torch.FloatTensor] = None, + scale: float = 1.0, + ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...], torch.FloatTensor]: + output_states = () + + for resnet in self.resnets: + hidden_states = resnet(hidden_states, temb, scale) + output_states += (hidden_states,) + + if self.downsamplers is not None: + hidden_states = self.resnet_down(hidden_states, temb, scale) + for downsampler in self.downsamplers: + skip_sample = downsampler(skip_sample) + + hidden_states = self.skip_conv(skip_sample) + hidden_states + + output_states += (hidden_states,) + + return hidden_states, output_states, skip_sample + + +class ResnetDownsampleBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + output_scale_factor: float = 1.0, + add_downsample: bool = True, + skip_time_act: bool = False, + ): + super().__init__() + resnets = [] + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + ) + ) + + self.resnets = nn.ModuleList(resnets) + + if add_downsample: + self.downsamplers = nn.ModuleList( + [ + ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + down=True, + ) + ] + ) + else: + self.downsamplers = None + + self.gradient_checkpointing = False + + def forward( + self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0 + ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]: + output_states = () + + for resnet in self.resnets: + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module): + def custom_forward(*inputs): + return module(*inputs) + + return custom_forward + + if is_torch_version(">=", "1.11.0"): + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb, use_reentrant=False + ) + else: + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb + ) + else: + hidden_states = resnet(hidden_states, temb, scale) + + output_states = output_states + (hidden_states,) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states, temb, scale) + + output_states = output_states + (hidden_states,) + + return hidden_states, output_states + + +class SimpleCrossAttnDownBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + cross_attention_dim: int = 1280, + output_scale_factor: float = 1.0, + add_downsample: bool = True, + skip_time_act: bool = False, + only_cross_attention: bool = False, + cross_attention_norm: Optional[str] = None, + ): + super().__init__() + + self.has_cross_attention = True + + resnets = [] + attentions = [] + + self.attention_head_dim = attention_head_dim + self.num_heads = out_channels // self.attention_head_dim + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + ) + ) + + processor = ( + AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor() + ) + + attentions.append( + Attention( + query_dim=out_channels, + cross_attention_dim=out_channels, + heads=self.num_heads, + dim_head=attention_head_dim, + added_kv_proj_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + bias=True, + upcast_softmax=True, + only_cross_attention=only_cross_attention, + cross_attention_norm=cross_attention_norm, + processor=processor, + ) + ) + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if add_downsample: + self.downsamplers = nn.ModuleList( + [ + ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + down=True, + ) + ] + ) + else: + self.downsamplers = None + + self.gradient_checkpointing = False + + def forward( + self, + hidden_states: torch.FloatTensor, + temb: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]: + output_states = () + cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {} + + lora_scale = cross_attention_kwargs.get("scale", 1.0) + + if attention_mask is None: + # if encoder_hidden_states is defined: we are doing cross-attn, so we should use cross-attn mask. + mask = None if encoder_hidden_states is None else encoder_attention_mask + else: + # when attention_mask is defined: we don't even check for encoder_attention_mask. + # this is to maintain compatibility with UnCLIP, which uses 'attention_mask' param for cross-attn masks. + # TODO: UnCLIP should express cross-attn mask via encoder_attention_mask param instead of via attention_mask. + # then we can simplify this whole if/else block to: + # mask = attention_mask if encoder_hidden_states is None else encoder_attention_mask + mask = attention_mask + + for resnet, attn in zip(self.resnets, self.attentions): + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=mask, + **cross_attention_kwargs, + ) + else: + hidden_states = resnet(hidden_states, temb, scale=lora_scale) + + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=mask, + **cross_attention_kwargs, + ) + + output_states = output_states + (hidden_states,) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states, temb, scale=lora_scale) + + output_states = output_states + (hidden_states,) + + return hidden_states, output_states + + +class KDownBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 4, + resnet_eps: float = 1e-5, + resnet_act_fn: str = "gelu", + resnet_group_size: int = 32, + add_downsample: bool = False, + ): + super().__init__() + resnets = [] + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + groups = in_channels // resnet_group_size + groups_out = out_channels // resnet_group_size + + resnets.append( + ResnetBlockCondNorm2D( + in_channels=in_channels, + out_channels=out_channels, + dropout=dropout, + temb_channels=temb_channels, + groups=groups, + groups_out=groups_out, + eps=resnet_eps, + non_linearity=resnet_act_fn, + time_embedding_norm="ada_group", + conv_shortcut_bias=False, + ) + ) + + self.resnets = nn.ModuleList(resnets) + + if add_downsample: + # YiYi's comments- might be able to use FirDownsample2D, look into details later + self.downsamplers = nn.ModuleList([KDownsample2D()]) + else: + self.downsamplers = None + + self.gradient_checkpointing = False + + def forward( + self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0 + ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]: + output_states = () + + for resnet in self.resnets: + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module): + def custom_forward(*inputs): + return module(*inputs) + + return custom_forward + + if is_torch_version(">=", "1.11.0"): + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb, use_reentrant=False + ) + else: + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb + ) + else: + hidden_states = resnet(hidden_states, temb, scale) + + output_states += (hidden_states,) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states) + + return hidden_states, output_states + + +class KCrossAttnDownBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + cross_attention_dim: int, + dropout: float = 0.0, + num_layers: int = 4, + resnet_group_size: int = 32, + add_downsample: bool = True, + attention_head_dim: int = 64, + add_self_attention: bool = False, + resnet_eps: float = 1e-5, + resnet_act_fn: str = "gelu", + ): + super().__init__() + resnets = [] + attentions = [] + + self.has_cross_attention = True + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + groups = in_channels // resnet_group_size + groups_out = out_channels // resnet_group_size + + resnets.append( + ResnetBlockCondNorm2D( + in_channels=in_channels, + out_channels=out_channels, + dropout=dropout, + temb_channels=temb_channels, + groups=groups, + groups_out=groups_out, + eps=resnet_eps, + non_linearity=resnet_act_fn, + time_embedding_norm="ada_group", + conv_shortcut_bias=False, + ) + ) + attentions.append( + KAttentionBlock( + out_channels, + out_channels // attention_head_dim, + attention_head_dim, + cross_attention_dim=cross_attention_dim, + temb_channels=temb_channels, + attention_bias=True, + add_self_attention=add_self_attention, + cross_attention_norm="layer_norm", + group_size=resnet_group_size, + ) + ) + + self.resnets = nn.ModuleList(resnets) + self.attentions = nn.ModuleList(attentions) + + if add_downsample: + self.downsamplers = nn.ModuleList([KDownsample2D()]) + else: + self.downsamplers = None + + self.gradient_checkpointing = False + + def forward( + self, + hidden_states: torch.FloatTensor, + temb: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]: + output_states = () + lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0 + + for resnet, attn in zip(self.resnets, self.attentions): + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {} + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), + hidden_states, + temb, + **ckpt_kwargs, + ) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + emb=temb, + attention_mask=attention_mask, + cross_attention_kwargs=cross_attention_kwargs, + encoder_attention_mask=encoder_attention_mask, + ) + else: + hidden_states = resnet(hidden_states, temb, scale=lora_scale) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + emb=temb, + attention_mask=attention_mask, + cross_attention_kwargs=cross_attention_kwargs, + encoder_attention_mask=encoder_attention_mask, + ) + + if self.downsamplers is None: + output_states += (None,) + else: + output_states += (hidden_states,) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states) + + return hidden_states, output_states + + +class AttnUpBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + prev_output_channel: int, + out_channels: int, + temb_channels: int, + resolution_idx: int = None, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + output_scale_factor: float = 1.0, + upsample_type: str = "conv", + ): + super().__init__() + resnets = [] + attentions = [] + + self.upsample_type = upsample_type + + if attention_head_dim is None: + logger.warn( + f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {out_channels}." + ) + attention_head_dim = out_channels + + for i in range(num_layers): + res_skip_channels = in_channels if (i == num_layers - 1) else out_channels + resnet_in_channels = prev_output_channel if i == 0 else out_channels + + resnets.append( + ResnetBlock2D( + in_channels=resnet_in_channels + res_skip_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + attentions.append( + Attention( + out_channels, + heads=out_channels // attention_head_dim, + dim_head=attention_head_dim, + rescale_output_factor=output_scale_factor, + eps=resnet_eps, + norm_num_groups=resnet_groups, + residual_connection=True, + bias=True, + upcast_softmax=True, + _from_deprecated_attn_block=True, + ) + ) + + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if upsample_type == "conv": + self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)]) + elif upsample_type == "resnet": + self.upsamplers = nn.ModuleList( + [ + ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + up=True, + ) + ] + ) + else: + self.upsamplers = None + + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + upsample_size: Optional[int] = None, + scale: float = 1.0, + ) -> torch.FloatTensor: + for resnet, attn in zip(self.resnets, self.attentions): + # pop res hidden states + res_hidden_states = res_hidden_states_tuple[-1] + res_hidden_states_tuple = res_hidden_states_tuple[:-1] + hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1) + + hidden_states = resnet(hidden_states, temb, scale=scale) + cross_attention_kwargs = {"scale": scale} + hidden_states = attn(hidden_states, **cross_attention_kwargs) + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + if self.upsample_type == "resnet": + hidden_states = upsampler(hidden_states, temb=temb, scale=scale) + else: + hidden_states = upsampler(hidden_states, scale=scale) + + return hidden_states + + +class CrossAttnUpBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + prev_output_channel: int, + temb_channels: int, + resolution_idx: Optional[int] = None, + dropout: float = 0.0, + num_layers: int = 1, + transformer_layers_per_block: Union[int, Tuple[int]] = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + num_attention_heads: int = 1, + cross_attention_dim: int = 1280, + output_scale_factor: float = 1.0, + add_upsample: bool = True, + dual_cross_attention: bool = False, + use_linear_projection: bool = False, + only_cross_attention: bool = False, + upcast_attention: bool = False, + attention_type: str = "default", + ): + super().__init__() + resnets = [] + attentions = [] + + self.has_cross_attention = True + self.num_attention_heads = num_attention_heads + + if isinstance(transformer_layers_per_block, int): + transformer_layers_per_block = [transformer_layers_per_block] * num_layers + + for i in range(num_layers): + res_skip_channels = in_channels if (i == num_layers - 1) else out_channels + resnet_in_channels = prev_output_channel if i == 0 else out_channels + + resnets.append( + ResnetBlock2D( + in_channels=resnet_in_channels + res_skip_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + if not dual_cross_attention: + attentions.append( + Transformer2DModel( + num_attention_heads, + out_channels // num_attention_heads, + in_channels=out_channels, + num_layers=transformer_layers_per_block[i], + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention, + upcast_attention=upcast_attention, + attention_type=attention_type, + ) + ) + else: + attentions.append( + DualTransformer2DModel( + num_attention_heads, + out_channels // num_attention_heads, + in_channels=out_channels, + num_layers=1, + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + ) + ) + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if add_upsample: + self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)]) + else: + self.upsamplers = None + + self.gradient_checkpointing = False + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + upsample_size: Optional[int] = None, + attention_mask: Optional[torch.FloatTensor] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + ) -> torch.FloatTensor: + lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0 + is_freeu_enabled = ( + getattr(self, "s1", None) + and getattr(self, "s2", None) + and getattr(self, "b1", None) + and getattr(self, "b2", None) + ) + + for resnet, attn in zip(self.resnets, self.attentions): + # pop res hidden states + res_hidden_states = res_hidden_states_tuple[-1] + res_hidden_states_tuple = res_hidden_states_tuple[:-1] + + # FreeU: Only operate on the first two stages + if is_freeu_enabled: + hidden_states, res_hidden_states = apply_freeu( + self.resolution_idx, + hidden_states, + res_hidden_states, + s1=self.s1, + s2=self.s2, + b1=self.b1, + b2=self.b2, + ) + + hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1) + + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {} + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), + hidden_states, + temb, + **ckpt_kwargs, + ) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + return_dict=False, + )[0] + else: + hidden_states = resnet(hidden_states, temb, scale=lora_scale) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + return_dict=False, + )[0] + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states, upsample_size, scale=lora_scale) + + return hidden_states + + +class UpBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + prev_output_channel: int, + out_channels: int, + temb_channels: int, + resolution_idx: Optional[int] = None, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + output_scale_factor: float = 1.0, + add_upsample: bool = True, + ): + super().__init__() + resnets = [] + + for i in range(num_layers): + res_skip_channels = in_channels if (i == num_layers - 1) else out_channels + resnet_in_channels = prev_output_channel if i == 0 else out_channels + + resnets.append( + ResnetBlock2D( + in_channels=resnet_in_channels + res_skip_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + self.resnets = nn.ModuleList(resnets) + + if add_upsample: + self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)]) + else: + self.upsamplers = None + + self.gradient_checkpointing = False + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + upsample_size: Optional[int] = None, + scale: float = 1.0, + ) -> torch.FloatTensor: + is_freeu_enabled = ( + getattr(self, "s1", None) + and getattr(self, "s2", None) + and getattr(self, "b1", None) + and getattr(self, "b2", None) + ) + + for resnet in self.resnets: + # pop res hidden states + res_hidden_states = res_hidden_states_tuple[-1] + res_hidden_states_tuple = res_hidden_states_tuple[:-1] + + # FreeU: Only operate on the first two stages + if is_freeu_enabled: + hidden_states, res_hidden_states = apply_freeu( + self.resolution_idx, + hidden_states, + res_hidden_states, + s1=self.s1, + s2=self.s2, + b1=self.b1, + b2=self.b2, + ) + + hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1) + + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module): + def custom_forward(*inputs): + return module(*inputs) + + return custom_forward + + if is_torch_version(">=", "1.11.0"): + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb, use_reentrant=False + ) + else: + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb + ) + else: + hidden_states = resnet(hidden_states, temb, scale=scale) + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states, upsample_size, scale=scale) + + return hidden_states + + +class UpDecoderBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + resolution_idx: Optional[int] = None, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", # default, spatial + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + output_scale_factor: float = 1.0, + add_upsample: bool = True, + temb_channels: Optional[int] = None, + ): + super().__init__() + resnets = [] + + for i in range(num_layers): + input_channels = in_channels if i == 0 else out_channels + + if resnet_time_scale_shift == "spatial": + resnets.append( + ResnetBlockCondNorm2D( + in_channels=input_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm="spatial", + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + ) + ) + else: + resnets.append( + ResnetBlock2D( + in_channels=input_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + self.resnets = nn.ModuleList(resnets) + + if add_upsample: + self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)]) + else: + self.upsamplers = None + + self.resolution_idx = resolution_idx + + def forward( + self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0 + ) -> torch.FloatTensor: + for resnet in self.resnets: + hidden_states = resnet(hidden_states, temb=temb, scale=scale) + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states) + + return hidden_states + + +class AttnUpDecoderBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + resolution_idx: Optional[int] = None, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + output_scale_factor: float = 1.0, + add_upsample: bool = True, + temb_channels: Optional[int] = None, + ): + super().__init__() + resnets = [] + attentions = [] + + if attention_head_dim is None: + logger.warn( + f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `out_channels`: {out_channels}." + ) + attention_head_dim = out_channels + + for i in range(num_layers): + input_channels = in_channels if i == 0 else out_channels + + if resnet_time_scale_shift == "spatial": + resnets.append( + ResnetBlockCondNorm2D( + in_channels=input_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm="spatial", + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + ) + ) + else: + resnets.append( + ResnetBlock2D( + in_channels=input_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + attentions.append( + Attention( + out_channels, + heads=out_channels // attention_head_dim, + dim_head=attention_head_dim, + rescale_output_factor=output_scale_factor, + eps=resnet_eps, + norm_num_groups=resnet_groups if resnet_time_scale_shift != "spatial" else None, + spatial_norm_dim=temb_channels if resnet_time_scale_shift == "spatial" else None, + residual_connection=True, + bias=True, + upcast_softmax=True, + _from_deprecated_attn_block=True, + ) + ) + + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if add_upsample: + self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)]) + else: + self.upsamplers = None + + self.resolution_idx = resolution_idx + + def forward( + self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0 + ) -> torch.FloatTensor: + for resnet, attn in zip(self.resnets, self.attentions): + hidden_states = resnet(hidden_states, temb=temb, scale=scale) + cross_attention_kwargs = {"scale": scale} + hidden_states = attn(hidden_states, temb=temb, **cross_attention_kwargs) + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states, scale=scale) + + return hidden_states + + +class AttnSkipUpBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + prev_output_channel: int, + out_channels: int, + temb_channels: int, + resolution_idx: Optional[int] = None, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + output_scale_factor: float = np.sqrt(2.0), + add_upsample: bool = True, + ): + super().__init__() + self.attentions = nn.ModuleList([]) + self.resnets = nn.ModuleList([]) + + for i in range(num_layers): + res_skip_channels = in_channels if (i == num_layers - 1) else out_channels + resnet_in_channels = prev_output_channel if i == 0 else out_channels + + self.resnets.append( + ResnetBlock2D( + in_channels=resnet_in_channels + res_skip_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=min(resnet_in_channels + res_skip_channels // 4, 32), + groups_out=min(out_channels // 4, 32), + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + if attention_head_dim is None: + logger.warn( + f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `out_channels`: {out_channels}." + ) + attention_head_dim = out_channels + + self.attentions.append( + Attention( + out_channels, + heads=out_channels // attention_head_dim, + dim_head=attention_head_dim, + rescale_output_factor=output_scale_factor, + eps=resnet_eps, + norm_num_groups=32, + residual_connection=True, + bias=True, + upcast_softmax=True, + _from_deprecated_attn_block=True, + ) + ) + + self.upsampler = FirUpsample2D(in_channels, out_channels=out_channels) + if add_upsample: + self.resnet_up = ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=min(out_channels // 4, 32), + groups_out=min(out_channels // 4, 32), + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + use_in_shortcut=True, + up=True, + kernel="fir", + ) + self.skip_conv = nn.Conv2d(out_channels, 3, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) + self.skip_norm = torch.nn.GroupNorm( + num_groups=min(out_channels // 4, 32), num_channels=out_channels, eps=resnet_eps, affine=True + ) + self.act = nn.SiLU() + else: + self.resnet_up = None + self.skip_conv = None + self.skip_norm = None + self.act = None + + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + skip_sample=None, + scale: float = 1.0, + ) -> Tuple[torch.FloatTensor, torch.FloatTensor]: + for resnet in self.resnets: + # pop res hidden states + res_hidden_states = res_hidden_states_tuple[-1] + res_hidden_states_tuple = res_hidden_states_tuple[:-1] + hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1) + + hidden_states = resnet(hidden_states, temb, scale=scale) + + cross_attention_kwargs = {"scale": scale} + hidden_states = self.attentions[0](hidden_states, **cross_attention_kwargs) + + if skip_sample is not None: + skip_sample = self.upsampler(skip_sample) + else: + skip_sample = 0 + + if self.resnet_up is not None: + skip_sample_states = self.skip_norm(hidden_states) + skip_sample_states = self.act(skip_sample_states) + skip_sample_states = self.skip_conv(skip_sample_states) + + skip_sample = skip_sample + skip_sample_states + + hidden_states = self.resnet_up(hidden_states, temb, scale=scale) + + return hidden_states, skip_sample + + +class SkipUpBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + prev_output_channel: int, + out_channels: int, + temb_channels: int, + resolution_idx: Optional[int] = None, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_pre_norm: bool = True, + output_scale_factor: float = np.sqrt(2.0), + add_upsample: bool = True, + upsample_padding: int = 1, + ): + super().__init__() + self.resnets = nn.ModuleList([]) + + for i in range(num_layers): + res_skip_channels = in_channels if (i == num_layers - 1) else out_channels + resnet_in_channels = prev_output_channel if i == 0 else out_channels + + self.resnets.append( + ResnetBlock2D( + in_channels=resnet_in_channels + res_skip_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=min((resnet_in_channels + res_skip_channels) // 4, 32), + groups_out=min(out_channels // 4, 32), + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + self.upsampler = FirUpsample2D(in_channels, out_channels=out_channels) + if add_upsample: + self.resnet_up = ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=min(out_channels // 4, 32), + groups_out=min(out_channels // 4, 32), + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + use_in_shortcut=True, + up=True, + kernel="fir", + ) + self.skip_conv = nn.Conv2d(out_channels, 3, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) + self.skip_norm = torch.nn.GroupNorm( + num_groups=min(out_channels // 4, 32), num_channels=out_channels, eps=resnet_eps, affine=True + ) + self.act = nn.SiLU() + else: + self.resnet_up = None + self.skip_conv = None + self.skip_norm = None + self.act = None + + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + skip_sample=None, + scale: float = 1.0, + ) -> Tuple[torch.FloatTensor, torch.FloatTensor]: + for resnet in self.resnets: + # pop res hidden states + res_hidden_states = res_hidden_states_tuple[-1] + res_hidden_states_tuple = res_hidden_states_tuple[:-1] + hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1) + + hidden_states = resnet(hidden_states, temb, scale=scale) + + if skip_sample is not None: + skip_sample = self.upsampler(skip_sample) + else: + skip_sample = 0 + + if self.resnet_up is not None: + skip_sample_states = self.skip_norm(hidden_states) + skip_sample_states = self.act(skip_sample_states) + skip_sample_states = self.skip_conv(skip_sample_states) + + skip_sample = skip_sample + skip_sample_states + + hidden_states = self.resnet_up(hidden_states, temb, scale=scale) + + return hidden_states, skip_sample + + +class ResnetUpsampleBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + prev_output_channel: int, + out_channels: int, + temb_channels: int, + resolution_idx: Optional[int] = None, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + output_scale_factor: float = 1.0, + add_upsample: bool = True, + skip_time_act: bool = False, + ): + super().__init__() + resnets = [] + + for i in range(num_layers): + res_skip_channels = in_channels if (i == num_layers - 1) else out_channels + resnet_in_channels = prev_output_channel if i == 0 else out_channels + + resnets.append( + ResnetBlock2D( + in_channels=resnet_in_channels + res_skip_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + ) + ) + + self.resnets = nn.ModuleList(resnets) + + if add_upsample: + self.upsamplers = nn.ModuleList( + [ + ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + up=True, + ) + ] + ) + else: + self.upsamplers = None + + self.gradient_checkpointing = False + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + upsample_size: Optional[int] = None, + scale: float = 1.0, + ) -> torch.FloatTensor: + for resnet in self.resnets: + # pop res hidden states + res_hidden_states = res_hidden_states_tuple[-1] + res_hidden_states_tuple = res_hidden_states_tuple[:-1] + hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1) + + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module): + def custom_forward(*inputs): + return module(*inputs) + + return custom_forward + + if is_torch_version(">=", "1.11.0"): + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb, use_reentrant=False + ) + else: + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb + ) + else: + hidden_states = resnet(hidden_states, temb, scale=scale) + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states, temb, scale=scale) + + return hidden_states + + +class SimpleCrossAttnUpBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + prev_output_channel: int, + temb_channels: int, + resolution_idx: Optional[int] = None, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + cross_attention_dim: int = 1280, + output_scale_factor: float = 1.0, + add_upsample: bool = True, + skip_time_act: bool = False, + only_cross_attention: bool = False, + cross_attention_norm: Optional[str] = None, + ): + super().__init__() + resnets = [] + attentions = [] + + self.has_cross_attention = True + self.attention_head_dim = attention_head_dim + + self.num_heads = out_channels // self.attention_head_dim + + for i in range(num_layers): + res_skip_channels = in_channels if (i == num_layers - 1) else out_channels + resnet_in_channels = prev_output_channel if i == 0 else out_channels + + resnets.append( + ResnetBlock2D( + in_channels=resnet_in_channels + res_skip_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + ) + ) + + processor = ( + AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor() + ) + + attentions.append( + Attention( + query_dim=out_channels, + cross_attention_dim=out_channels, + heads=self.num_heads, + dim_head=self.attention_head_dim, + added_kv_proj_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + bias=True, + upcast_softmax=True, + only_cross_attention=only_cross_attention, + cross_attention_norm=cross_attention_norm, + processor=processor, + ) + ) + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if add_upsample: + self.upsamplers = nn.ModuleList( + [ + ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + up=True, + ) + ] + ) + else: + self.upsamplers = None + + self.gradient_checkpointing = False + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + upsample_size: Optional[int] = None, + attention_mask: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + ) -> torch.FloatTensor: + cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {} + + lora_scale = cross_attention_kwargs.get("scale", 1.0) + if attention_mask is None: + # if encoder_hidden_states is defined: we are doing cross-attn, so we should use cross-attn mask. + mask = None if encoder_hidden_states is None else encoder_attention_mask + else: + # when attention_mask is defined: we don't even check for encoder_attention_mask. + # this is to maintain compatibility with UnCLIP, which uses 'attention_mask' param for cross-attn masks. + # TODO: UnCLIP should express cross-attn mask via encoder_attention_mask param instead of via attention_mask. + # then we can simplify this whole if/else block to: + # mask = attention_mask if encoder_hidden_states is None else encoder_attention_mask + mask = attention_mask + + for resnet, attn in zip(self.resnets, self.attentions): + # resnet + # pop res hidden states + res_hidden_states = res_hidden_states_tuple[-1] + res_hidden_states_tuple = res_hidden_states_tuple[:-1] + hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1) + + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=mask, + **cross_attention_kwargs, + ) + else: + hidden_states = resnet(hidden_states, temb, scale=lora_scale) + + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=mask, + **cross_attention_kwargs, + ) + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states, temb, scale=lora_scale) + + return hidden_states + + +class KUpBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + resolution_idx: int, + dropout: float = 0.0, + num_layers: int = 5, + resnet_eps: float = 1e-5, + resnet_act_fn: str = "gelu", + resnet_group_size: Optional[int] = 32, + add_upsample: bool = True, + ): + super().__init__() + resnets = [] + k_in_channels = 2 * out_channels + k_out_channels = in_channels + num_layers = num_layers - 1 + + for i in range(num_layers): + in_channels = k_in_channels if i == 0 else out_channels + groups = in_channels // resnet_group_size + groups_out = out_channels // resnet_group_size + + resnets.append( + ResnetBlockCondNorm2D( + in_channels=in_channels, + out_channels=k_out_channels if (i == num_layers - 1) else out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=groups, + groups_out=groups_out, + dropout=dropout, + non_linearity=resnet_act_fn, + time_embedding_norm="ada_group", + conv_shortcut_bias=False, + ) + ) + + self.resnets = nn.ModuleList(resnets) + + if add_upsample: + self.upsamplers = nn.ModuleList([KUpsample2D()]) + else: + self.upsamplers = None + + self.gradient_checkpointing = False + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + upsample_size: Optional[int] = None, + scale: float = 1.0, + ) -> torch.FloatTensor: + res_hidden_states_tuple = res_hidden_states_tuple[-1] + if res_hidden_states_tuple is not None: + hidden_states = torch.cat([hidden_states, res_hidden_states_tuple], dim=1) + + for resnet in self.resnets: + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module): + def custom_forward(*inputs): + return module(*inputs) + + return custom_forward + + if is_torch_version(">=", "1.11.0"): + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb, use_reentrant=False + ) + else: + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb + ) + else: + hidden_states = resnet(hidden_states, temb, scale=scale) + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states) + + return hidden_states + + +class KCrossAttnUpBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + resolution_idx: int, + dropout: float = 0.0, + num_layers: int = 4, + resnet_eps: float = 1e-5, + resnet_act_fn: str = "gelu", + resnet_group_size: int = 32, + attention_head_dim: int = 1, # attention dim_head + cross_attention_dim: int = 768, + add_upsample: bool = True, + upcast_attention: bool = False, + ): + super().__init__() + resnets = [] + attentions = [] + + is_first_block = in_channels == out_channels == temb_channels + is_middle_block = in_channels != out_channels + add_self_attention = True if is_first_block else False + + self.has_cross_attention = True + self.attention_head_dim = attention_head_dim + + # in_channels, and out_channels for the block (k-unet) + k_in_channels = out_channels if is_first_block else 2 * out_channels + k_out_channels = in_channels + + num_layers = num_layers - 1 + + for i in range(num_layers): + in_channels = k_in_channels if i == 0 else out_channels + groups = in_channels // resnet_group_size + groups_out = out_channels // resnet_group_size + + if is_middle_block and (i == num_layers - 1): + conv_2d_out_channels = k_out_channels + else: + conv_2d_out_channels = None + + resnets.append( + ResnetBlockCondNorm2D( + in_channels=in_channels, + out_channels=out_channels, + conv_2d_out_channels=conv_2d_out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=groups, + groups_out=groups_out, + dropout=dropout, + non_linearity=resnet_act_fn, + time_embedding_norm="ada_group", + conv_shortcut_bias=False, + ) + ) + attentions.append( + KAttentionBlock( + k_out_channels if (i == num_layers - 1) else out_channels, + k_out_channels // attention_head_dim + if (i == num_layers - 1) + else out_channels // attention_head_dim, + attention_head_dim, + cross_attention_dim=cross_attention_dim, + temb_channels=temb_channels, + attention_bias=True, + add_self_attention=add_self_attention, + cross_attention_norm="layer_norm", + upcast_attention=upcast_attention, + ) + ) + + self.resnets = nn.ModuleList(resnets) + self.attentions = nn.ModuleList(attentions) + + if add_upsample: + self.upsamplers = nn.ModuleList([KUpsample2D()]) + else: + self.upsamplers = None + + self.gradient_checkpointing = False + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + upsample_size: Optional[int] = None, + attention_mask: Optional[torch.FloatTensor] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + ) -> torch.FloatTensor: + res_hidden_states_tuple = res_hidden_states_tuple[-1] + if res_hidden_states_tuple is not None: + hidden_states = torch.cat([hidden_states, res_hidden_states_tuple], dim=1) + + lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0 + for resnet, attn in zip(self.resnets, self.attentions): + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {} + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), + hidden_states, + temb, + **ckpt_kwargs, + ) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + emb=temb, + attention_mask=attention_mask, + cross_attention_kwargs=cross_attention_kwargs, + encoder_attention_mask=encoder_attention_mask, + ) + else: + hidden_states = resnet(hidden_states, temb, scale=lora_scale) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + emb=temb, + attention_mask=attention_mask, + cross_attention_kwargs=cross_attention_kwargs, + encoder_attention_mask=encoder_attention_mask, + ) + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states) + + return hidden_states + + +# can potentially later be renamed to `No-feed-forward` attention +class KAttentionBlock(nn.Module): + r""" + A basic Transformer block. + + Parameters: + dim (`int`): The number of channels in the input and output. + num_attention_heads (`int`): The number of heads to use for multi-head attention. + attention_head_dim (`int`): The number of channels in each head. + dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. + cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention. + attention_bias (`bool`, *optional*, defaults to `False`): + Configure if the attention layers should contain a bias parameter. + upcast_attention (`bool`, *optional*, defaults to `False`): + Set to `True` to upcast the attention computation to `float32`. + temb_channels (`int`, *optional*, defaults to 768): + The number of channels in the token embedding. + add_self_attention (`bool`, *optional*, defaults to `False`): + Set to `True` to add self-attention to the block. + cross_attention_norm (`str`, *optional*, defaults to `None`): + The type of normalization to use for the cross attention. Can be `None`, `layer_norm`, or `group_norm`. + group_size (`int`, *optional*, defaults to 32): + The number of groups to separate the channels into for group normalization. + """ + + def __init__( + self, + dim: int, + num_attention_heads: int, + attention_head_dim: int, + dropout: float = 0.0, + cross_attention_dim: Optional[int] = None, + attention_bias: bool = False, + upcast_attention: bool = False, + temb_channels: int = 768, # for ada_group_norm + add_self_attention: bool = False, + cross_attention_norm: Optional[str] = None, + group_size: int = 32, + ): + super().__init__() + self.add_self_attention = add_self_attention + + # 1. Self-Attn + if add_self_attention: + self.norm1 = AdaGroupNorm(temb_channels, dim, max(1, dim // group_size)) + self.attn1 = Attention( + query_dim=dim, + heads=num_attention_heads, + dim_head=attention_head_dim, + dropout=dropout, + bias=attention_bias, + cross_attention_dim=None, + cross_attention_norm=None, + ) + + # 2. Cross-Attn + self.norm2 = AdaGroupNorm(temb_channels, dim, max(1, dim // group_size)) + self.attn2 = Attention( + query_dim=dim, + cross_attention_dim=cross_attention_dim, + heads=num_attention_heads, + dim_head=attention_head_dim, + dropout=dropout, + bias=attention_bias, + upcast_attention=upcast_attention, + cross_attention_norm=cross_attention_norm, + ) + + def _to_3d(self, hidden_states: torch.FloatTensor, height: int, weight: int) -> torch.FloatTensor: + return hidden_states.permute(0, 2, 3, 1).reshape(hidden_states.shape[0], height * weight, -1) + + def _to_4d(self, hidden_states: torch.FloatTensor, height: int, weight: int) -> torch.FloatTensor: + return hidden_states.permute(0, 2, 1).reshape(hidden_states.shape[0], -1, height, weight) + + def forward( + self, + hidden_states: torch.FloatTensor, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + # TODO: mark emb as non-optional (self.norm2 requires it). + # requires assessing impact of change to positional param interface. + emb: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + ) -> torch.FloatTensor: + cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {} + + # 1. Self-Attention + if self.add_self_attention: + norm_hidden_states = self.norm1(hidden_states, emb) + + height, weight = norm_hidden_states.shape[2:] + norm_hidden_states = self._to_3d(norm_hidden_states, height, weight) + + attn_output = self.attn1( + norm_hidden_states, + encoder_hidden_states=None, + attention_mask=attention_mask, + **cross_attention_kwargs, + ) + attn_output = self._to_4d(attn_output, height, weight) + + hidden_states = attn_output + hidden_states + + # 2. Cross-Attention/None + norm_hidden_states = self.norm2(hidden_states, emb) + + height, weight = norm_hidden_states.shape[2:] + norm_hidden_states = self._to_3d(norm_hidden_states, height, weight) + attn_output = self.attn2( + norm_hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=attention_mask if encoder_hidden_states is None else encoder_attention_mask, + **cross_attention_kwargs, + ) + attn_output = self._to_4d(attn_output, height, weight) + + hidden_states = attn_output + hidden_states + + return hidden_states diff --git a/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/models/unets/unet_2d_condition.py b/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/models/unets/unet_2d_condition.py new file mode 100644 index 0000000000000000000000000000000000000000..bbe078d297cf740e02198ff0a8b02133b5d7b00b --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/models/unets/unet_2d_condition.py @@ -0,0 +1,1218 @@ +# Copyright 2023 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from dataclasses import dataclass +from typing import Any, Dict, List, Optional, Tuple, Union + +import torch +import torch.nn as nn +import torch.utils.checkpoint + +from diffusers.configuration_utils import ConfigMixin, register_to_config +from diffusers.loaders import PeftAdapterMixin, UNet2DConditionLoadersMixin +from diffusers.utils import USE_PEFT_BACKEND, BaseOutput, deprecate, logging, scale_lora_layers, unscale_lora_layers +from stablediffusion.layers.activations import get_activation +from stablediffusion.layers.attention_processor import ( + ADDED_KV_ATTENTION_PROCESSORS, + CROSS_ATTENTION_PROCESSORS, + Attention, + AttentionProcessor, + AttnAddedKVProcessor, + AttnProcessor, +) +from stablediffusion.layers.embeddings import ( + GaussianFourierProjection, + GLIGENTextBoundingboxProjection, + ImageHintTimeEmbedding, + ImageProjection, + ImageTimeEmbedding, + TextImageProjection, + TextImageTimeEmbedding, + TextTimeEmbedding, + TimestepEmbedding, + Timesteps, +) +from .unet_2d_blocks import ( + UNetMidBlock2D, + UNetMidBlock2DCrossAttn, + UNetMidBlock2DSimpleCrossAttn, + get_down_block, + get_up_block, +) +from stablediffusion.models.model_utils import ModelMixin + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +@dataclass +class UNet2DConditionOutput(BaseOutput): + """ + The output of [`UNet2DConditionModel`]. + + Args: + sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`): + The hidden states output conditioned on `encoder_hidden_states` input. Output of last layer of model. + """ + + sample: torch.FloatTensor = None + + +class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin, PeftAdapterMixin): + r""" + A conditional 2D UNet model that takes a noisy sample, conditional state, and a timestep and returns a sample + shaped output. + + This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented + for all models (such as downloading or saving). + + Parameters: + sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`): + Height and width of input/output sample. + in_channels (`int`, *optional*, defaults to 4): Number of channels in the input sample. + out_channels (`int`, *optional*, defaults to 4): Number of channels in the output. + center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample. + flip_sin_to_cos (`bool`, *optional*, defaults to `False`): + Whether to flip the sin to cos in the time embedding. + freq_shift (`int`, *optional*, defaults to 0): The frequency shift to apply to the time embedding. + down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`): + The tuple of downsample blocks to use. + mid_block_type (`str`, *optional*, defaults to `"UNetMidBlock2DCrossAttn"`): + Block type for middle of UNet, it can be one of `UNetMidBlock2DCrossAttn`, `UNetMidBlock2D`, or + `UNetMidBlock2DSimpleCrossAttn`. If `None`, the mid block layer is skipped. + up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D")`): + The tuple of upsample blocks to use. + only_cross_attention(`bool` or `Tuple[bool]`, *optional*, default to `False`): + Whether to include self-attention in the basic transformer blocks, see + [`~models.attention.BasicTransformerBlock`]. + block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`): + The tuple of output channels for each block. + layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block. + downsample_padding (`int`, *optional*, defaults to 1): The padding to use for the downsampling convolution. + mid_block_scale_factor (`float`, *optional*, defaults to 1.0): The scale factor to use for the mid block. + dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. + act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use. + norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for the normalization. + If `None`, normalization and activation layers is skipped in post-processing. + norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization. + cross_attention_dim (`int` or `Tuple[int]`, *optional*, defaults to 1280): + The dimension of the cross attention features. + transformer_layers_per_block (`int`, `Tuple[int]`, or `Tuple[Tuple]` , *optional*, defaults to 1): + The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for + [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`], + [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`]. + reverse_transformer_layers_per_block : (`Tuple[Tuple]`, *optional*, defaults to None): + The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`], in the upsampling + blocks of the U-Net. Only relevant if `transformer_layers_per_block` is of type `Tuple[Tuple]` and for + [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`], + [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`]. + encoder_hid_dim (`int`, *optional*, defaults to None): + If `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim` + dimension to `cross_attention_dim`. + encoder_hid_dim_type (`str`, *optional*, defaults to `None`): + If given, the `encoder_hidden_states` and potentially other embeddings are down-projected to text + embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`. + attention_head_dim (`int`, *optional*, defaults to 8): The dimension of the attention heads. + num_attention_heads (`int`, *optional*): + The number of attention heads. If not defined, defaults to `attention_head_dim` + resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config + for ResNet blocks (see [`~models.resnet.ResnetBlock2D`]). Choose from `default` or `scale_shift`. + class_embed_type (`str`, *optional*, defaults to `None`): + The type of class embedding to use which is ultimately summed with the time embeddings. Choose from `None`, + `"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`. + addition_embed_type (`str`, *optional*, defaults to `None`): + Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or + "text". "text" will use the `TextTimeEmbedding` layer. + addition_time_embed_dim: (`int`, *optional*, defaults to `None`): + Dimension for the timestep embeddings. + num_class_embeds (`int`, *optional*, defaults to `None`): + Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing + class conditioning with `class_embed_type` equal to `None`. + time_embedding_type (`str`, *optional*, defaults to `positional`): + The type of position embedding to use for timesteps. Choose from `positional` or `fourier`. + time_embedding_dim (`int`, *optional*, defaults to `None`): + An optional override for the dimension of the projected time embedding. + time_embedding_act_fn (`str`, *optional*, defaults to `None`): + Optional activation function to use only once on the time embeddings before they are passed to the rest of + the UNet. Choose from `silu`, `mish`, `gelu`, and `swish`. + timestep_post_act (`str`, *optional*, defaults to `None`): + The second activation function to use in timestep embedding. Choose from `silu`, `mish` and `gelu`. + time_cond_proj_dim (`int`, *optional*, defaults to `None`): + The dimension of `cond_proj` layer in the timestep embedding. + conv_in_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_in` layer. conv_out_kernel (`int`, + *optional*, default to `3`): The kernel size of `conv_out` layer. projection_class_embeddings_input_dim (`int`, + *optional*): The dimension of the `class_labels` input when + `class_embed_type="projection"`. Required when `class_embed_type="projection"`. + class_embeddings_concat (`bool`, *optional*, defaults to `False`): Whether to concatenate the time + embeddings with the class embeddings. + mid_block_only_cross_attention (`bool`, *optional*, defaults to `None`): + Whether to use cross attention with the mid block when using the `UNetMidBlock2DSimpleCrossAttn`. If + `only_cross_attention` is given as a single boolean and `mid_block_only_cross_attention` is `None`, the + `only_cross_attention` value is used as the value for `mid_block_only_cross_attention`. Default to `False` + otherwise. + """ + + _supports_gradient_checkpointing = True + + @register_to_config + def __init__( + self, + sample_size: Optional[int] = None, + in_channels: int = 4, + out_channels: int = 4, + center_input_sample: bool = False, + flip_sin_to_cos: bool = True, + freq_shift: int = 0, + down_block_types: Tuple[str] = ( + "CrossAttnDownBlock2D", + "CrossAttnDownBlock2D", + "CrossAttnDownBlock2D", + "DownBlock2D", + ), + mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn", + up_block_types: Tuple[str] = ("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D"), + only_cross_attention: Union[bool, Tuple[bool]] = False, + block_out_channels: Tuple[int] = (320, 640, 1280, 1280), + layers_per_block: Union[int, Tuple[int]] = 2, + downsample_padding: int = 1, + mid_block_scale_factor: float = 1, + dropout: float = 0.0, + act_fn: str = "silu", + norm_num_groups: Optional[int] = 32, + norm_eps: float = 1e-5, + cross_attention_dim: Union[int, Tuple[int]] = 1280, + transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple]] = 1, + reverse_transformer_layers_per_block: Optional[Tuple[Tuple[int]]] = None, + encoder_hid_dim: Optional[int] = None, + encoder_hid_dim_type: Optional[str] = None, + attention_head_dim: Union[int, Tuple[int]] = 8, + num_attention_heads: Optional[Union[int, Tuple[int]]] = None, + dual_cross_attention: bool = False, + use_linear_projection: bool = False, + class_embed_type: Optional[str] = None, + addition_embed_type: Optional[str] = None, + addition_time_embed_dim: Optional[int] = None, + num_class_embeds: Optional[int] = None, + upcast_attention: bool = False, + resnet_time_scale_shift: str = "default", + resnet_skip_time_act: bool = False, + resnet_out_scale_factor: int = 1.0, + time_embedding_type: str = "positional", + time_embedding_dim: Optional[int] = None, + time_embedding_act_fn: Optional[str] = None, + timestep_post_act: Optional[str] = None, + time_cond_proj_dim: Optional[int] = None, + conv_in_kernel: int = 3, + conv_out_kernel: int = 3, + projection_class_embeddings_input_dim: Optional[int] = None, + attention_type: str = "default", + class_embeddings_concat: bool = False, + mid_block_only_cross_attention: Optional[bool] = None, + cross_attention_norm: Optional[str] = None, + addition_embed_type_num_heads=64, + ): + super().__init__() + + self.sample_size = sample_size + + if num_attention_heads is not None: + raise ValueError( + "At the moment it is not possible to define the number of attention heads via `num_attention_heads` because of a naming issue as described in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131. Passing `num_attention_heads` will only be supported in diffusers v0.19." + ) + + # If `num_attention_heads` is not defined (which is the case for most models) + # it will default to `attention_head_dim`. This looks weird upon first reading it and it is. + # The reason for this behavior is to correct for incorrectly named variables that were introduced + # when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131 + # Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking + # which is why we correct for the naming here. + num_attention_heads = num_attention_heads or attention_head_dim + + # Check inputs + if len(down_block_types) != len(up_block_types): + raise ValueError( + f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}." + ) + + if len(block_out_channels) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}." + ) + + if not isinstance(only_cross_attention, bool) and len(only_cross_attention) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `only_cross_attention` as `down_block_types`. `only_cross_attention`: {only_cross_attention}. `down_block_types`: {down_block_types}." + ) + + if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}." + ) + + if not isinstance(attention_head_dim, int) and len(attention_head_dim) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `attention_head_dim` as `down_block_types`. `attention_head_dim`: {attention_head_dim}. `down_block_types`: {down_block_types}." + ) + + if isinstance(cross_attention_dim, list) and len(cross_attention_dim) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `cross_attention_dim` as `down_block_types`. `cross_attention_dim`: {cross_attention_dim}. `down_block_types`: {down_block_types}." + ) + + if not isinstance(layers_per_block, int) and len(layers_per_block) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `layers_per_block` as `down_block_types`. `layers_per_block`: {layers_per_block}. `down_block_types`: {down_block_types}." + ) + if isinstance(transformer_layers_per_block, list) and reverse_transformer_layers_per_block is None: + for layer_number_per_block in transformer_layers_per_block: + if isinstance(layer_number_per_block, list): + raise ValueError("Must provide 'reverse_transformer_layers_per_block` if using asymmetrical UNet.") + + # input + conv_in_padding = (conv_in_kernel - 1) // 2 + self.conv_in = nn.Conv2d( + in_channels, block_out_channels[0], kernel_size=conv_in_kernel, padding=conv_in_padding + ) + + # time + if time_embedding_type == "fourier": + time_embed_dim = time_embedding_dim or block_out_channels[0] * 2 + if time_embed_dim % 2 != 0: + raise ValueError(f"`time_embed_dim` should be divisible by 2, but is {time_embed_dim}.") + self.time_proj = GaussianFourierProjection( + time_embed_dim // 2, set_W_to_weight=False, log=False, flip_sin_to_cos=flip_sin_to_cos + ) + timestep_input_dim = time_embed_dim + elif time_embedding_type == "positional": + time_embed_dim = time_embedding_dim or block_out_channels[0] * 4 + + self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift) + timestep_input_dim = block_out_channels[0] + else: + raise ValueError( + f"{time_embedding_type} does not exist. Please make sure to use one of `fourier` or `positional`." + ) + + self.time_embedding = TimestepEmbedding( + timestep_input_dim, + time_embed_dim, + act_fn=act_fn, + post_act_fn=timestep_post_act, + cond_proj_dim=time_cond_proj_dim, + ) + + if encoder_hid_dim_type is None and encoder_hid_dim is not None: + encoder_hid_dim_type = "text_proj" + self.register_to_config(encoder_hid_dim_type=encoder_hid_dim_type) + logger.info("encoder_hid_dim_type defaults to 'text_proj' as `encoder_hid_dim` is defined.") + + if encoder_hid_dim is None and encoder_hid_dim_type is not None: + raise ValueError( + f"`encoder_hid_dim` has to be defined when `encoder_hid_dim_type` is set to {encoder_hid_dim_type}." + ) + + if encoder_hid_dim_type == "text_proj": + self.encoder_hid_proj = nn.Linear(encoder_hid_dim, cross_attention_dim) + elif encoder_hid_dim_type == "text_image_proj": + # image_embed_dim DOESN'T have to be `cross_attention_dim`. To not clutter the __init__ too much + # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use + # case when `addition_embed_type == "text_image_proj"` (Kadinsky 2.1)` + self.encoder_hid_proj = TextImageProjection( + text_embed_dim=encoder_hid_dim, + image_embed_dim=cross_attention_dim, + cross_attention_dim=cross_attention_dim, + ) + elif encoder_hid_dim_type == "image_proj": + # Kandinsky 2.2 + self.encoder_hid_proj = ImageProjection( + image_embed_dim=encoder_hid_dim, + cross_attention_dim=cross_attention_dim, + ) + elif encoder_hid_dim_type is not None: + raise ValueError( + f"encoder_hid_dim_type: {encoder_hid_dim_type} must be None, 'text_proj' or 'text_image_proj'." + ) + else: + self.encoder_hid_proj = None + + # class embedding + if class_embed_type is None and num_class_embeds is not None: + self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim) + elif class_embed_type == "timestep": + self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim, act_fn=act_fn) + elif class_embed_type == "identity": + self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim) + elif class_embed_type == "projection": + if projection_class_embeddings_input_dim is None: + raise ValueError( + "`class_embed_type`: 'projection' requires `projection_class_embeddings_input_dim` be set" + ) + # The projection `class_embed_type` is the same as the timestep `class_embed_type` except + # 1. the `class_labels` inputs are not first converted to sinusoidal embeddings + # 2. it projects from an arbitrary input dimension. + # + # Note that `TimestepEmbedding` is quite general, being mainly linear layers and activations. + # When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings. + # As a result, `TimestepEmbedding` can be passed arbitrary vectors. + self.class_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim) + elif class_embed_type == "simple_projection": + if projection_class_embeddings_input_dim is None: + raise ValueError( + "`class_embed_type`: 'simple_projection' requires `projection_class_embeddings_input_dim` be set" + ) + self.class_embedding = nn.Linear(projection_class_embeddings_input_dim, time_embed_dim) + else: + self.class_embedding = None + + if addition_embed_type == "text": + if encoder_hid_dim is not None: + text_time_embedding_from_dim = encoder_hid_dim + else: + text_time_embedding_from_dim = cross_attention_dim + + self.add_embedding = TextTimeEmbedding( + text_time_embedding_from_dim, time_embed_dim, num_heads=addition_embed_type_num_heads + ) + elif addition_embed_type == "text_image": + # text_embed_dim and image_embed_dim DON'T have to be `cross_attention_dim`. To not clutter the __init__ too much + # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use + # case when `addition_embed_type == "text_image"` (Kadinsky 2.1)` + self.add_embedding = TextImageTimeEmbedding( + text_embed_dim=cross_attention_dim, image_embed_dim=cross_attention_dim, time_embed_dim=time_embed_dim + ) + elif addition_embed_type == "text_time": + self.add_time_proj = Timesteps(addition_time_embed_dim, flip_sin_to_cos, freq_shift) + self.add_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim) + elif addition_embed_type == "image": + # Kandinsky 2.2 + self.add_embedding = ImageTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim) + elif addition_embed_type == "image_hint": + # Kandinsky 2.2 ControlNet + self.add_embedding = ImageHintTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim) + elif addition_embed_type is not None: + raise ValueError(f"addition_embed_type: {addition_embed_type} must be None, 'text' or 'text_image'.") + + if time_embedding_act_fn is None: + self.time_embed_act = None + else: + self.time_embed_act = get_activation(time_embedding_act_fn) + + self.down_blocks = nn.ModuleList([]) + self.up_blocks = nn.ModuleList([]) + + if isinstance(only_cross_attention, bool): + if mid_block_only_cross_attention is None: + mid_block_only_cross_attention = only_cross_attention + + only_cross_attention = [only_cross_attention] * len(down_block_types) + + if mid_block_only_cross_attention is None: + mid_block_only_cross_attention = False + + if isinstance(num_attention_heads, int): + num_attention_heads = (num_attention_heads,) * len(down_block_types) + + if isinstance(attention_head_dim, int): + attention_head_dim = (attention_head_dim,) * len(down_block_types) + + if isinstance(cross_attention_dim, int): + cross_attention_dim = (cross_attention_dim,) * len(down_block_types) + + if isinstance(layers_per_block, int): + layers_per_block = [layers_per_block] * len(down_block_types) + + if isinstance(transformer_layers_per_block, int): + transformer_layers_per_block = [transformer_layers_per_block] * len(down_block_types) + + if class_embeddings_concat: + # The time embeddings are concatenated with the class embeddings. The dimension of the + # time embeddings passed to the down, middle, and up blocks is twice the dimension of the + # regular time embeddings + blocks_time_embed_dim = time_embed_dim * 2 + else: + blocks_time_embed_dim = time_embed_dim + + # down + output_channel = block_out_channels[0] + for i, down_block_type in enumerate(down_block_types): + input_channel = output_channel + output_channel = block_out_channels[i] + is_final_block = i == len(block_out_channels) - 1 + + down_block = get_down_block( + down_block_type, + num_layers=layers_per_block[i], + transformer_layers_per_block=transformer_layers_per_block[i], + in_channels=input_channel, + out_channels=output_channel, + temb_channels=blocks_time_embed_dim, + add_downsample=not is_final_block, + resnet_eps=norm_eps, + resnet_act_fn=act_fn, + resnet_groups=norm_num_groups, + cross_attention_dim=cross_attention_dim[i], + num_attention_heads=num_attention_heads[i], + downsample_padding=downsample_padding, + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention[i], + upcast_attention=upcast_attention, + resnet_time_scale_shift=resnet_time_scale_shift, + attention_type=attention_type, + resnet_skip_time_act=resnet_skip_time_act, + resnet_out_scale_factor=resnet_out_scale_factor, + cross_attention_norm=cross_attention_norm, + attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel, + dropout=dropout, + ) + self.down_blocks.append(down_block) + + # mid + if mid_block_type == "UNetMidBlock2DCrossAttn": + self.mid_block = UNetMidBlock2DCrossAttn( + transformer_layers_per_block=transformer_layers_per_block[-1], + in_channels=block_out_channels[-1], + temb_channels=blocks_time_embed_dim, + dropout=dropout, + resnet_eps=norm_eps, + resnet_act_fn=act_fn, + output_scale_factor=mid_block_scale_factor, + resnet_time_scale_shift=resnet_time_scale_shift, + cross_attention_dim=cross_attention_dim[-1], + num_attention_heads=num_attention_heads[-1], + resnet_groups=norm_num_groups, + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, + upcast_attention=upcast_attention, + attention_type=attention_type, + ) + elif mid_block_type == "UNetMidBlock2DSimpleCrossAttn": + self.mid_block = UNetMidBlock2DSimpleCrossAttn( + in_channels=block_out_channels[-1], + temb_channels=blocks_time_embed_dim, + dropout=dropout, + resnet_eps=norm_eps, + resnet_act_fn=act_fn, + output_scale_factor=mid_block_scale_factor, + cross_attention_dim=cross_attention_dim[-1], + attention_head_dim=attention_head_dim[-1], + resnet_groups=norm_num_groups, + resnet_time_scale_shift=resnet_time_scale_shift, + skip_time_act=resnet_skip_time_act, + only_cross_attention=mid_block_only_cross_attention, + cross_attention_norm=cross_attention_norm, + ) + elif mid_block_type == "UNetMidBlock2D": + self.mid_block = UNetMidBlock2D( + in_channels=block_out_channels[-1], + temb_channels=blocks_time_embed_dim, + dropout=dropout, + num_layers=0, + resnet_eps=norm_eps, + resnet_act_fn=act_fn, + output_scale_factor=mid_block_scale_factor, + resnet_groups=norm_num_groups, + resnet_time_scale_shift=resnet_time_scale_shift, + add_attention=False, + ) + elif mid_block_type is None: + self.mid_block = None + else: + raise ValueError(f"unknown mid_block_type : {mid_block_type}") + + # count how many layers upsample the images + self.num_upsamplers = 0 + + # up + reversed_block_out_channels = list(reversed(block_out_channels)) + reversed_num_attention_heads = list(reversed(num_attention_heads)) + reversed_layers_per_block = list(reversed(layers_per_block)) + reversed_cross_attention_dim = list(reversed(cross_attention_dim)) + reversed_transformer_layers_per_block = ( + list(reversed(transformer_layers_per_block)) + if reverse_transformer_layers_per_block is None + else reverse_transformer_layers_per_block + ) + only_cross_attention = list(reversed(only_cross_attention)) + + output_channel = reversed_block_out_channels[0] + for i, up_block_type in enumerate(up_block_types): + is_final_block = i == len(block_out_channels) - 1 + + prev_output_channel = output_channel + output_channel = reversed_block_out_channels[i] + input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)] + + # add upsample block for all BUT final layer + if not is_final_block: + add_upsample = True + self.num_upsamplers += 1 + else: + add_upsample = False + + up_block = get_up_block( + up_block_type, + num_layers=reversed_layers_per_block[i] + 1, + transformer_layers_per_block=reversed_transformer_layers_per_block[i], + in_channels=input_channel, + out_channels=output_channel, + prev_output_channel=prev_output_channel, + temb_channels=blocks_time_embed_dim, + add_upsample=add_upsample, + resnet_eps=norm_eps, + resnet_act_fn=act_fn, + resolution_idx=i, + resnet_groups=norm_num_groups, + cross_attention_dim=reversed_cross_attention_dim[i], + num_attention_heads=reversed_num_attention_heads[i], + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention[i], + upcast_attention=upcast_attention, + resnet_time_scale_shift=resnet_time_scale_shift, + attention_type=attention_type, + resnet_skip_time_act=resnet_skip_time_act, + resnet_out_scale_factor=resnet_out_scale_factor, + cross_attention_norm=cross_attention_norm, + attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel, + dropout=dropout, + ) + self.up_blocks.append(up_block) + prev_output_channel = output_channel + + # out + if norm_num_groups is not None: + self.conv_norm_out = nn.GroupNorm( + num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=norm_eps + ) + + self.conv_act = get_activation(act_fn) + + else: + self.conv_norm_out = None + self.conv_act = None + + conv_out_padding = (conv_out_kernel - 1) // 2 + self.conv_out = nn.Conv2d( + block_out_channels[0], out_channels, kernel_size=conv_out_kernel, padding=conv_out_padding + ) + + if attention_type in ["gated", "gated-text-image"]: + positive_len = 768 + if isinstance(cross_attention_dim, int): + positive_len = cross_attention_dim + elif isinstance(cross_attention_dim, tuple) or isinstance(cross_attention_dim, list): + positive_len = cross_attention_dim[0] + + feature_type = "text-only" if attention_type == "gated" else "text-image" + self.position_net = GLIGENTextBoundingboxProjection( + positive_len=positive_len, out_dim=cross_attention_dim, feature_type=feature_type + ) + + @property + def attn_processors(self) -> Dict[str, AttentionProcessor]: + r""" + Returns: + `dict` of attention processors: A dictionary containing all attention processors used in the model with + indexed by its weight name. + """ + # set recursively + processors = {} + + def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]): + if hasattr(module, "get_processor"): + processors[f"{name}.processor"] = module.get_processor(return_deprecated_lora=True) + + for sub_name, child in module.named_children(): + fn_recursive_add_processors(f"{name}.{sub_name}", child, processors) + + return processors + + for name, module in self.named_children(): + fn_recursive_add_processors(name, module, processors) + + return processors + + def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]): + r""" + Sets the attention processor to use to compute attention. + + Parameters: + processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`): + The instantiated processor class or a dictionary of processor classes that will be set as the processor + for **all** `Attention` layers. + + If `processor` is a dict, the key needs to define the path to the corresponding cross attention + processor. This is strongly recommended when setting trainable attention processors. + + """ + count = len(self.attn_processors.keys()) + + if isinstance(processor, dict) and len(processor) != count: + raise ValueError( + f"A dict of processors was passed, but the number of processors {len(processor)} does not match the" + f" number of attention layers: {count}. Please make sure to pass {count} processor classes." + ) + + def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor): + if hasattr(module, "set_processor"): + if not isinstance(processor, dict): + module.set_processor(processor) + else: + module.set_processor(processor.pop(f"{name}.processor")) + + for sub_name, child in module.named_children(): + fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor) + + for name, module in self.named_children(): + fn_recursive_attn_processor(name, module, processor) + + def set_default_attn_processor(self): + """ + Disables custom attention processors and sets the default attention implementation. + """ + if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()): + processor = AttnAddedKVProcessor() + elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()): + processor = AttnProcessor() + else: + raise ValueError( + f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}" + ) + + self.set_attn_processor(processor) + + def set_attention_slice(self, slice_size): + r""" + Enable sliced attention computation. + + When this option is enabled, the attention module splits the input tensor in slices to compute attention in + several steps. This is useful for saving some memory in exchange for a small decrease in speed. + + Args: + slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`): + When `"auto"`, input to the attention heads is halved, so attention is computed in two steps. If + `"max"`, maximum amount of memory is saved by running only one slice at a time. If a number is + provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim` + must be a multiple of `slice_size`. + """ + sliceable_head_dims = [] + + def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module): + if hasattr(module, "set_attention_slice"): + sliceable_head_dims.append(module.sliceable_head_dim) + + for child in module.children(): + fn_recursive_retrieve_sliceable_dims(child) + + # retrieve number of attention layers + for module in self.children(): + fn_recursive_retrieve_sliceable_dims(module) + + num_sliceable_layers = len(sliceable_head_dims) + + if slice_size == "auto": + # half the attention head size is usually a good trade-off between + # speed and memory + slice_size = [dim // 2 for dim in sliceable_head_dims] + elif slice_size == "max": + # make smallest slice possible + slice_size = num_sliceable_layers * [1] + + slice_size = num_sliceable_layers * [slice_size] if not isinstance(slice_size, list) else slice_size + + if len(slice_size) != len(sliceable_head_dims): + raise ValueError( + f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different" + f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}." + ) + + for i in range(len(slice_size)): + size = slice_size[i] + dim = sliceable_head_dims[i] + if size is not None and size > dim: + raise ValueError(f"size {size} has to be smaller or equal to {dim}.") + + # Recursively walk through all the children. + # Any children which exposes the set_attention_slice method + # gets the message + def fn_recursive_set_attention_slice(module: torch.nn.Module, slice_size: List[int]): + if hasattr(module, "set_attention_slice"): + module.set_attention_slice(slice_size.pop()) + + for child in module.children(): + fn_recursive_set_attention_slice(child, slice_size) + + reversed_slice_size = list(reversed(slice_size)) + for module in self.children(): + fn_recursive_set_attention_slice(module, reversed_slice_size) + + def _set_gradient_checkpointing(self, module, value=False): + if hasattr(module, "gradient_checkpointing"): + module.gradient_checkpointing = value + + def enable_freeu(self, s1, s2, b1, b2): + r"""Enables the FreeU mechanism from https://arxiv.org/abs/2309.11497. + + The suffixes after the scaling factors represent the stage blocks where they are being applied. + + Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of values that + are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL. + + Args: + s1 (`float`): + Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to + mitigate the "oversmoothing effect" in the enhanced denoising process. + s2 (`float`): + Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to + mitigate the "oversmoothing effect" in the enhanced denoising process. + b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features. + b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features. + """ + for i, upsample_block in enumerate(self.up_blocks): + setattr(upsample_block, "s1", s1) + setattr(upsample_block, "s2", s2) + setattr(upsample_block, "b1", b1) + setattr(upsample_block, "b2", b2) + + def disable_freeu(self): + """Disables the FreeU mechanism.""" + freeu_keys = {"s1", "s2", "b1", "b2"} + for i, upsample_block in enumerate(self.up_blocks): + for k in freeu_keys: + if hasattr(upsample_block, k) or getattr(upsample_block, k, None) is not None: + setattr(upsample_block, k, None) + + def fuse_qkv_projections(self): + """ + Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, + key, value) are fused. For cross-attention modules, key and value projection matrices are fused. + + + + This API is 🧪 experimental. + + + """ + self.original_attn_processors = None + + for _, attn_processor in self.attn_processors.items(): + if "Added" in str(attn_processor.__class__.__name__): + raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.") + + self.original_attn_processors = self.attn_processors + + for module in self.modules(): + if isinstance(module, Attention): + module.fuse_projections(fuse=True) + + def unfuse_qkv_projections(self): + """Disables the fused QKV projection if enabled. + + + + This API is 🧪 experimental. + + + + """ + if self.original_attn_processors is not None: + self.set_attn_processor(self.original_attn_processors) + + def unload_lora(self): + """Unloads LoRA weights.""" + deprecate( + "unload_lora", + "0.28.0", + "Calling `unload_lora()` is deprecated and will be removed in a future version. Please install `peft` and then call `disable_adapters().", + ) + for module in self.modules(): + if hasattr(module, "set_lora_layer"): + module.set_lora_layer(None) + + def forward( + self, + sample: torch.FloatTensor, + timestep: Union[torch.Tensor, float, int], + encoder_hidden_states: torch.Tensor, + class_labels: Optional[torch.Tensor] = None, + timestep_cond: Optional[torch.Tensor] = None, + attention_mask: Optional[torch.Tensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None, + down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None, + mid_block_additional_residual: Optional[torch.Tensor] = None, + down_intrablock_additional_residuals: Optional[Tuple[torch.Tensor]] = None, + encoder_attention_mask: Optional[torch.Tensor] = None, + return_dict: bool = True, + ) -> Union[UNet2DConditionOutput, Tuple]: + r""" + The [`UNet2DConditionModel`] forward method. + + Args: + sample (`torch.FloatTensor`): + The noisy input tensor with the following shape `(batch, channel, height, width)`. + timestep (`torch.FloatTensor` or `float` or `int`): The number of timesteps to denoise an input. + encoder_hidden_states (`torch.FloatTensor`): + The encoder hidden states with shape `(batch, sequence_length, feature_dim)`. + class_labels (`torch.Tensor`, *optional*, defaults to `None`): + Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings. + timestep_cond: (`torch.Tensor`, *optional*, defaults to `None`): + Conditional embeddings for timestep. If provided, the embeddings will be summed with the samples passed + through the `self.time_embedding` layer to obtain the timestep embeddings. + attention_mask (`torch.Tensor`, *optional*, defaults to `None`): + An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask + is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large + negative values to the attention scores corresponding to "discard" tokens. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + added_cond_kwargs: (`dict`, *optional*): + A kwargs dictionary containing additional embeddings that if specified are added to the embeddings that + are passed along to the UNet blocks. + down_block_additional_residuals: (`tuple` of `torch.Tensor`, *optional*): + A tuple of tensors that if specified are added to the residuals of down unet blocks. + mid_block_additional_residual: (`torch.Tensor`, *optional*): + A tensor that if specified is added to the residual of the middle unet block. + encoder_attention_mask (`torch.Tensor`): + A cross-attention mask of shape `(batch, sequence_length)` is applied to `encoder_hidden_states`. If + `True` the mask is kept, otherwise if `False` it is discarded. Mask will be converted into a bias, + which adds large negative values to the attention scores corresponding to "discard" tokens. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] instead of a plain + tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttnProcessor`]. + added_cond_kwargs: (`dict`, *optional*): + A kwargs dictionary containin additional embeddings that if specified are added to the embeddings that + are passed along to the UNet blocks. + down_block_additional_residuals (`tuple` of `torch.Tensor`, *optional*): + additional residuals to be added to UNet long skip connections from down blocks to up blocks for + example from ControlNet side model(s) + mid_block_additional_residual (`torch.Tensor`, *optional*): + additional residual to be added to UNet mid block output, for example from ControlNet side model + down_intrablock_additional_residuals (`tuple` of `torch.Tensor`, *optional*): + additional residuals to be added within UNet down blocks, for example from T2I-Adapter side model(s) + + Returns: + [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] or `tuple`: + If `return_dict` is True, an [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] is returned, otherwise + a `tuple` is returned where the first element is the sample tensor. + """ + # By default samples have to be AT least a multiple of the overall upsampling factor. + # The overall upsampling factor is equal to 2 ** (# num of upsampling layers). + # However, the upsampling interpolation output size can be forced to fit any upsampling size + # on the fly if necessary. + default_overall_up_factor = 2**self.num_upsamplers + + # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor` + forward_upsample_size = False + upsample_size = None + + for dim in sample.shape[-2:]: + if dim % default_overall_up_factor != 0: + # Forward upsample size to force interpolation output size. + forward_upsample_size = True + break + + # ensure attention_mask is a bias, and give it a singleton query_tokens dimension + # expects mask of shape: + # [batch, key_tokens] + # adds singleton query_tokens dimension: + # [batch, 1, key_tokens] + # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes: + # [batch, heads, query_tokens, key_tokens] (e.g. torch sdp attn) + # [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn) + if attention_mask is not None: + # assume that mask is expressed as: + # (1 = keep, 0 = discard) + # convert mask into a bias that can be added to attention scores: + # (keep = +0, discard = -10000.0) + attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0 + attention_mask = attention_mask.unsqueeze(1) + + # convert encoder_attention_mask to a bias the same way we do for attention_mask + if encoder_attention_mask is not None: + encoder_attention_mask = (1 - encoder_attention_mask.to(sample.dtype)) * -10000.0 + encoder_attention_mask = encoder_attention_mask.unsqueeze(1) + + # 0. center input if necessary + if self.config.center_input_sample: + sample = 2 * sample - 1.0 + + # 1. time + timesteps = timestep + if not torch.is_tensor(timesteps): + # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can + # This would be a good case for the `match` statement (Python 3.10+) + is_mps = sample.device.type == "mps" + if isinstance(timestep, float): + dtype = torch.float32 if is_mps else torch.float64 + else: + dtype = torch.int32 if is_mps else torch.int64 + timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device) + elif len(timesteps.shape) == 0: + timesteps = timesteps[None].to(sample.device) + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timesteps = timesteps.expand(sample.shape[0]) + + t_emb = self.time_proj(timesteps) + + # `Timesteps` does not contain any weights and will always return f32 tensors + # but time_embedding might actually be running in fp16. so we need to cast here. + # there might be better ways to encapsulate this. + t_emb = t_emb.to(dtype=sample.dtype) + + emb = self.time_embedding(t_emb, timestep_cond) + aug_emb = None + + if self.class_embedding is not None: + if class_labels is None: + raise ValueError("class_labels should be provided when num_class_embeds > 0") + + if self.config.class_embed_type == "timestep": + class_labels = self.time_proj(class_labels) + + # `Timesteps` does not contain any weights and will always return f32 tensors + # there might be better ways to encapsulate this. + class_labels = class_labels.to(dtype=sample.dtype) + + class_emb = self.class_embedding(class_labels).to(dtype=sample.dtype) + + if self.config.class_embeddings_concat: + emb = torch.cat([emb, class_emb], dim=-1) + else: + emb = emb + class_emb + + if self.config.addition_embed_type == "text": + aug_emb = self.add_embedding(encoder_hidden_states) + elif self.config.addition_embed_type == "text_image": + # Kandinsky 2.1 - style + if "image_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `addition_embed_type` set to 'text_image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`" + ) + + image_embs = added_cond_kwargs.get("image_embeds") + text_embs = added_cond_kwargs.get("text_embeds", encoder_hidden_states) + aug_emb = self.add_embedding(text_embs, image_embs) + elif self.config.addition_embed_type == "text_time": + # SDXL - style + if "text_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`" + ) + text_embeds = added_cond_kwargs.get("text_embeds") + if "time_ids" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`" + ) + time_ids = added_cond_kwargs.get("time_ids") + time_embeds = self.add_time_proj(time_ids.flatten()) + time_embeds = time_embeds.reshape((text_embeds.shape[0], -1)) + add_embeds = torch.concat([text_embeds, time_embeds], dim=-1) + add_embeds = add_embeds.to(emb.dtype) + aug_emb = self.add_embedding(add_embeds) + elif self.config.addition_embed_type == "image": + # Kandinsky 2.2 - style + if "image_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `addition_embed_type` set to 'image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`" + ) + image_embs = added_cond_kwargs.get("image_embeds") + aug_emb = self.add_embedding(image_embs) + elif self.config.addition_embed_type == "image_hint": + # Kandinsky 2.2 - style + if "image_embeds" not in added_cond_kwargs or "hint" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `addition_embed_type` set to 'image_hint' which requires the keyword arguments `image_embeds` and `hint` to be passed in `added_cond_kwargs`" + ) + image_embs = added_cond_kwargs.get("image_embeds") + hint = added_cond_kwargs.get("hint") + aug_emb, hint = self.add_embedding(image_embs, hint) + sample = torch.cat([sample, hint], dim=1) + + emb = emb + aug_emb if aug_emb is not None else emb + + if self.time_embed_act is not None: + emb = self.time_embed_act(emb) + + if self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_proj": + encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states) + elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_image_proj": + # Kadinsky 2.1 - style + if "image_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'text_image_proj' which requires the keyword argument `image_embeds` to be passed in `added_conditions`" + ) + + image_embeds = added_cond_kwargs.get("image_embeds") + encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states, image_embeds) + elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "image_proj": + # Kandinsky 2.2 - style + if "image_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'image_proj' which requires the keyword argument `image_embeds` to be passed in `added_conditions`" + ) + image_embeds = added_cond_kwargs.get("image_embeds") + encoder_hidden_states = self.encoder_hid_proj(image_embeds) + elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "ip_image_proj": + if "image_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'ip_image_proj' which requires the keyword argument `image_embeds` to be passed in `added_conditions`" + ) + image_embeds = added_cond_kwargs.get("image_embeds") + image_embeds = self.encoder_hid_proj(image_embeds) + encoder_hidden_states = (encoder_hidden_states, image_embeds) + + # 2. pre-process + sample = self.conv_in(sample) + + # 2.5 GLIGEN position net + if cross_attention_kwargs is not None and cross_attention_kwargs.get("gligen", None) is not None: + cross_attention_kwargs = cross_attention_kwargs.copy() + gligen_args = cross_attention_kwargs.pop("gligen") + cross_attention_kwargs["gligen"] = {"objs": self.position_net(**gligen_args)} + + # 3. down + lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0 + if USE_PEFT_BACKEND: + # weight the lora layers by setting `lora_scale` for each PEFT layer + scale_lora_layers(self, lora_scale) + + is_controlnet = mid_block_additional_residual is not None and down_block_additional_residuals is not None + # using new arg down_intrablock_additional_residuals for T2I-Adapters, to distinguish from controlnets + is_adapter = down_intrablock_additional_residuals is not None + # maintain backward compatibility for legacy usage, where + # T2I-Adapter and ControlNet both use down_block_additional_residuals arg + # but can only use one or the other + if not is_adapter and mid_block_additional_residual is None and down_block_additional_residuals is not None: + deprecate( + "T2I should not use down_block_additional_residuals", + "1.3.0", + "Passing intrablock residual connections with `down_block_additional_residuals` is deprecated \ + and will be removed in diffusers 1.3.0. `down_block_additional_residuals` should only be used \ + for ControlNet. Please make sure use `down_intrablock_additional_residuals` instead. ", + standard_warn=False, + ) + down_intrablock_additional_residuals = down_block_additional_residuals + is_adapter = True + + down_block_res_samples = (sample,) + for downsample_block in self.down_blocks: + if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention: + # For t2i-adapter CrossAttnDownBlock2D + additional_residuals = {} + if is_adapter and len(down_intrablock_additional_residuals) > 0: + additional_residuals["additional_residuals"] = down_intrablock_additional_residuals.pop(0) + + sample, res_samples = downsample_block( + hidden_states=sample, + temb=emb, + encoder_hidden_states=encoder_hidden_states, + attention_mask=attention_mask, + cross_attention_kwargs=cross_attention_kwargs, + encoder_attention_mask=encoder_attention_mask, + **additional_residuals, + ) + else: + sample, res_samples = downsample_block(hidden_states=sample, temb=emb, scale=lora_scale) + if is_adapter and len(down_intrablock_additional_residuals) > 0: + sample += down_intrablock_additional_residuals.pop(0) + + down_block_res_samples += res_samples + + if is_controlnet: + new_down_block_res_samples = () + + for down_block_res_sample, down_block_additional_residual in zip( + down_block_res_samples, down_block_additional_residuals + ): + down_block_res_sample = down_block_res_sample + down_block_additional_residual + new_down_block_res_samples = new_down_block_res_samples + (down_block_res_sample,) + + down_block_res_samples = new_down_block_res_samples + + # 4. mid + if self.mid_block is not None: + if hasattr(self.mid_block, "has_cross_attention") and self.mid_block.has_cross_attention: + sample = self.mid_block( + sample, + emb, + encoder_hidden_states=encoder_hidden_states, + attention_mask=attention_mask, + cross_attention_kwargs=cross_attention_kwargs, + encoder_attention_mask=encoder_attention_mask, + ) + else: + sample = self.mid_block(sample, emb) + + # To support T2I-Adapter-XL + if ( + is_adapter + and len(down_intrablock_additional_residuals) > 0 + and sample.shape == down_intrablock_additional_residuals[0].shape + ): + sample += down_intrablock_additional_residuals.pop(0) + + if is_controlnet: + sample = sample + mid_block_additional_residual + + # 5. up + for i, upsample_block in enumerate(self.up_blocks): + is_final_block = i == len(self.up_blocks) - 1 + + res_samples = down_block_res_samples[-len(upsample_block.resnets) :] + down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)] + + # if we have not reached the final block and need to forward the + # upsample size, we do it here + if not is_final_block and forward_upsample_size: + upsample_size = down_block_res_samples[-1].shape[2:] + + if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention: + sample = upsample_block( + hidden_states=sample, + temb=emb, + res_hidden_states_tuple=res_samples, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + upsample_size=upsample_size, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + ) + else: + sample = upsample_block( + hidden_states=sample, + temb=emb, + res_hidden_states_tuple=res_samples, + upsample_size=upsample_size, + scale=lora_scale, + ) + + # 6. post-process + if self.conv_norm_out: + sample = self.conv_norm_out(sample) + sample = self.conv_act(sample) + sample = self.conv_out(sample) + + if USE_PEFT_BACKEND: + # remove `lora_scale` from each PEFT layer + unscale_lora_layers(self, lora_scale) + + if not return_dict: + return (sample,) + + return UNet2DConditionOutput(sample=sample) diff --git a/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/pipeline/__init__.py b/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/pipeline/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..efd186f76b5e7023389c7d4c5f6358e44492d874 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/pipeline/__init__.py @@ -0,0 +1 @@ +from .pipeline_stable_diffusion import StableDiffusionPipeline, DiffusionPipeline \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/pipeline/pipeline_stable_diffusion.py b/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/pipeline/pipeline_stable_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..5d00932bcf2274801db295625c428aade682b856 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/pipeline/pipeline_stable_diffusion.py @@ -0,0 +1,1088 @@ +# Copyright 2023 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable, Dict, List, Optional, Union + +import torch +from packaging import version +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection + +from diffusers.configuration_utils import FrozenDict +from diffusers.image_processor import PipelineImageInput, VaeImageProcessor +from diffusers.loaders import FromSingleFileMixin, IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin +from diffusers.models import AutoencoderKL, ImageProjection, UNet2DConditionModel +from diffusers.models.attention_processor import FusedAttnProcessor2_0 +from diffusers.models.lora import adjust_lora_scale_text_encoder +from diffusers.schedulers import KarrasDiffusionSchedulers +from diffusers.utils import ( + USE_PEFT_BACKEND, + deprecate, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from diffusers.utils.torch_utils import randn_tensor +from diffusers.pipelines.pipeline_utils import DiffusionPipeline +from diffusers.pipelines.stable_diffusion.pipeline_output import StableDiffusionPipelineOutput +from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import StableDiffusionPipeline + + >>> pipe = StableDiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16) + >>> pipe = pipe.to("cuda") + + >>> prompt = "a photo of an astronaut riding a horse on mars" + >>> image = pipe(prompt).images[0] + ``` +""" + + +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + """ + Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4 + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +def retrieve_timesteps( + scheduler, + num_inference_steps: Optional[int] = None, + device: Optional[Union[str, torch.device]] = None, + timesteps: Optional[List[int]] = None, + **kwargs, +): + """ + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, + `timesteps` must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`List[int]`, *optional*): + Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default + timestep spacing strategy of the scheduler is used. If `timesteps` is passed, `num_inference_steps` + must be `None`. + + Returns: + `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class StableDiffusionPipeline( + DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin, IPAdapterMixin, FromSingleFileMixin +): + r""" + Pipeline for text-to-image generation using Stable Diffusion. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details + about a model's potential harms. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + """ + + model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae" + _optional_components = ["safety_checker", "feature_extractor", "image_encoder"] + _exclude_from_cpu_offload = ["safety_checker"] + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + image_encoder: CLIPVisionModelWithProjection = None, + requires_safety_checker: bool = True, + ): + super().__init__() + + if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`" + f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure " + "to update the config accordingly as leaving `steps_offset` might led to incorrect results" + " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub," + " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`" + " file" + ) + deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["steps_offset"] = 1 + scheduler._internal_dict = FrozenDict(new_config) + + if hasattr(scheduler.config, "clip_sample") and scheduler.config.clip_sample is True: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`." + " `clip_sample` should be set to False in the configuration file. Please make sure to update the" + " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in" + " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very" + " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file" + ) + deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["clip_sample"] = False + scheduler._internal_dict = FrozenDict(new_config) + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + is_unet_version_less_0_9_0 = hasattr(unet.config, "_diffusers_version") and version.parse( + version.parse(unet.config._diffusers_version).base_version + ) < version.parse("0.9.0.dev0") + is_unet_sample_size_less_64 = hasattr(unet.config, "sample_size") and unet.config.sample_size < 64 + if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64: + deprecation_message = ( + "The configuration file of the unet has set the default `sample_size` to smaller than" + " 64 which seems highly unlikely. If your checkpoint is a fine-tuned version of any of the" + " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-" + " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- runwayml/stable-diffusion-v1-5" + " \n- runwayml/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the" + " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`" + " in the config might lead to incorrect results in future versions. If you have downloaded this" + " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for" + " the `unet/config.json` file" + ) + deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(unet.config) + new_config["sample_size"] = 64 + unet._internal_dict = FrozenDict(new_config) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + image_encoder=image_encoder, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + def enable_vae_slicing(self): + r""" + Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to + compute decoding in several steps. This is useful to save some memory and allow larger batch sizes. + """ + self.vae.enable_slicing() + + def disable_vae_slicing(self): + r""" + Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to + computing decoding in one step. + """ + self.vae.disable_slicing() + + def enable_vae_tiling(self): + r""" + Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to + compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow + processing larger images. + """ + self.vae.enable_tiling() + + def disable_vae_tiling(self): + r""" + Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to + computing decoding in one step. + """ + self.vae.disable_tiling() + + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.FloatTensor] = None, + negative_prompt_embeds: Optional[torch.FloatTensor] = None, + lora_scale: Optional[float] = None, + **kwargs, + ): + deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple." + deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False) + + prompt_embeds_tuple = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + **kwargs, + ) + + # concatenate for backwards comp + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + + return prompt_embeds + + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.FloatTensor] = None, + negative_prompt_embeds: Optional[torch.FloatTensor] = None, + lora_scale: Optional[float] = None, + clip_skip: Optional[int] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, LoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: procecss multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: List[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: procecss multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + def prepare_ip_adapter_image_embeds(self, ip_adapter_image, device, num_images_per_prompt): + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + image_embeds = [] + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers + ): + output_hidden_state = not isinstance(image_proj_layer, ImageProjection) + single_image_embeds, single_negative_image_embeds = self.encode_image( + single_ip_adapter_image, device, 1, output_hidden_state + ) + single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0) + single_negative_image_embeds = torch.stack([single_negative_image_embeds] * num_images_per_prompt, dim=0) + + if self.do_classifier_free_guidance: + single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds]) + single_image_embeds = single_image_embeds.to(device) + + image_embeds.append(single_image_embeds) + + return image_embeds + + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + def decode_latents(self, latents): + deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead" + deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False) + + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + height, + width, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + def enable_freeu(self, s1: float, s2: float, b1: float, b2: float): + r"""Enables the FreeU mechanism as in https://arxiv.org/abs/2309.11497. + + The suffixes after the scaling factors represent the stages where they are being applied. + + Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of the values + that are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL. + + Args: + s1 (`float`): + Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to + mitigate "oversmoothing effect" in the enhanced denoising process. + s2 (`float`): + Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to + mitigate "oversmoothing effect" in the enhanced denoising process. + b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features. + b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features. + """ + if not hasattr(self, "unet"): + raise ValueError("The pipeline must have `unet` for using FreeU.") + self.unet.enable_freeu(s1=s1, s2=s2, b1=b1, b2=b2) + + def disable_freeu(self): + """Disables the FreeU mechanism if enabled.""" + self.unet.disable_freeu() + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.fuse_qkv_projections + def fuse_qkv_projections(self, unet: bool = True, vae: bool = True): + """ + Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, + key, value) are fused. For cross-attention modules, key and value projection matrices are fused. + + + + This API is 🧪 experimental. + + + + Args: + unet (`bool`, defaults to `True`): To apply fusion on the UNet. + vae (`bool`, defaults to `True`): To apply fusion on the VAE. + """ + self.fusing_unet = False + self.fusing_vae = False + + if unet: + self.fusing_unet = True + self.unet.fuse_qkv_projections() + self.unet.set_attn_processor(FusedAttnProcessor2_0()) + + if vae: + if not isinstance(self.vae, AutoencoderKL): + raise ValueError("`fuse_qkv_projections()` is only supported for the VAE of type `AutoencoderKL`.") + + self.fusing_vae = True + self.vae.fuse_qkv_projections() + self.vae.set_attn_processor(FusedAttnProcessor2_0()) + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.unfuse_qkv_projections + def unfuse_qkv_projections(self, unet: bool = True, vae: bool = True): + """Disable QKV projection fusion if enabled. + + + + This API is 🧪 experimental. + + + + Args: + unet (`bool`, defaults to `True`): To apply fusion on the UNet. + vae (`bool`, defaults to `True`): To apply fusion on the VAE. + + """ + if unet: + if not self.fusing_unet: + logger.warning("The UNet was not initially fused for QKV projections. Doing nothing.") + else: + self.unet.unfuse_qkv_projections() + self.fusing_unet = False + + if vae: + if not self.fusing_vae: + logger.warning("The VAE was not initially fused for QKV projections. Doing nothing.") + else: + self.vae.unfuse_qkv_projections() + self.fusing_vae = False + + # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding + def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32): + """ + See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 + + Args: + timesteps (`torch.Tensor`): + generate embedding vectors at these timesteps + embedding_dim (`int`, *optional*, defaults to 512): + dimension of the embeddings to generate + dtype: + data type of the generated embeddings + + Returns: + `torch.FloatTensor`: Embedding vectors with shape `(len(timesteps), embedding_dim)` + """ + assert len(w.shape) == 1 + w = w * 1000.0 + + half_dim = embedding_dim // 2 + emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) + emb = w.to(dtype)[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0, 1)) + assert emb.shape == (w.shape[0], embedding_dim) + return emb + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def guidance_rescale(self): + return self._guidance_rescale + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None + + @property + def cross_attention_kwargs(self): + return self._cross_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Union[str, List[str]] = None, + height: Optional[int] = None, + width: Optional[int] = None, + num_inference_steps: int = 50, + timesteps: List[int] = None, + guidance_scale: float = 7.5, + negative_prompt: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.FloatTensor] = None, + prompt_embeds: Optional[torch.FloatTensor] = None, + negative_prompt_embeds: Optional[torch.FloatTensor] = None, + ip_adapter_image: Optional[PipelineImageInput] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + guidance_rescale: float = 0.0, + clip_skip: Optional[int] = None, + callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None, + callback_on_step_end_tensor_inputs: List[str] = ["latents"], + **kwargs, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`List[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies + to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + guidance_rescale (`float`, *optional*, defaults to 0.0): + Guidance rescale factor from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://arxiv.org/pdf/2305.08891.pdf). Guidance rescale factor should fix overexposure when + using zero terminal SNR. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + """ + + callback = kwargs.pop("callback", None) + callback_steps = kwargs.pop("callback_steps", None) + + if callback is not None: + deprecate( + "callback", + "1.0.0", + "Passing `callback` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`", + ) + if callback_steps is not None: + deprecate( + "callback_steps", + "1.0.0", + "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`", + ) + + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + # to deal with lora scaling and other possible forward hooks + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + height, + width, + callback_steps, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + callback_on_step_end_tensor_inputs, + ) + + self._guidance_scale = guidance_scale + self._guidance_rescale = guidance_rescale + self._clip_skip = clip_skip + self._cross_attention_kwargs = cross_attention_kwargs + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Encode input prompt + lora_scale = ( + self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None + ) + + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + self.do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + clip_skip=self.clip_skip, + ) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + if ip_adapter_image is not None: + image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, device, batch_size * num_images_per_prompt + ) + + # 4. Prepare timesteps + timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps) + + # 5. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 6.1 Add image embeds for IP-Adapter + added_cond_kwargs = {"image_embeds": image_embeds} if ip_adapter_image is not None else None + + # 6.2 Optionally get Guidance Scale Embedding + timestep_cond = None + if self.unet.config.time_cond_proj_dim is not None: + guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt) + timestep_cond = self.get_guidance_scale_embedding( + guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim + ).to(device=device, dtype=latents.dtype) + + # 7. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + timestep_cond=timestep_cond, + cross_attention_kwargs=self.cross_attention_kwargs, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + + if self.do_classifier_free_guidance and self.guidance_rescale > 0.0: + # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf + noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, generator=generator)[ + 0 + ] + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + else: + image = latents + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/schedulers/__init__.py b/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/schedulers/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..6f853e7d075eb2528f2ef9e4e230a082af79fb80 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/schedulers/__init__.py @@ -0,0 +1 @@ +from .scheduling_pndm import PNDMScheduler, SchedulerMixin \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/schedulers/scheduling_pndm.py b/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/schedulers/scheduling_pndm.py new file mode 100644 index 0000000000000000000000000000000000000000..7d50d01e9602ee2bfa4b612baadcfdc8128deff1 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/schedulers/scheduling_pndm.py @@ -0,0 +1,478 @@ +# Copyright 2023 Zhejiang University Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# DISCLAIMER: This file is strongly influenced by https://github.com/ermongroup/ddim + +import math +from typing import List, Optional, Tuple, Union + +import numpy as np +import torch + +from diffusers.configuration_utils import ConfigMixin, register_to_config +from diffusers.schedulers.scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin, SchedulerOutput + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps, + max_beta=0.999, + alpha_transform_type="cosine", +): + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + + Args: + num_diffusion_timesteps (`int`): the number of betas to produce. + max_beta (`float`): the maximum beta to use; use values lower than 1 to + prevent singularities. + alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar. + Choose from `cosine` or `exp` + + Returns: + betas (`np.ndarray`): the betas used by the scheduler to step the model outputs + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_tranform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +class PNDMScheduler(SchedulerMixin, ConfigMixin): + """ + `PNDMScheduler` uses pseudo numerical methods for diffusion models such as the Runge-Kutta and linear multi-step + method. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to 0.0001): + The starting `beta` value of inference. + beta_end (`float`, defaults to 0.02): + The final `beta` value. + beta_schedule (`str`, defaults to `"linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear`, `scaled_linear`, or `squaredcos_cap_v2`. + trained_betas (`np.ndarray`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + skip_prk_steps (`bool`, defaults to `False`): + Allows the scheduler to skip the Runge-Kutta steps defined in the original paper as being required before + PLMS steps. + set_alpha_to_one (`bool`, defaults to `False`): + Each diffusion step uses the alphas product value at that step and at the previous one. For the final step + there is no previous alpha. When this option is `True` the previous alpha product is fixed to `1`, + otherwise it uses the alpha value at step 0. + prediction_type (`str`, defaults to `epsilon`, *optional*): + Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process) + or `v_prediction` (see section 2.4 of [Imagen Video](https://imagen.research.google/video/paper.pdf) + paper). + timestep_spacing (`str`, defaults to `"leading"`): + The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + steps_offset (`int`, defaults to 0): + An offset added to the inference steps. You can use a combination of `offset=1` and + `set_alpha_to_one=False` to make the last step use step 0 for the previous alpha product like in Stable + Diffusion. + """ + + _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.0001, + beta_end: float = 0.02, + beta_schedule: str = "linear", + trained_betas: Optional[Union[np.ndarray, List[float]]] = None, + skip_prk_steps: bool = False, + set_alpha_to_one: bool = False, + prediction_type: str = "epsilon", + timestep_spacing: str = "leading", + steps_offset: int = 0, + ): + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2 + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + else: + raise NotImplementedError(f"{beta_schedule} does is not implemented for {self.__class__}") + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + + self.final_alpha_cumprod = torch.tensor(1.0) if set_alpha_to_one else self.alphas_cumprod[0] + + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + # For now we only support F-PNDM, i.e. the runge-kutta method + # For more information on the algorithm please take a look at the paper: https://arxiv.org/pdf/2202.09778.pdf + # mainly at formula (9), (12), (13) and the Algorithm 2. + self.pndm_order = 4 + + # running values + self.cur_model_output = 0 + self.counter = 0 + self.cur_sample = None + self.ets = [] + + # setable values + self.num_inference_steps = None + self._timesteps = np.arange(0, num_train_timesteps)[::-1].copy() + self.prk_timesteps = None + self.plms_timesteps = None + self.timesteps = None + + def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + """ + + self.num_inference_steps = num_inference_steps + # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://arxiv.org/abs/2305.08891 + if self.config.timestep_spacing == "linspace": + self._timesteps = ( + np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps).round().astype(np.int64) + ) + elif self.config.timestep_spacing == "leading": + step_ratio = self.config.num_train_timesteps // self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + self._timesteps = (np.arange(0, num_inference_steps) * step_ratio).round() + self._timesteps += self.config.steps_offset + elif self.config.timestep_spacing == "trailing": + step_ratio = self.config.num_train_timesteps / self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + self._timesteps = np.round(np.arange(self.config.num_train_timesteps, 0, -step_ratio))[::-1].astype( + np.int64 + ) + self._timesteps -= 1 + else: + raise ValueError( + f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'." + ) + + if self.config.skip_prk_steps: + # for some models like stable diffusion the prk steps can/should be skipped to + # produce better results. When using PNDM with `self.config.skip_prk_steps` the implementation + # is based on crowsonkb's PLMS sampler implementation: https://github.com/CompVis/latent-diffusion/pull/51 + self.prk_timesteps = np.array([]) + self.plms_timesteps = np.concatenate([self._timesteps[:-1], self._timesteps[-2:-1], self._timesteps[-1:]])[ + ::-1 + ].copy() + else: + prk_timesteps = np.array(self._timesteps[-self.pndm_order :]).repeat(2) + np.tile( + np.array([0, self.config.num_train_timesteps // num_inference_steps // 2]), self.pndm_order + ) + self.prk_timesteps = (prk_timesteps[:-1].repeat(2)[1:-1])[::-1].copy() + self.plms_timesteps = self._timesteps[:-3][ + ::-1 + ].copy() # we copy to avoid having negative strides which are not supported by torch.from_numpy + + timesteps = np.concatenate([self.prk_timesteps, self.plms_timesteps]).astype(np.int64) + self.timesteps = torch.from_numpy(timesteps).to(device) + + self.ets = [] + self.counter = 0 + self.cur_model_output = 0 + + def step( + self, + model_output: torch.FloatTensor, + timestep: int, + sample: torch.FloatTensor, + return_dict: bool = True, + ) -> Union[SchedulerOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise), and calls [`~PNDMScheduler.step_prk`] + or [`~PNDMScheduler.step_plms`] depending on the internal variable `counter`. + + Args: + model_output (`torch.FloatTensor`): + The direct output from learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + sample (`torch.FloatTensor`): + A current instance of a sample created by the diffusion process. + return_dict (`bool`): + Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`. + + Returns: + [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + + """ + if self.counter < len(self.prk_timesteps) and not self.config.skip_prk_steps: + return self.step_prk(model_output=model_output, timestep=timestep, sample=sample, return_dict=return_dict) + else: + return self.step_plms(model_output=model_output, timestep=timestep, sample=sample, return_dict=return_dict) + + def step_prk( + self, + model_output: torch.FloatTensor, + timestep: int, + sample: torch.FloatTensor, + return_dict: bool = True, + ) -> Union[SchedulerOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with + the Runge-Kutta method. It performs four forward passes to approximate the solution to the differential + equation. + + Args: + model_output (`torch.FloatTensor`): + The direct output from learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + sample (`torch.FloatTensor`): + A current instance of a sample created by the diffusion process. + return_dict (`bool`): + Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or tuple. + + Returns: + [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + + """ + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + diff_to_prev = 0 if self.counter % 2 else self.config.num_train_timesteps // self.num_inference_steps // 2 + prev_timestep = timestep - diff_to_prev + timestep = self.prk_timesteps[self.counter // 4 * 4] + + if self.counter % 4 == 0: + self.cur_model_output += 1 / 6 * model_output + self.ets.append(model_output) + self.cur_sample = sample + elif (self.counter - 1) % 4 == 0: + self.cur_model_output += 1 / 3 * model_output + elif (self.counter - 2) % 4 == 0: + self.cur_model_output += 1 / 3 * model_output + elif (self.counter - 3) % 4 == 0: + model_output = self.cur_model_output + 1 / 6 * model_output + self.cur_model_output = 0 + + # cur_sample should not be `None` + cur_sample = self.cur_sample if self.cur_sample is not None else sample + + prev_sample = self._get_prev_sample(cur_sample, timestep, prev_timestep, model_output) + self.counter += 1 + + if not return_dict: + return (prev_sample,) + + return SchedulerOutput(prev_sample=prev_sample) + + def step_plms( + self, + model_output: torch.FloatTensor, + timestep: int, + sample: torch.FloatTensor, + return_dict: bool = True, + ) -> Union[SchedulerOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with + the linear multistep method. It performs one forward pass multiple times to approximate the solution. + + Args: + model_output (`torch.FloatTensor`): + The direct output from learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + sample (`torch.FloatTensor`): + A current instance of a sample created by the diffusion process. + return_dict (`bool`): + Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or tuple. + + Returns: + [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + + """ + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + if not self.config.skip_prk_steps and len(self.ets) < 3: + raise ValueError( + f"{self.__class__} can only be run AFTER scheduler has been run " + "in 'prk' mode for at least 12 iterations " + "See: https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/pipeline_pndm.py " + "for more information." + ) + + prev_timestep = timestep - self.config.num_train_timesteps // self.num_inference_steps + + if self.counter != 1: + self.ets = self.ets[-3:] + self.ets.append(model_output) + else: + prev_timestep = timestep + timestep = timestep + self.config.num_train_timesteps // self.num_inference_steps + + if len(self.ets) == 1 and self.counter == 0: + model_output = model_output + self.cur_sample = sample + elif len(self.ets) == 1 and self.counter == 1: + model_output = (model_output + self.ets[-1]) / 2 + sample = self.cur_sample + self.cur_sample = None + elif len(self.ets) == 2: + model_output = (3 * self.ets[-1] - self.ets[-2]) / 2 + elif len(self.ets) == 3: + model_output = (23 * self.ets[-1] - 16 * self.ets[-2] + 5 * self.ets[-3]) / 12 + else: + model_output = (1 / 24) * (55 * self.ets[-1] - 59 * self.ets[-2] + 37 * self.ets[-3] - 9 * self.ets[-4]) + + prev_sample = self._get_prev_sample(sample, timestep, prev_timestep, model_output) + self.counter += 1 + + if not return_dict: + return (prev_sample,) + + return SchedulerOutput(prev_sample=prev_sample) + + def scale_model_input(self, sample: torch.FloatTensor, *args, **kwargs) -> torch.FloatTensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.FloatTensor`): + The input sample. + + Returns: + `torch.FloatTensor`: + A scaled input sample. + """ + return sample + + def _get_prev_sample(self, sample, timestep, prev_timestep, model_output): + # See formula (9) of PNDM paper https://arxiv.org/pdf/2202.09778.pdf + # this function computes x_(t−δ) using the formula of (9) + # Note that x_t needs to be added to both sides of the equation + + # Notation ( -> + # alpha_prod_t -> α_t + # alpha_prod_t_prev -> α_(t−δ) + # beta_prod_t -> (1 - α_t) + # beta_prod_t_prev -> (1 - α_(t−δ)) + # sample -> x_t + # model_output -> e_θ(x_t, t) + # prev_sample -> x_(t−δ) + alpha_prod_t = self.alphas_cumprod[timestep] + alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod + beta_prod_t = 1 - alpha_prod_t + beta_prod_t_prev = 1 - alpha_prod_t_prev + + if self.config.prediction_type == "v_prediction": + model_output = (alpha_prod_t**0.5) * model_output + (beta_prod_t**0.5) * sample + elif self.config.prediction_type != "epsilon": + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon` or `v_prediction`" + ) + + # corresponds to (α_(t−δ) - α_t) divided by + # denominator of x_t in formula (9) and plus 1 + # Note: (α_(t−δ) - α_t) / (sqrt(α_t) * (sqrt(α_(t−δ)) + sqr(α_t))) = + # sqrt(α_(t−δ)) / sqrt(α_t)) + sample_coeff = (alpha_prod_t_prev / alpha_prod_t) ** (0.5) + + # corresponds to denominator of e_θ(x_t, t) in formula (9) + model_output_denom_coeff = alpha_prod_t * beta_prod_t_prev ** (0.5) + ( + alpha_prod_t * beta_prod_t * alpha_prod_t_prev + ) ** (0.5) + + # full formula (9) + prev_sample = ( + sample_coeff * sample - (alpha_prod_t_prev - alpha_prod_t) * model_output / model_output_denom_coeff + ) + + return prev_sample + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.add_noise + def add_noise( + self, + original_samples: torch.FloatTensor, + noise: torch.FloatTensor, + timesteps: torch.IntTensor, + ) -> torch.FloatTensor: + # Make sure alphas_cumprod and timestep have same device and dtype as original_samples + # Move the self.alphas_cumprod to device to avoid redundant CPU to GPU data movement + # for the subsequent add_noise calls + self.alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device) + alphas_cumprod = self.alphas_cumprod.to(dtype=original_samples.dtype) + timesteps = timesteps.to(original_samples.device) + + sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(original_samples.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise + return noisy_samples + + def __len__(self): + return self.config.num_train_timesteps diff --git a/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/vae/__init__.py b/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/vae/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..58bbb8f14ee8e394705159b0ebf877de963c143a --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/vae/__init__.py @@ -0,0 +1 @@ +from .autoencoder_kl import AutoencoderKL \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/vae/autoencoder_kl.py b/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/vae/autoencoder_kl.py new file mode 100644 index 0000000000000000000000000000000000000000..0a9d02770b66dd1874224d14fad5f6d46f7f0370 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-2.1/stablediffusion/vae/autoencoder_kl.py @@ -0,0 +1,488 @@ +# Copyright 2023 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from typing import Dict, Optional, Tuple, Union + +import torch +import torch.nn as nn + +from diffusers.configuration_utils import ConfigMixin, register_to_config +from diffusers.loaders import FromOriginalVAEMixin +from diffusers.utils.accelerate_utils import apply_forward_hook +from ..layers.attention_processor import ( + ADDED_KV_ATTENTION_PROCESSORS, + CROSS_ATTENTION_PROCESSORS, + Attention, + AttentionProcessor, + AttnAddedKVProcessor, + AttnProcessor, +) +from diffusers.models.modeling_outputs import AutoencoderKLOutput +from diffusers.models.autoencoders.vae import Decoder, DecoderOutput, DiagonalGaussianDistribution, Encoder + +from stablediffusion.models.model_utils import ModelMixin + + +class AutoencoderKL(ModelMixin, ConfigMixin, FromOriginalVAEMixin): + r""" + A VAE model with KL loss for encoding images into latents and decoding latent representations into images. + + This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented + for all models (such as downloading or saving). + + Parameters: + in_channels (int, *optional*, defaults to 3): Number of channels in the input image. + out_channels (int, *optional*, defaults to 3): Number of channels in the output. + down_block_types (`Tuple[str]`, *optional*, defaults to `("DownEncoderBlock2D",)`): + Tuple of downsample block types. + up_block_types (`Tuple[str]`, *optional*, defaults to `("UpDecoderBlock2D",)`): + Tuple of upsample block types. + block_out_channels (`Tuple[int]`, *optional*, defaults to `(64,)`): + Tuple of block output channels. + act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use. + latent_channels (`int`, *optional*, defaults to 4): Number of channels in the latent space. + sample_size (`int`, *optional*, defaults to `32`): Sample input size. + scaling_factor (`float`, *optional*, defaults to 0.18215): + The component-wise standard deviation of the trained latent space computed using the first batch of the + training set. This is used to scale the latent space to have unit variance when training the diffusion + model. The latents are scaled with the formula `z = z * scaling_factor` before being passed to the + diffusion model. When decoding, the latents are scaled back to the original scale with the formula: `z = 1 + / scaling_factor * z`. For more details, refer to sections 4.3.2 and D.1 of the [High-Resolution Image + Synthesis with Latent Diffusion Models](https://arxiv.org/abs/2112.10752) paper. + force_upcast (`bool`, *optional*, default to `True`): + If enabled it will force the VAE to run in float32 for high image resolution pipelines, such as SD-XL. VAE + can be fine-tuned / trained to a lower range without loosing too much precision in which case + `force_upcast` can be set to `False` - see: https://huggingface.co/madebyollin/sdxl-vae-fp16-fix + """ + + _supports_gradient_checkpointing = True + + @register_to_config + def __init__( + self, + in_channels: int = 3, + out_channels: int = 3, + down_block_types: Tuple[str] = ("DownEncoderBlock2D",), + up_block_types: Tuple[str] = ("UpDecoderBlock2D",), + block_out_channels: Tuple[int] = (64,), + layers_per_block: int = 1, + act_fn: str = "silu", + latent_channels: int = 4, + norm_num_groups: int = 32, + sample_size: int = 32, + scaling_factor: float = 0.18215, + force_upcast: float = True, + ): + super().__init__() + + # pass init params to Encoder + self.encoder = Encoder( + in_channels=in_channels, + out_channels=latent_channels, + down_block_types=down_block_types, + block_out_channels=block_out_channels, + layers_per_block=layers_per_block, + act_fn=act_fn, + norm_num_groups=norm_num_groups, + double_z=True, + ) + + # pass init params to Decoder + self.decoder = Decoder( + in_channels=latent_channels, + out_channels=out_channels, + up_block_types=up_block_types, + block_out_channels=block_out_channels, + layers_per_block=layers_per_block, + norm_num_groups=norm_num_groups, + act_fn=act_fn, + ) + + self.quant_conv = nn.Conv2d(2 * latent_channels, 2 * latent_channels, 1) + self.post_quant_conv = nn.Conv2d(latent_channels, latent_channels, 1) + + self.use_slicing = False + self.use_tiling = False + + # only relevant if vae tiling is enabled + self.tile_sample_min_size = self.config.sample_size + sample_size = ( + self.config.sample_size[0] + if isinstance(self.config.sample_size, (list, tuple)) + else self.config.sample_size + ) + self.tile_latent_min_size = int(sample_size / (2 ** (len(self.config.block_out_channels) - 1))) + self.tile_overlap_factor = 0.25 + + def _set_gradient_checkpointing(self, module, value=False): + if isinstance(module, (Encoder, Decoder)): + module.gradient_checkpointing = value + + def enable_tiling(self, use_tiling: bool = True): + r""" + Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to + compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow + processing larger images. + """ + self.use_tiling = use_tiling + + def disable_tiling(self): + r""" + Disable tiled VAE decoding. If `enable_tiling` was previously enabled, this method will go back to computing + decoding in one step. + """ + self.enable_tiling(False) + + def enable_slicing(self): + r""" + Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to + compute decoding in several steps. This is useful to save some memory and allow larger batch sizes. + """ + self.use_slicing = True + + def disable_slicing(self): + r""" + Disable sliced VAE decoding. If `enable_slicing` was previously enabled, this method will go back to computing + decoding in one step. + """ + self.use_slicing = False + + @property + # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors + def attn_processors(self) -> Dict[str, AttentionProcessor]: + r""" + Returns: + `dict` of attention processors: A dictionary containing all attention processors used in the model with + indexed by its weight name. + """ + # set recursively + processors = {} + + def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]): + if hasattr(module, "get_processor"): + processors[f"{name}.processor"] = module.get_processor(return_deprecated_lora=True) + + for sub_name, child in module.named_children(): + fn_recursive_add_processors(f"{name}.{sub_name}", child, processors) + + return processors + + for name, module in self.named_children(): + fn_recursive_add_processors(name, module, processors) + + return processors + + # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor + def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]): + r""" + Sets the attention processor to use to compute attention. + + Parameters: + processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`): + The instantiated processor class or a dictionary of processor classes that will be set as the processor + for **all** `Attention` layers. + + If `processor` is a dict, the key needs to define the path to the corresponding cross attention + processor. This is strongly recommended when setting trainable attention processors. + + """ + count = len(self.attn_processors.keys()) + + if isinstance(processor, dict) and len(processor) != count: + raise ValueError( + f"A dict of processors was passed, but the number of processors {len(processor)} does not match the" + f" number of attention layers: {count}. Please make sure to pass {count} processor classes." + ) + + def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor): + if hasattr(module, "set_processor"): + if not isinstance(processor, dict): + module.set_processor(processor) + else: + module.set_processor(processor.pop(f"{name}.processor")) + + for sub_name, child in module.named_children(): + fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor) + + for name, module in self.named_children(): + fn_recursive_attn_processor(name, module, processor) + + # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor + def set_default_attn_processor(self): + """ + Disables custom attention processors and sets the default attention implementation. + """ + if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()): + processor = AttnAddedKVProcessor() + elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()): + processor = AttnProcessor() + else: + raise ValueError( + f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}" + ) + + self.set_attn_processor(processor) + + @apply_forward_hook + def encode( + self, x: torch.FloatTensor, return_dict: bool = True + ) -> Union[AutoencoderKLOutput, Tuple[DiagonalGaussianDistribution]]: + """ + Encode a batch of images into latents. + + Args: + x (`torch.FloatTensor`): Input batch of images. + return_dict (`bool`, *optional*, defaults to `True`): + Whether to return a [`~models.autoencoder_kl.AutoencoderKLOutput`] instead of a plain tuple. + + Returns: + The latent representations of the encoded images. If `return_dict` is True, a + [`~models.autoencoder_kl.AutoencoderKLOutput`] is returned, otherwise a plain `tuple` is returned. + """ + if self.use_tiling and (x.shape[-1] > self.tile_sample_min_size or x.shape[-2] > self.tile_sample_min_size): + return self.tiled_encode(x, return_dict=return_dict) + + if self.use_slicing and x.shape[0] > 1: + encoded_slices = [self.encoder(x_slice) for x_slice in x.split(1)] + h = torch.cat(encoded_slices) + else: + h = self.encoder(x) + + moments = self.quant_conv(h) + posterior = DiagonalGaussianDistribution(moments) + + if not return_dict: + return (posterior,) + + return AutoencoderKLOutput(latent_dist=posterior) + + def _decode(self, z: torch.FloatTensor, return_dict: bool = True) -> Union[DecoderOutput, torch.FloatTensor]: + if self.use_tiling and (z.shape[-1] > self.tile_latent_min_size or z.shape[-2] > self.tile_latent_min_size): + return self.tiled_decode(z, return_dict=return_dict) + + z = self.post_quant_conv(z) + dec = self.decoder(z) + + if not return_dict: + return (dec,) + + return DecoderOutput(sample=dec) + + @apply_forward_hook + def decode( + self, z: torch.FloatTensor, return_dict: bool = True, generator=None + ) -> Union[DecoderOutput, torch.FloatTensor]: + """ + Decode a batch of images. + + Args: + z (`torch.FloatTensor`): Input batch of latent vectors. + return_dict (`bool`, *optional*, defaults to `True`): + Whether to return a [`~models.vae.DecoderOutput`] instead of a plain tuple. + + Returns: + [`~models.vae.DecoderOutput`] or `tuple`: + If return_dict is True, a [`~models.vae.DecoderOutput`] is returned, otherwise a plain `tuple` is + returned. + + """ + if self.use_slicing and z.shape[0] > 1: + decoded_slices = [self._decode(z_slice).sample for z_slice in z.split(1)] + decoded = torch.cat(decoded_slices) + else: + decoded = self._decode(z).sample + + if not return_dict: + return (decoded,) + + return DecoderOutput(sample=decoded) + + def blend_v(self, a: torch.Tensor, b: torch.Tensor, blend_extent: int) -> torch.Tensor: + blend_extent = min(a.shape[2], b.shape[2], blend_extent) + for y in range(blend_extent): + b[:, :, y, :] = a[:, :, -blend_extent + y, :] * (1 - y / blend_extent) + b[:, :, y, :] * (y / blend_extent) + return b + + def blend_h(self, a: torch.Tensor, b: torch.Tensor, blend_extent: int) -> torch.Tensor: + blend_extent = min(a.shape[3], b.shape[3], blend_extent) + for x in range(blend_extent): + b[:, :, :, x] = a[:, :, :, -blend_extent + x] * (1 - x / blend_extent) + b[:, :, :, x] * (x / blend_extent) + return b + + def tiled_encode(self, x: torch.FloatTensor, return_dict: bool = True) -> AutoencoderKLOutput: + r"""Encode a batch of images using a tiled encoder. + + When this option is enabled, the VAE will split the input tensor into tiles to compute encoding in several + steps. This is useful to keep memory use constant regardless of image size. The end result of tiled encoding is + different from non-tiled encoding because each tile uses a different encoder. To avoid tiling artifacts, the + tiles overlap and are blended together to form a smooth output. You may still see tile-sized changes in the + output, but they should be much less noticeable. + + Args: + x (`torch.FloatTensor`): Input batch of images. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~models.autoencoder_kl.AutoencoderKLOutput`] instead of a plain tuple. + + Returns: + [`~models.autoencoder_kl.AutoencoderKLOutput`] or `tuple`: + If return_dict is True, a [`~models.autoencoder_kl.AutoencoderKLOutput`] is returned, otherwise a plain + `tuple` is returned. + """ + overlap_size = int(self.tile_sample_min_size * (1 - self.tile_overlap_factor)) + blend_extent = int(self.tile_latent_min_size * self.tile_overlap_factor) + row_limit = self.tile_latent_min_size - blend_extent + + # Split the image into 512x512 tiles and encode them separately. + rows = [] + for i in range(0, x.shape[2], overlap_size): + row = [] + for j in range(0, x.shape[3], overlap_size): + tile = x[:, :, i : i + self.tile_sample_min_size, j : j + self.tile_sample_min_size] + tile = self.encoder(tile) + tile = self.quant_conv(tile) + row.append(tile) + rows.append(row) + result_rows = [] + for i, row in enumerate(rows): + result_row = [] + for j, tile in enumerate(row): + # blend the above tile and the left tile + # to the current tile and add the current tile to the result row + if i > 0: + tile = self.blend_v(rows[i - 1][j], tile, blend_extent) + if j > 0: + tile = self.blend_h(row[j - 1], tile, blend_extent) + result_row.append(tile[:, :, :row_limit, :row_limit]) + result_rows.append(torch.cat(result_row, dim=3)) + + moments = torch.cat(result_rows, dim=2) + posterior = DiagonalGaussianDistribution(moments) + + if not return_dict: + return (posterior,) + + return AutoencoderKLOutput(latent_dist=posterior) + + def tiled_decode(self, z: torch.FloatTensor, return_dict: bool = True) -> Union[DecoderOutput, torch.FloatTensor]: + r""" + Decode a batch of images using a tiled decoder. + + Args: + z (`torch.FloatTensor`): Input batch of latent vectors. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~models.vae.DecoderOutput`] instead of a plain tuple. + + Returns: + [`~models.vae.DecoderOutput`] or `tuple`: + If return_dict is True, a [`~models.vae.DecoderOutput`] is returned, otherwise a plain `tuple` is + returned. + """ + overlap_size = int(self.tile_latent_min_size * (1 - self.tile_overlap_factor)) + blend_extent = int(self.tile_sample_min_size * self.tile_overlap_factor) + row_limit = self.tile_sample_min_size - blend_extent + + # Split z into overlapping 64x64 tiles and decode them separately. + # The tiles have an overlap to avoid seams between tiles. + rows = [] + for i in range(0, z.shape[2], overlap_size): + row = [] + for j in range(0, z.shape[3], overlap_size): + tile = z[:, :, i : i + self.tile_latent_min_size, j : j + self.tile_latent_min_size] + tile = self.post_quant_conv(tile) + decoded = self.decoder(tile) + row.append(decoded) + rows.append(row) + result_rows = [] + for i, row in enumerate(rows): + result_row = [] + for j, tile in enumerate(row): + # blend the above tile and the left tile + # to the current tile and add the current tile to the result row + if i > 0: + tile = self.blend_v(rows[i - 1][j], tile, blend_extent) + if j > 0: + tile = self.blend_h(row[j - 1], tile, blend_extent) + result_row.append(tile[:, :, :row_limit, :row_limit]) + result_rows.append(torch.cat(result_row, dim=3)) + + dec = torch.cat(result_rows, dim=2) + if not return_dict: + return (dec,) + + return DecoderOutput(sample=dec) + + def forward( + self, + sample: torch.FloatTensor, + sample_posterior: bool = False, + return_dict: bool = True, + generator: Optional[torch.Generator] = None, + ) -> Union[DecoderOutput, torch.FloatTensor]: + r""" + Args: + sample (`torch.FloatTensor`): Input sample. + sample_posterior (`bool`, *optional*, defaults to `False`): + Whether to sample from the posterior. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`DecoderOutput`] instead of a plain tuple. + """ + x = sample + posterior = self.encode(x).latent_dist + if sample_posterior: + z = posterior.sample(generator=generator) + else: + z = posterior.mode() + dec = self.decode(z).sample + + if not return_dict: + return (dec,) + + return DecoderOutput(sample=dec) + + # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.fuse_qkv_projections + def fuse_qkv_projections(self): + """ + Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, + key, value) are fused. For cross-attention modules, key and value projection matrices are fused. + + + + This API is 🧪 experimental. + + + """ + self.original_attn_processors = None + + for _, attn_processor in self.attn_processors.items(): + if "Added" in str(attn_processor.__class__.__name__): + raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.") + + self.original_attn_processors = self.attn_processors + + for module in self.modules(): + if isinstance(module, Attention): + module.fuse_projections(fuse=True) + + # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.unfuse_qkv_projections + def unfuse_qkv_projections(self): + """Disables the fused QKV projection if enabled. + + + + This API is 🧪 experimental. + + + + """ + if self.original_attn_processors is not None: + self.set_attn_processor(self.original_attn_processors) diff --git a/MindIE/MultiModal/StableDiffusion-3/README.md b/MindIE/MultiModal/StableDiffusion-3/README.md new file mode 100644 index 0000000000000000000000000000000000000000..26ebb8ec1e856af4315242a64096fa6d234f1cc3 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-3/README.md @@ -0,0 +1,254 @@ +--- +license: apache-2.0 +--- + + + +# Stable diffusion 3 + +## 一、介绍 +此仓库是开源模型Stable diffusion 3， 基于MindIE SD 的实现。 可以实现更高效的推理性能。运行此仓库代码，需要安装MindIE SD 及其依赖。 + + +## 二、安装依赖 + +MindIE SD是MindIE的视图生成推理模型套件，其目标是为稳定扩散（Stable Diffusion, SD）系列大模型推理任务提供在昇腾硬件及其软件栈上的端到端解决方案，软件系统内部集成各功能模块，对外呈现统一的编程接口。 + +MindIE-SD其依赖组件为driver驱动包、firmware固件包、CANN开发套件包、推理引擎MindIE包，使用MindIE-SD前请提前安装这些依赖。 + +| 简称 | 安装包全名 | 默认安装路径 | 版本约束 | +| --------------- |---------------------------------------------------------------------------|--------------------------------------|-----------------------------------| +| driver驱动包 | 昇腾310P处理器对应驱动软件包：Ascend-hdk-310p-npu-driver_\{version\}\_{os}\-{arch}.run | /usr/local/Ascend | 24.0.rc1及以上 | +| firmware固件包 | 昇腾310P处理器对应固件软件包：Ascend-hdk-310p-npu-firmware_\{version\}.run | /usr/local/Ascend | 24.0.rc1及以上 | +| CANN开发套件包 | Ascend-cann-toolkit\_{version}_linux-{arch}.run | /usr/local/Ascend/ascend-toolkit/latest | 8.0.RC1及以上 | +| 推理引擎MindIE包 | Ascend-mindie\_\{version}_linux-\{arch}.run | /usr/local/Ascend/mindie/latest | 和mindietorch严格配套使用 | +| torch | Python的whl包：torch-{version}-cp310-cp310-{os}_{arch}.whl | - | Python版本3.10.x，torch版本支持2.1.0 | + +- {version}为软件包版本 +- {os}为系统名称，如Linux +- {arch}为架构名称，如x86_64 + +### 2.1 安装驱动和固件 + +1. 获取地址 +- [Atlas 800I A2](https://www.hiascend.com/hardware/firmware-drivers/community?product=4&model=32&cann=8.0.RC1.beta1&driver=1.0.RC1.alpha) +2. [安装指导手册](https://www.hiascend.com/document/detail/zh/CANNCommunityEdition/80RC2alpha002/softwareinst/instg/instg_0019.html) +### 2.2 CANN开发套件包+kernel包+MindIE包下载 +1. 下载： +- [Atlas 800I A2](https://www.hiascend.com/developer/download/community/result?module=pt+ie+cann&product=4&model=32) +2. [环境准备指导](https://www.hiascend.com/document/detail/zh/CANNCommunityEdition/80RC2alpha002/softwareinst/instg/instg_0001.html) + +3. 快速安装： +- CANN开发套件包+kernel包安装 +```commandline +# 增加软件包可执行权限，{version}表示软件版本号，{arch}表示CPU架构，{soc}表示昇腾AI处理器的版本。 +chmod +x ./Ascend-cann-toolkit_{version}_linux-{arch}.run +chmod +x ./Ascend-cann-kernels-{soc}_{version}_linux.run +# 校验软件包安装文件的一致性和完整性 +./Ascend-cann-toolkit_{version}_linux-{arch}.run --check +./Ascend-cann-kernels-{soc}_{version}_linux.run --check +# 安装 +./Ascend-cann-toolkit_{version}_linux-{arch}.run --install +./Ascend-cann-kernels-{soc}_{version}_linux.run --install + +# 设置环境变量 +source /usr/local/Ascend/ascend-toolkit/set_env.sh +``` +- MindIE包安装 +```commandline +# 增加软件包可执行权限，{version}表示软件版本号，{arch}表示CPU架构。 +chmod +x ./Ascend-mindie_${version}_linux-${arch}.run +./Ascend-mindie_${version}_linux-${arch}.run --check + +# 方式一：默认路径安装 +./Ascend-mindie_${version}_linux-${arch}.run --install +# 设置环境变量 +cd /usr/local/Ascend/mindie && source set_env.sh + +# 方式二：指定路径安装 +./Ascend-mindie_${version}_linux-${arch}.run --install-path=${AieInstallPath} +# 设置环境变量 +cd ${AieInstallPath}/mindie && source set_env.sh +``` + +- MindIE SD不需要单独安装，安装MindIE时将会自动安装 +- torch_npu 安装: +下载 pytorch_v{pytorchversion}_py{pythonversion}.tar.gz +```commandline +tar -xzvf pytorch_v{pytorchversion}_py{pythonversion}.tar.gz +# 解压后，会有whl包 +pip install torch_npu-{pytorchversion}.xxxx.{arch}.whl +``` + +### 2.3 pytorch框架(支持版本为：2.1.0) +[安装包下载](https://download.pytorch.org/whl/cpu/torch/) + +使用pip安装 +```shell +# {version}表示软件版本号，{arch}表示CPU架构。 +pip install torch-${version}-cp310-cp310-linux_${arch}.whl +``` + +### 2.4 安装依赖库 +安装MindIE-SD的依赖库。 +``` +pip install -r requirements.txt +``` + +## 三、Stable diffusion 3 + +### 3.1 权重及配置文件说明 + +权重下载路径 +``` +https://huggingface.co/stabilityai/stable-diffusion-3-medium-diffusers +``` +将下载的权重文件夹中的model_index.josn 替换为，本代码仓中的model_index.josn文件 + +将权重的文件夹权限设置为安全权限 +```bash +chmod -R 640 ./stable-diffusion-3-medium-diffusers +``` + +### 3.2 执行推理脚本 +```bash +python3 inference_sd3.py \ + --model ./path/to/model \ + --prompt_file ./prompts.txt \ + --prompt_file_type plain \ + --device 0 \ + --save_dir ./results \ + --steps 28 \ + --height 1024 \ + --width 1024 \ + --batch_size 1 +``` + +参数说明： +- --model：模型权重路径。 +- --prompt_file：提示词文件。 +- --prompt_file_type: prompt文件类型，用于指定读取方式，可选plain，parti，hpsv2。 +- --save_dir：生成图片的存放目录。 +- --batch_size：模型batch size。 +- --steps：生成图片迭代次数。 +- --device：推理设备ID； +- --height：生成图像高度，当前只支持1024 +- --width：生成图像宽度，当前只支持1024 + +### 3.3 精度验证 + +由于生成的图片存在随机性，提供两种精度验证方法： + 1. CLIP-score（文图匹配度量）：评估图片和输入文本的相关性，分数的取值范围为[-1, 1]，越高越好。使用Parti数据集进行验证。 + 2. HPSv2（图片美学度量）：评估生成图片的人类偏好评分，分数的取值范围为[0, 1]，越高越好。使用HPSv2数据集进行验证 + + 注意，由于要生成的图片数量较多，进行完整的精度验证需要耗费很长的时间。 + + 1. 下载Parti数据集和hpsv2数据集 + + ```bash + # 下载Parti数据集 + wget https://raw.githubusercontent.com/google-research/parti/main/PartiPrompts.tsv --no-check-certificate + ``` + hpsv2数据集下载链接：https://gitee.com/ascend/ModelZoo-PyTorch/blob/master/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion_xl/hpsv2_benchmark_prompts.json + + 2. 下载模型权重 + + ```bash + # Clip Score和HPSv2均需要使用的权重 + GIT_LFS_SKIP_SMUDGE=1 + git clone https://huggingface.co/laion/CLIP-ViT-H-14-laion2B-s32B-b79K + cd ./CLIP-ViT-H-14-laion2B-s32B-b79K + + # HPSv2权重 + wget https://huggingface.co/spaces/xswu/HPSv2/resolve/main/HPS_v2_compressed.pt --no-check-certificate + ``` + 也可手动下载[权重](https://huggingface.co/laion/CLIP-ViT-H-14-laion2B-s32B-b79K/blob/main/open_clip_pytorch_model.bin) + 将权重放到`CLIP-ViT-H-14-laion2B-s32B-b79K`目录下，手动下载[HPSv2权重](https://huggingface.co/spaces/xswu/HPSv2/resolve/main/HPS_v2_compressed.pt)放到当前路径 + + 3. 使用推理脚本读取Parti数据集，生成图片 + + ```bash + python3 inference_sd3.py \ + --model ./path/to/model \ + --prompt_file ./PartiPrompts.tsv \ + --prompt_file_type parti \ + --num_images_per_prompt 4 \ + --max_num_prompts 0 \ + --device 0 \ + --save_dir ./results_PartiPrompts \ + --steps 28 \ + --output_dir ./models \ + --height 1024 \ + --width 1024 \ + --batch_size 1 + ``` + + 参数说明： + - --model：模型权重路径。 + - --prompt_file：提示词文件。 + - --prompt_file_type: prompt文件类型，用于指定读取方式，可选plain，parti，hpsv2。注意使用hpsv2时，设置num_images_per_prompt=1即可。 + - --num_images_per_prompt: 每个prompt生成的图片数量。注意使用hpsv2时，设置num_images_per_prompt=1即可。 + - --max_num_prompts：限制prompt数量为前X个，0表示不限制。 + - --save_dir：生成图片的存放目录。 + - --batch_size：模型batch size。 + - --steps：生成图片迭代次数。 + - --device：推理设备ID + + 执行完成后在`./results_PartiPrompts`目录下生成推理图片，在当前目录生成一个`image_info.json`文件，记录着图片和prompt的对应关系，并在终端显示推理时间。 + +4. 使用推理脚本读取hpsv2数据集，生成图片 + + ```bash + # 不使用并行 + python3 inference_sd3.py \ + --model ${model_base} \ + --prompt_file_type hpsv2 \ + --num_images_per_prompt 1 \ + --info_file_save_path ./image_info_hpsv2.json \ + --device 0 \ + --save_dir ./results_hpsv2 \ + --steps 28 \ + --output_dir ./models \ + --height 1024 \ + --width 1024 \ + --batch_size 1 + ``` + 参数说明： + - --info_file_save_path：生成图片信息的json文件路径。 + + 执行完成后在`./results_hpsv2`目录下生成推理图片，在当前目录生成一个`image_info_hpsv2.json`文件，记录着图片和prompt的对应关系，并在终端显示推理时间。 + +5. 计算精度指标 + 1. CLIP-score + ```bash + python3 clip_score.py \ + --device=cpu \ + --image_info="image_info.json" \ + --model_name="ViT-H-14" \ + --model_weights_path="./CLIP-ViT-H-14-laion2B-s32B-b79K/open_clip_pytorch_model.bin" + ``` + + 参数说明： + - --device: 推理设备，默认为"cpu"，如果是cuda设备可设置为"cuda"。 + - --image_info: 上一步生成的`image_info.json`文件。 + - --model_name: Clip模型名称。 + - --model_weights_path: Clip模型权重文件路径。 + + clip_score.py脚本可参考[SDXL](https://gitee.com/ascend/ModelZoo-PyTorch/blob/master/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion_xl/clip_score.py)，执行完成后会在屏幕打印出精度计算结果。 + + 2. HPSv2 + ```bash + python3 hpsv2_score.py \ + --image_info="image_info_hpsv2.json" \ + --HPSv2_checkpoint="./HPS_v2_compressed.pt" \ + --clip_checkpoint="./CLIP-ViT-H-14-laion2B-s32B-b79K/open_clip_pytorch_model.bin" + ``` + + 参数说明： + - --image_info: 上一步生成的`image_info_hpsv2.json`文件。 + - --HPSv2_checkpoint: HPSv2模型权重文件路径。 + - --clip_checkpointh: Clip模型权重文件路径。 + + hpsv2_score.py脚本可参考[SDXL](https://gitee.com/ascend/ModelZoo-PyTorch/blob/master/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion_xl/hpsv2_score.py)，执行完成后会在屏幕打印出精度计算结果。 + +性能测试需要独占npu和cpu \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-3/__init__.py b/MindIE/MultiModal/StableDiffusion-3/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..b0a69f2c91b5f23b70942db4be66df70ccd4acaf --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-3/__init__.py @@ -0,0 +1,11 @@ +from .models.dit import SD3DiTModel +from .models.vae import AutoencoderKL +from .pipepline.stable_diffusion_pipeline import StableDiffusion3Pipeline +from .schedulers.flow_match_euler_discrete import FlowMatchEulerDiscreteScheduler + +__all__ = [ + "SD3DiTModel", + "AutoencoderKL", + "StableDiffusion3Pipeline", + "FlowMatchEulerDiscreteScheduler", +] \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-3/clip_score.py b/MindIE/MultiModal/StableDiffusion-3/clip_score.py new file mode 100644 index 0000000000000000000000000000000000000000..069f5d6e9a9baaa61b9a3537bcab6f637605858e --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-3/clip_score.py @@ -0,0 +1,140 @@ +# Copyright 2023 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import os +import json +import time +import argparse + +import open_clip +import numpy as np +from PIL import Image +import torch +import torch.nn.functional as F + + +def clip_score(model_clip, tokenizer, preprocess, prompt, image_files, device): + imgs = [] + texts = [] + for image_file in image_files: + img = preprocess(Image.open(image_file)).unsqueeze(0).to(device) + imgs.append(img) + text = tokenizer([prompt]).to(device) + texts.append(text) + + img = torch.cat(imgs) # [bs, 3, 224, 224] + text = torch.cat(texts) # [bs, 77] + + with torch.no_grad(): + text_ft = model_clip.encode_text(text).float() + img_ft = model_clip.encode_image(img).float() + score = F.cosine_similarity(img_ft, text_ft).squeeze() + + return score.cpu() + + +def main(): + args = parse_arguments() + + if args.device is None: + device = torch.device('cuda' if (torch.cuda.is_available()) else 'cpu') + else: + device = torch.device(args.device) + + t_b = time.time() + print(f"Load clip model...") + model_clip, _, preprocess = open_clip.create_model_and_transforms( + args.model_name, pretrained=args.model_weights_path, device=device) + model_clip.eval() + print(f">done. elapsed time: {(time.time() - t_b):.3f} s") + + tokenizer = open_clip.get_tokenizer(args.model_name) + + with os.fdopen(os.open(args.image_info, os.O_RDONLY), "r") as f: + image_info = json.load(f) + + t_b = time.time() + print(f"Calc clip score...") + all_scores = [] + cat_scores = {} + + for i, info in enumerate(image_info): + image_files = info['images'] + category = info['category'] + prompt = info['prompt'] + + print(f"[{i + 1}/{len(image_info)}] {prompt}") + + image_scores = clip_score(model_clip, + tokenizer, + preprocess, + prompt, + image_files, + device) + if len(image_files) > 1: + best_score = max(image_scores) + else: + best_score = image_scores + + print(f"image scores: {image_scores}") + print(f"best score: {best_score}") + + all_scores.append(best_score) + if category not in cat_scores: + cat_scores[category] = [] + cat_scores[category].append(best_score) + print(f">done. elapsed time: {(time.time() - t_b):.3f} s") + + average_score = np.average(all_scores) + print(f"====================================") + print(f"average score: {average_score:.3f}") + print(f"category average scores:") + cat_average_scores = {} + for category, scores in cat_scores.items(): + cat_average_scores[category] = np.average(scores) + print(f"[{category}], average score: {cat_average_scores[category]:.3f}") + + +def parse_arguments(): + parser = argparse.ArgumentParser() + parser.add_argument( + "--device", + type=str, + default="cpu", + choices=["cpu", "cuda"], + help="device for torch.", + ) + parser.add_argument( + "--image_info", + type=str, + default="./image_info.json", + help="Image_info.json file.", + ) + parser.add_argument( + "--model_name", + type=str, + default="ViT-H-14", + help="open clip model name", + ) + parser.add_argument( + "--model_weights_path", + type=str, + default="./CLIP-ViT-H-14-laion2B-s32B-b79K/open_clip_pytorch_model.bin", + help="open clip model weights", + ) + return parser.parse_args() + + +if __name__ == '__main__': + main() \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-3/hpsv2_score.py b/MindIE/MultiModal/StableDiffusion-3/hpsv2_score.py new file mode 100644 index 0000000000000000000000000000000000000000..04e9bd8d8f82ece84c642520b001b62901286eda --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-3/hpsv2_score.py @@ -0,0 +1,123 @@ +# Copyright 2024 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import argparse +import os +from typing import Union +import json + +from clint.textui import progress +import hpsv2 +from hpsv2.utils import root_path, hps_version_map +from hpsv2.src.open_clip import create_model_and_transforms, get_tokenizer +import huggingface_hub +from PIL import Image +import requests +import torch + + +def initialize_model(pretrained_path, device): + model, _, preprocess_val = create_model_and_transforms( + "ViT-H-14", pretrained=pretrained_path, precision='amp', + device=device, + jit=False, + force_quick_gelu=False, + force_custom_text=False, + force_patch_dropout=False, + force_image_size=None, + pretrained_image=False, + image_mean=None, + image_std=None, + light_augmentation=True, + aug_cfg={}, + output_dict=True, + with_score_predictor=False, + with_region_predictor=False + ) + return model, preprocess_val + + +def parse_arguments(): + parser = argparse.ArgumentParser() + parser.add_argument( + "--image_info", + type=str, + default="./image_info.json", + help="Image_info.json file.", + ) + parser.add_argument( + "--HPSv2_checkpoint", + type=str, + default="./HPS_v2_compressed.pt", + help="HPS_v2 model weights", + ) + parser.add_argument( + "--clip_checkpoint", + type=str, + default="./CLIP-ViT-H-14-laion2B-s32B-b79K/open_clip_pytorch_model.bin", + help="open clip model weights", + ) + return parser.parse_args() + + +def main(): + args = parse_arguments() + + device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') + + model, preprocess_val = initialize_model(args.clip_checkpoint, device) + + checkpoint = torch.load(args.HPSv2_checkpoint, map_location=device) + model.load_state_dict(checkpoint['state_dict']) + tokenizer = get_tokenizer('ViT-H-14') + model = model.to(device) + model.eval() + + with os.fdopen(os.open(args.image_info, os.O_RDONLY), "r") as f: + image_info = json.load(f) + + result = [] + for i, info in enumerate(image_info): + image_file = info['images'][0] + prompt = info['prompt'] + + # Load your image and prompt + with torch.no_grad(): + # Process the image + if isinstance(image_file, str): + image = preprocess_val(Image.open(image_file)) + elif isinstance(image_file, Image.Image): + image = preprocess_val(image_file) + else: + raise TypeError('The type of parameter img_path is illegal.') + image = image.unsqueeze(0).to(device=device, non_blocking=True) + # Process the prompt + text = tokenizer([prompt]).to(device=device, non_blocking=True) + # Calculate the HPS + with torch.cuda.amp.autocast(): + outputs = model(image, text) + image_features = outputs["image_features"] + text_features = outputs["text_features"] + logits_per_image = image_features @ text_features.T + + hps_score = torch.diagonal(logits_per_image).cpu().numpy() + print(f"image {i} hps_score: ", hps_score[0]) + + result.append(hps_score[0]) + + print('avg HPSv2 score:', sum(result) / len(result)) + + +if __name__ == '__main__': + main() \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-3/inference_sd3.py b/MindIE/MultiModal/StableDiffusion-3/inference_sd3.py new file mode 100644 index 0000000000000000000000000000000000000000..0b81cc178a97cfed77098a0f9cc56c6df0ef1294 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-3/inference_sd3.py @@ -0,0 +1,287 @@ +#!/usr/bin/env python +# coding=utf-8 +# Copyright 2024 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import argparse +import time +import os +import json +import csv + +import torch +import torch_npu + +from stable_diffusion3 import StableDiffusion3Pipeline +from stable_diffusion3.utils import set_random_seed + + +class PromptLoader: + def __init__( + self, + prompt_file: str, + prompt_file_type: str, + batch_size: int, + num_images_per_prompt: int = 1, + max_num_prompts: int = 0 + ): + self.prompts = [] + self.catagories = ['Not_specified'] + self.batch_size = batch_size + self.num_images_per_prompt = num_images_per_prompt + + if prompt_file_type == 'plain': + self.load_prompts_plain(prompt_file, max_num_prompts) + elif prompt_file_type == 'parti': + self.load_prompts_parti(prompt_file, max_num_prompts) + elif prompt_file_type == 'hpsv2': + self.load_prompts_hpsv2(max_num_prompts) + else: + print("This operation is not supported!") + + self.current_id = 0 + self.inner_id = 0 + + def __len__(self): + return len(self.prompts) * self.num_images_per_prompt + + def __iter__(self): + return self + + def __next__(self): + if self.current_id == len(self.prompts): + raise StopIteration + + ret = { + 'prompts': [], + 'catagories': [], + 'save_names': [], + 'n_prompts': self.batch_size, + } + for _ in range(self.batch_size): + if self.current_id == len(self.prompts): + ret['prompts'].append('') + ret['save_names'].append('') + ret['catagories'].append('') + ret['n_prompts'] -= 1 + + else: + prompt, catagory_id = self.prompts[self.current_id] + ret['prompts'].append(prompt) + ret['catagories'].append(self.catagories[catagory_id]) + ret['save_names'].append(f'{self.current_id}_{self.inner_id}') + + self.inner_id += 1 + if self.inner_id == self.num_images_per_prompt: + self.inner_id = 0 + self.current_id += 1 + + return ret + + def load_prompts_plain(self, file_path: str, max_num_prompts: int): + with os.fdopen(os.open(file_path, os.O_RDONLY), "r") as f: + for i, line in enumerate(f): + if max_num_prompts and i == max_num_prompts: + break + + prompt = line.strip() + self.prompts.append((prompt, 0)) + + def load_prompts_parti(self, file_path: str, max_num_prompts: int): + with os.fdopen(os.open(file_path, os.O_RDONLY), "r") as f: + # Skip the first line + next(f) + tsv_file = csv.reader(f, delimiter="\t") + for i, line in enumerate(tsv_file): + if max_num_prompts and i == max_num_prompts: + break + + prompt = line[0] + catagory = line[1] + if catagory not in self.catagories: + self.catagories.append(catagory) + + catagory_id = self.catagories.index(catagory) + self.prompts.append((prompt, catagory_id)) + + def load_prompts_hpsv2(self, max_num_prompts: int): + with open('hpsv2_benchmark_prompts.json', 'r') as file: + all_prompts = json.load(file) + count = 0 + for style, prompts in all_prompts.items(): + for prompt in prompts: + count += 1 + if max_num_prompts and count >= max_num_prompts: + break + + if style not in self.catagories: + self.catagories.append(style) + + catagory_id = self.catagories.index(style) + self.prompts.append((prompt, catagory_id)) + +def parse_arguments(): + parser = argparse.ArgumentParser() + parser.add_argument( + "-m", + "--model", + type=str, + default="./stable-diffusion-3-medium-diffusers", + help="Path or name of the pre-trained model.", + ) + parser.add_argument( + "--prompt_file", + type=str, + default="./prompts.txt", + help="A text file of prompts for generating images.", + ) + parser.add_argument( + "--prompt_file_type", + choices=["plain", "parti", "hpsv2"], + default="plain", + help="Type of prompt file.", + ) + parser.add_argument( + "--save_dir", + type=str, + default="./results", + help="Path to save result images.", + ) + parser.add_argument( + "--info_file_save_path", + type=str, + default="./image_info.json", + help="Path to save image information file.", + ) + parser.add_argument( + "--steps", + type=int, + default=28, + help="Number of inference steps.", + ) + parser.add_argument( + "--device", + type=int, + default=0, + help="NPU device id. Give 2 ids to enable parallel inferencing.", + ) + parser.add_argument( + "--num_images_per_prompt", + default=1, + type=int, + help="Number of images generated for each prompt.", + ) + parser.add_argument( + "--max_num_prompts", + default=0, + type=int, + help="Limit the number of prompts (0: no limit).", + ) + parser.add_argument( + "-bs", + "--batch_size", + type=int, + default=1, + help="Batch size." + ) + + parser.add_argument( + "--height", + default=1024, + type=int, + help="image height", + ) + parser.add_argument( + "--width", + default=1024, + type=int, + help="image width" + ) + parser.add_argument( + '--seed', + type=int, + default=1234, + help='Random seed' + ) + + return parser.parse_args() + + +if __name__ == "__main__": + args = parse_arguments() + path=args.model + + torch.npu.set_device(args.device) + save_dir = args.save_dir + if not os.path.exists(args.save_dir): + os.makedirs(args.save_dir) + + prompt_loader = PromptLoader(args.prompt_file, + args.prompt_file_type, + args.batch_size, + args.num_images_per_prompt, + args.max_num_prompts) + + pipe = StableDiffusion3Pipeline.from_pretrained(path, dtype=torch.float16) + pipe.to("npu") + set_random_seed(args.seed) + + use_time = 0 + infer_num = 0 + image_info = [] + current_prompt = None + for i, input_info in enumerate(prompt_loader): + prompts = input_info['prompts'] + catagories = input_info['catagories'] + save_names = input_info['save_names'] + n_prompts = input_info['n_prompts'] + + print(f"[{infer_num + n_prompts}/{len(prompt_loader)}]: {prompts}") + infer_num += args.batch_size + + start_time = time.time() + images = pipe( + prompts, + negative_prompt="", + height=args.height, + width=args.width, + num_inference_steps=args.steps, + guidance_scale=7.5, + output_type="pil" + ) + if i > 4: # do not count the time spent inferring the first 0 to 4 images + use_time += time.time() - start_time + + for j in range(n_prompts): + image_save_path = os.path.join(save_dir, f"{save_names[j]}.png") + image = images[0][j] + image.save(image_save_path) + + if current_prompt != prompts[j]: + current_prompt = prompts[j] + image_info.append({'images': [], 'prompt': current_prompt, 'category': catagories[j]}) + + image_info[-1]['images'].append(image_save_path) + + + infer_num = infer_num - 5 # do not count the time spent inferring the first 5 images + print(f"[info] infer number: {infer_num}; use time: {use_time:.3f}s\n" + f"average time: {use_time / infer_num:.3f}s\n") + + # Save image information to a json file + if os.path.exists(args.info_file_save_path): + os.remove(args.info_file_save_path) + + with os.fdopen(os.open(args.info_file_save_path, os.O_RDWR | os.O_CREAT, 0o640), "w") as f: + json.dump(image_info, f) \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-3/layers/activation.py b/MindIE/MultiModal/StableDiffusion-3/layers/activation.py new file mode 100644 index 0000000000000000000000000000000000000000..f84e230e7dfdc8dfac5bc9ce03b87073ef91797e --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-3/layers/activation.py @@ -0,0 +1,57 @@ +#!/usr/bin/env python +# coding=utf-8 +# Copyright 2024 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import torch.nn as nn +import torch_npu + + +class GEGLU(nn.Module): + + def __init__(self): + super().__init__() + + def forward(self, hidden_states): + + return torch_npu.npu_geglu(hidden_states, dim=-1, approximate=1)[0] + + +ACTIVATION_FUNCTIONS = { + "swish": nn.SiLU(), + "silu": nn.SiLU(), + "mish": nn.Mish(), + "gelu": nn.GELU(), + "relu": nn.ReLU(), + "gelu-approximate": nn.GELU(approximate="tanh"), + "geglu": GEGLU() +} + + + +def get_activation_fn(act_fn: str) -> nn.Module: + """Helper function to get activation function from string. + + Args: + act_fn (str): Name of activation function. + + Returns: + nn.Module: Activation function. + """ + + act_fn = act_fn.lower() + if act_fn in ACTIVATION_FUNCTIONS: + return ACTIVATION_FUNCTIONS[act_fn] + else: + raise ValueError(f"Unsupported activation function: {act_fn}") \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-3/layers/attention.py b/MindIE/MultiModal/StableDiffusion-3/layers/attention.py new file mode 100644 index 0000000000000000000000000000000000000000..b561ba2d6bd7b3f6b1758b915803bfe0c3aa7ef6 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-3/layers/attention.py @@ -0,0 +1,275 @@ +#!/usr/bin/env python +# coding=utf-8 +# Copyright 2024 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# This source code is licensed under the license found in the +# LICENSE file in the root directory of this source tree. + +import math +import inspect +from typing import Optional, Tuple + +import torch +import torch.nn as nn +import torch_npu + + +class ReconstitutionAttention(nn.Module): + r""" + Attention layer. + """ + def __init__( + self, + attention_dim: int, + cross_attention_dim: Optional[int] = None, + num_heads: int = 8, + head_dim: int = 64, + qkv_bias: bool = True, + out_proj_bias: bool = True, + num_norm_groups: Optional[int] = None, + add_proj_dim: Optional[int] = None, + add_proj_bias: bool = True, + enable_add_out_proj: bool = True, + scale_attention: bool = True, + eps: float = 1e-5, + processor: Optional["JointAttnProcessor"] = None, + ): + r""" + Attention layer init function. + Args: + attention_dim (`int`): + The number of channels in the hidden_states. + cross_attention_dim (`int`, *optional*, defaults to `None`): + The number of channels in the encoder_hidden_states. If not `None`, means cross attention. + num_heads (`int`, *optional*, defaults to 8): + The number of attention heads used in the multi-head attention layers. + head_dim (`int`, *optional*, defaults to 64): + The number of dims in each head. + qkv_bias (`bool`, *optional*, defaults to `True`): + Whether or not the quert, key and value linear layer to contain a bias parameter. + out_proj_bias (`bool`, *optional*, defaults to `True`): + Whether or not the out projection layer to contain a bias parameter. + num_norm_groups (`int`, *optional*, defaults to `None`): + The number of groups to use for the `GroupNorm` in attention. + If `None`, no `GroupNorm` is used. + attention_norm (`str`, *optional*, defaults to `None`): + The type of normalization to use for the query and key in attention. + Can be `None`, `layer_norm`, or `llama_rms_norm`. + position_embedding (`str`, *optional*, defaults to `None`): + The type of position embdding to use for the query and key in attention. Can be `None`, `rope`. + add_proj_dim (`int`, *optional*, defaults to `None`): + The number of channels to use for the additional projections. If `None`, no projection is used. + add_proj_bias (`bool`, *optional*, defaults to `True`): + Whether or not the additional projection layer to contain a bias parameter. + enable_add_out_proj (`bool`, *optional*, defaults to `True`): + Whether or not use the additional out projection. + scale_attention (`bool`, *optional*, defaults to `True`): + Set `True` to scale the query @ key result with by `1 / sqrt(head_dim)`. + eps (`float`, *optional*, defaults to 1e-5): + The additional value added to eh denominator in normalization. + processor (`AttnProcessor`, *optional*, defaults to `None`): + The attention processor to use. If `None`, `AttnProcessor` will be used. + """ + super().__init__() + + self.num_heads = num_heads + if head_dim <= 0: + raise ValueError(f"Input head_dim should be greater than zero, but got {head_dim}.") + self.head_dim = head_dim + self.pad_dim = self.head_dim + + self.is_cross_attention = cross_attention_dim is not None + + self.scale_value = 1 / math.sqrt(head_dim) if scale_attention else 1.0 + + # `hidden_size` is calculated by num_heads * head_dim -> H = N * D + hidden_size = num_heads * head_dim + + + # Normalize hidden states by group_norm + self.group_norm = nn.GroupNorm(num_channels=hidden_size, num_groups=num_norm_groups, eps=eps, affine=True) \ + if num_norm_groups is not None else None + + # QKVLinear + if self.is_cross_attention: + self.q_proj = nn.Linear(attention_dim, hidden_size, bias=qkv_bias) + self.kv_proj = nn.Linear(cross_attention_dim, 2 * hidden_size, bias=qkv_bias) + else: + self.qkv_proj = nn.Linear(attention_dim, 3 * hidden_size, bias=qkv_bias) + + + + # Additional qkv linear for Multi-Modal Diffusion Transformer + if add_proj_dim is not None: + self.add_qkv_proj = nn.Linear(add_proj_dim, 3 * hidden_size, bias=add_proj_bias) # 3: qkv + + # OutLinear + self.out_proj = nn.Linear(hidden_size, attention_dim, bias=out_proj_bias) + + # Additional out linear for Multi-Modal Diffusion Transformer + if add_proj_dim is not None: + # For the last attention layer in Multi-Modal Diffusion Transformer, + # no need to calculate the additional out linear + self.add_out_proj = nn.Linear(hidden_size, attention_dim, bias=out_proj_bias) \ + if enable_add_out_proj else nn.Identity() + + # Set default processor by AttnProcessor + attn_processor = processor if processor is not None else AttnProcessor() + self.set_processor(attn_processor) + + def set_processor(self, processor: "AttnProcessor"): + """ + Set the attention processor. + Users can develop different attention processor for `Attention` to achieve different functions. + Args: + processor: ("AttnProcessor"): + The attention processor to used for attention forward. + """ + self.processor = processor + + def forward( + self, + hidden_states: torch.Tensor, + encoder_hidden_states: torch.Tensor = None, + attention_mask: Optional[torch.Tensor] = None, + **kwargs, + ) -> torch.Tensor: + """ + Attention forward function. + Args: + hidden_states (`torch.Tensor`): + The hidden states of attention query. + encoder_hidden_states (`torch.Tensor`, *optional*, defaults to `None`): + The hidden states of the encoder. + attention_mask (`torch.Tensor`, *optional*, defaults to `None`): + The mask of attention. + **kwargs: + The additional arguments to the attention processors. + For standard attention use `AttnProcessor`, kwargs is empty. + Returns: + `torch.Tensor`: The output of the attention layer. + """ + attn_parameters = set(inspect.signature(self.processor.__call__).parameters.keys()) + attn_kwargs = {key: value for key, value in kwargs.items() if key in attn_parameters} + + return self.processor( + self, + hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=attention_mask, + **attn_kwargs + ) + + +class AttnProcessor: + """ + The standard attention processor. + """ + def __call__( + self, + attn: ReconstitutionAttention, + hidden_states: torch.Tensor, + encoder_hidden_states: Optional[torch.Tensor] = None, + attention_mask: Optional[torch.Tensor] = None, + **kwargs, + ) -> torch.Tensor: + if hidden_states is None: + raise ValueError("Input hidden_states should not be none.") + + # only support BNC now. + if hidden_states.ndim != 3: # 3: BNC + raise ValueError(f"The dimensions of hidden_states should be 3, but got {hidden_states.ndim}") + + batch_size = hidden_states.shape[0] + + if attn.group_norm is not None: + # In `BSH`, `H` represents channel, so it needs to be transposed. + hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2) + + if attn.is_cross_attention: + query = attn.q_proj(hidden_states) + query = query.reshape(batch_size, -1, attn.num_heads, attn.head_dim).transpose(1, 2) # B S N D -> B N S D + + kv = attn.kv_proj(encoder_hidden_states) + kv = kv.reshape(batch_size, -1, 2, attn.num_heads, attn.head_dim) + key, value = kv.permute(2, 0, 3, 1, 4).unbind(0) # B S 2 N D -> 2 B N S D -> 2 * B N S D + else: + qkv = attn.qkv_proj(hidden_states) + qkv = qkv.reshape(batch_size, -1, 3, attn.num_heads, attn.head_dim) # 3: q,k,v + query, key, value = qkv.permute(2, 0, 3, 1, 4).unbind(0) # B S 3 N D -> 3 B N S D -> 3 * B N S D + + hidden_states = torch_npu.npu_prompt_flash_attention( + query, key, value, + num_heads=query.shape[1], + input_layout="BNSD", + atten_mask=attention_mask, + scale_value=attn.scale_value) + # transform the hidden_states layout from BNSD to BSH + hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.num_heads * attn.head_dim) + hidden_states = attn.out_proj(hidden_states) + return hidden_states + + +class JointAttnProcessor: + """ + The joint attention processor for Multi-Modal Diffusion Transformer. + """ + def __call__( + self, + attn: ReconstitutionAttention, + hidden_states: torch.Tensor, + encoder_hidden_states: torch.Tensor, + attention_mask: Optional[torch.Tensor] = None, + **kwargs, + ) -> Tuple[torch.Tensor, torch.Tensor]: + if hidden_states is None: + raise ValueError("Input hidden_states should not be none.") + + # only support BNC now. + if hidden_states.ndim != 3: # 3: BNC + raise ValueError(f"The dimensions of hidden_states should be 3, but got {hidden_states.ndim}") + + if encoder_hidden_states is None: + raise ValueError("Input encoder_hidden_states should not be none.") + + # only support BNC now. + if encoder_hidden_states.ndim != 3: # 3: BNC + raise ValueError( + f"The dimensions of encoder_hidden_states should be 3, but got {encoder_hidden_states.ndim}") + + batch_size, seqlen, _ = hidden_states.shape + encoder_seqlen = encoder_hidden_states.shape[1] + + # joint attention linear + qkv = attn.qkv_proj(hidden_states) + add_qkv = attn.add_qkv_proj(encoder_hidden_states) + qkv = torch.cat([qkv, add_qkv], dim=1) + qkv = qkv.reshape(batch_size, -1, 3, attn.num_heads, attn.head_dim) # 3: q,k,v + + query, key, value = qkv.permute(2, 0, 3, 1, 4).unbind(0) # B S 3 N D -> 3 B N S D -> 3 * B N S D + + # need replaced by dispatch flash_attention function + hidden_states = torch_npu.npu_prompt_flash_attention( + query, key, value, + num_heads=query.shape[1], + input_layout="BNSD", + scale_value=attn.scale_value) + # transform the hidden_states layout from BNSD to BSH + hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.num_heads * attn.head_dim) + + hidden_states, encoder_hidden_states = hidden_states.split([seqlen, encoder_seqlen], dim=1) + hidden_states = attn.out_proj(hidden_states) + encoder_hidden_states = attn.add_out_proj(encoder_hidden_states) + return hidden_states, encoder_hidden_states \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-3/layers/embedding.py b/MindIE/MultiModal/StableDiffusion-3/layers/embedding.py new file mode 100644 index 0000000000000000000000000000000000000000..0baa0235c70e0f68eae580cd307c2c3358110b76 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-3/layers/embedding.py @@ -0,0 +1,256 @@ +#!/usr/bin/env python +# coding=utf-8 +# Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# This source code is licensed under the license found in the +# LICENSE file in the root directory of this source tree. + +import math + +import torch +import torch.nn as nn +import numpy as np + +from .mlp import Mlp + + +def timestep_embedding( + timesteps: torch.Tensor, + output_dims: int, + flip: bool = False, + max_period: int = 10000, +) -> torch.Tensor: + """ + Create sinusoidal timestep embeddings. + Args: + timesteps (torch.Tensor): + Time steps sequence. Must be a 1D tensor. + output_dims (int): + The dimension of output. + flip (bool): + If true, return [cos, sin], else return [sin, cos] + max_period (int): + The maximum frequency of the embeddings. + Returns: + torch.Tensor: The tensor of positional embeddings. + """ + if output_dims <= 0: + raise ValueError("output_dims is invalid") + + half_dim = output_dims // 2 + + # generate frequency vectors + exponent = -math.log(max_period) * torch.arange( + start=0, end=half_dim, dtype=torch.float32, device=timesteps.device + ) + exponent = exponent / (half_dim) + + emb = torch.exp(exponent) + emb = timesteps[:, None].float() * emb[None, :] + + # flip sine and cosine embeddings + if flip: + emb = torch.cat([torch.cos(emb), torch.sin(emb)], dim=-1) + else: + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=-1) + + # zero pad + if output_dims % 2 == 1: + emb = torch.nn.functional.pad(emb, (0, 1, 0, 0)) + return emb + + +class CombinedTimestepTextProjEmbeddings(nn.Module): + def __init__(self, embedding_dim, pooled_projection_dim): + super().__init__() + + self.timestep_embedder = Mlp(features_in=256, features_hidden=embedding_dim, features_out=embedding_dim, act_layer="silu") + + self.text_embedder = Mlp(features_in=pooled_projection_dim, features_hidden=embedding_dim, features_out=embedding_dim, act_layer="silu") + + def forward(self, timesteps, pooled_projection): + embedding = timestep_embedding(timesteps, 256, flip=True) + timesteps_emb = self.timestep_embedder(embedding.to(dtype=pooled_projection.dtype)) # (N, D) + + pooled_projections = self.text_embedder(pooled_projection) + conditioning = timesteps_emb + pooled_projections + + return conditioning + + +class PatchEmbed2D(nn.Module): + """ + 2D Image to Patch Embedding + """ + def __init__( + self, + height=224, + width=224, + patch_size=16, + in_channels=3, + embed_dim=768, + flatten=True, + bias=True, + pos_embed_type="sincos", + pos_embed_max_size=None, + ) -> None: + super().__init__() + num_patches = (height // patch_size) * (width // patch_size) + self.proj = nn.Conv2d( + in_channels, embed_dim, kernel_size=(patch_size, patch_size), stride=patch_size, bias=bias + ) + self.height, self.width = height // patch_size, width // patch_size + self.patch_size = patch_size + self.base_size = height // patch_size + # Calculate positional embeddings based on max size or default + if pos_embed_max_size: + grid_size = pos_embed_max_size + else: + grid_size = int(num_patches**0.5) + self.pos_embed_max_size = pos_embed_max_size + self.pos_embed_type = pos_embed_type + self.flatten = flatten + if pos_embed_type is None: + self.pos_embed = None + elif pos_embed_type == "sincos": + pos_embed = get_2d_sincos_pos_embed( + embed_dim, grid_size, base_size=self.base_size + ) + persistent = True if pos_embed_max_size else False + self.register_buffer("pos_embed", torch.from_numpy(pos_embed).float().unsqueeze(0), persistent=persistent) + + else: + raise ValueError(f"Unsupported pos_embed_type: {pos_embed_type}") + + def cropped_pos_embed(self, height, width): + """Crops positional embeddings for SD3 compatibility.""" + + height = height // self.patch_size + width = width // self.patch_size + if height > self.pos_embed_max_size: + raise ValueError( + f"Height ({height}) cannot be greater than `pos_embed_max_size`: {self.pos_embed_max_size}." + ) + if width > self.pos_embed_max_size: + raise ValueError( + f"Width ({width}) cannot be greater than `pos_embed_max_size`: {self.pos_embed_max_size}." + ) + # Use the center point of the patch as the initial position + top = (self.pos_embed_max_size - height) // 2 + left = (self.pos_embed_max_size - width) // 2 + # Directly use the initialized position encoding + spatial_pos_embed = self.pos_embed.reshape(1, self.pos_embed_max_size, self.pos_embed_max_size, -1) + # Spatial position encoding, only take the center part + spatial_pos_embed = spatial_pos_embed[:, top : top + height, left : left + width, :] + # 1, H*W, D after cropping + spatial_pos_embed = spatial_pos_embed.reshape(1, -1, spatial_pos_embed.shape[-1]) + return spatial_pos_embed + + # Input latent, split latent according to patch and encode + def forward(self, latent: torch.Tensor): + + # This parameter is used to determine whether to use the maximum value to generate position encoding + if self.pos_embed_max_size is not None: + height, width = latent.shape[-2:] + else: + # If not using the maximum, use the result after patching + height, width = latent.shape[-2] // self.patch_size, latent.shape[-1] // self.patch_size + + dtype=latent.dtype + dtype_weight = self.pos_embed.dtype + if dtype != dtype_weight: + latent = latent.to(dtype_weight) + # self.proj.to(torch.float32) + # latent = latent.to(torch.float32) + latent = self.proj(latent) # embedding + if self.flatten: + # BCHW -> BNC + latent = latent.flatten(2).transpose(1, 2) + # If no position encoding is done, return directly + if self.pos_embed_type is None: + return latent + # Position encoding, interpolation or cropping + if self.pos_embed_max_size: + # self.pos_embed.to(torch.float32) + pos_embed = self.cropped_pos_embed(height, width) + + else: + # If it is the result of patching + if self.height != height or self.width != width: + pos_embed = get_2d_sincos_pos_embed( + embed_dim=self.pos_embed.shape[-1], + grid_size=(height, width), + base_size=self.base_size, + ) + pos_embed = torch.from_numpy(pos_embed).float().unsqueeze(0).to(latent.device) + else: + pos_embed = self.pos_embed + + return (latent + pos_embed).to(dtype) + + +def get_2d_sincos_pos_embed( + embed_dim, grid_size, base_size=16 +): + """ + grid_size: int of the grid height and width return: pos_embed: [grid_size*grid_size, embed_dim] or + [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token) + """ + if isinstance(grid_size, int): + grid_size = (grid_size, grid_size) + + grid_h = np.arange(grid_size[0], dtype=np.float32) / (grid_size[0] / base_size) + grid_w = np.arange(grid_size[1], dtype=np.float32) / (grid_size[1] / base_size) + # Generate coordinate matrix + grid = np.meshgrid(grid_w, grid_h) # here w goes first + grid = np.stack(grid, axis=0) + + grid = grid.reshape([2, 1, grid_size[1], grid_size[0]]) + # Get the cos_sin position encoding in the grid + pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid) + return pos_embed + + +def get_2d_sincos_pos_embed_from_grid(embed_dim, grid): + if embed_dim % 2 != 0: + raise ValueError("embed_dim must be divisible by 2") + + # use half of dimensions to encode grid_h + emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0]) # (H*W, D/2) + emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1]) # (H*W, D/2) + + emb = np.concatenate([emb_h, emb_w], axis=1) # (H*W, D) + return emb + + +def get_1d_sincos_pos_embed_from_grid(embed_dim, pos): + """ + embed_dim: output dimension for each position pos: a list of positions to be encoded: size (M,) out: (M, D) + """ + if embed_dim % 2 != 0: + raise ValueError("embed_dim must be divisible by 2") + + omega = np.arange(embed_dim // 2, dtype=np.float64) + omega /= embed_dim / 2.0 + omega = 1.0 / 10000**omega # (D/2,) + + pos = pos.reshape(-1) # (M,) + out = np.einsum("m,d->md", pos, omega) # (M, D/2), outer product + + emb_sin = np.sin(out) # (M, D/2) + emb_cos = np.cos(out) # (M, D/2) + + emb = np.concatenate([emb_sin, emb_cos], axis=1) # (M, D) + return emb \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-3/layers/mlp.py b/MindIE/MultiModal/StableDiffusion-3/layers/mlp.py new file mode 100644 index 0000000000000000000000000000000000000000..b6a804fb5a1b0b896cb45f6efbdc0539ed96ef4b --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-3/layers/mlp.py @@ -0,0 +1,62 @@ +#!/usr/bin/env python +# coding=utf-8 +# Copyright 2024 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import collections.abc +from itertools import repeat +from functools import partial + +import torch.nn as nn +from .activation import get_activation_fn + + +class Mlp(nn.Module): + """ MLP as used in Vision Transformer, MLP-Mixer and related networks + """ + + def __init__( + self, + features_in, + act_layer, + features_hidden=None, + features_out=None, + norm_layer=None, + bias=True, + ): + super().__init__() + features_out = features_out or features_in + features_hidden = features_hidden or features_in + to_2tuple = self._ntuple(2) + bias = to_2tuple(bias) + linear_layer = nn.Linear + + self.fc1 = linear_layer(features_in, features_hidden, bias=bias[0]) + self.act = get_activation_fn(act_layer) + self.norm = norm_layer(features_hidden) if norm_layer is not None else nn.Identity() + self.fc2 = linear_layer(features_hidden, features_out, bias=bias[1]) + + def forward(self, x): + x = self.fc1(x) + x = self.act(x) + x = self.norm(x) + x = self.fc2(x) + return x + + def _ntuple(self, n): + def parse(x): + if isinstance(x, collections.abc.Iterable) and not isinstance(x, str): + return tuple(x) + return tuple(repeat(x, n)) + return parse \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-3/layers/norm.py b/MindIE/MultiModal/StableDiffusion-3/layers/norm.py new file mode 100644 index 0000000000000000000000000000000000000000..e371f299fe954c73c14f8423f928c9ce61985453 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-3/layers/norm.py @@ -0,0 +1,78 @@ +#!/usr/bin/env python +# coding=utf-8 +# Copyright 2024 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from typing import Tuple + +import torch +import torch.nn as nn + + +class AdaLayerNormZero(nn.Module): + r""" + Norm layer adaptive layer norm zero (adaLN-Zero). + + Parameters: + embedding_dim (`int`): The size of each embedding vector. + """ + + def __init__(self, embedding_dim: int, eps=1e-5, bias=True): + super().__init__() + self.embedding_dim = embedding_dim + self.eps = eps + self.silu = nn.SiLU() + self.linear = nn.Linear(embedding_dim, 6 * embedding_dim, bias=bias) + self.norm = nn.LayerNorm(embedding_dim, elementwise_affine=False, eps=1e-6) + + def forward( + self, + x: torch.Tensor, + emb: torch.Tensor, + ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]: + + emb = self.linear(self.silu(emb)) + + shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = emb.chunk(6, dim=1) + x = self.norm(x) * (1 + scale_msa[:, None]) + shift_msa[:, None] + + return x, gate_msa, shift_mlp, scale_mlp, gate_mlp + + +class AdaLayerNormContinuous(nn.Module): + def __init__( + self, + embedding_dim: int, + conditioning_embedding_dim: int, + eps=1e-5, + bias=True, + ): + super().__init__() + self.embedding_dim = embedding_dim + self.eps = eps + self.silu = nn.SiLU() + self.linear = nn.Linear(conditioning_embedding_dim, 2 * embedding_dim, bias=bias) + self.norm = nn.LayerNorm(embedding_dim, elementwise_affine=False, eps=eps) + + def forward( + self, + x: torch.Tensor, + conditioning_embedding: torch.Tensor, + ) -> torch.Tensor: + # conditioning_embedding float32, x maybe float16 bfloat16 + emb = self.linear(self.silu(conditioning_embedding).to(x.dtype)) + + scale, shift = torch.chunk(emb, 2, dim=1) + x = self.norm(x) * (1 + scale)[:, None, :] + shift[:, None, :] + return x \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-3/layers/resnet.py b/MindIE/MultiModal/StableDiffusion-3/layers/resnet.py new file mode 100644 index 0000000000000000000000000000000000000000..06f899ffef94e0a31952440b3ae0e79fb1d38e61 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-3/layers/resnet.py @@ -0,0 +1,138 @@ +#!/usr/bin/env python +# coding=utf-8 +# Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# This source code is licensed under the license found in the +# LICENSE file in the root directory of this source tree. + +from typing import Optional + +import torch +import torch.nn as nn + +from .activation import get_activation_fn +from .sampling import Downsample2D, Upsample2D +from ..utils.log import checkif_error + + +DEFAULT = "default" + + +class ResnetBlock2D(nn.Module): + + def __init__( + self, + *, + in_channels: int, + out_channels: Optional[int] = None, + conv_shortcut: bool = False, + temb_channels: int = 512, + groups: int = 32, + groups_out: Optional[int] = None, + pre_norm: bool = True, + eps: float = 1e-6, + non_linearity: str = "swish", + skip_time_act: bool = False, + time_embedding_norm: str = DEFAULT, # default, scale_shift, + output_scale_factor: float = 1.0, + use_in_shortcut: Optional[bool] = None, + conv_shortcut_bias: bool = True, + conv_2d_out_channels: Optional[int] = None, + ): + super().__init__() + + + self.pre_norm = True + self.in_channels = in_channels + out_channels = in_channels if out_channels is None else out_channels + self.out_channels = out_channels + self.use_conv_shortcut = conv_shortcut + self.output_scale_factor = output_scale_factor + self.time_embedding_norm = time_embedding_norm + self.skip_time_act = skip_time_act + + if groups_out is None: + groups_out = groups + + self.norm1 = torch.nn.GroupNorm(num_groups=groups, num_channels=in_channels, eps=eps, affine=True) + + self.conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=1, padding=1) + + if temb_channels is not None: + checkif_error(time_embedding_norm not in (DEFAULT, "scale_shift"), ValueError, + "This class only support `time_embedding_norm` is default or scale_shift") + if self.time_embedding_norm == DEFAULT: + self.time_emb_proj = nn.Linear(temb_channels, out_channels) + elif self.time_embedding_norm == "scale_shift": + self.time_emb_proj = nn.Linear(temb_channels, 2 * out_channels) + else: + self.time_emb_proj = None + + self.norm2 = torch.nn.GroupNorm(num_groups=groups_out, num_channels=out_channels, eps=eps, affine=True) + + conv_2d_out_channels = conv_2d_out_channels or out_channels + self.conv2 = nn.Conv2d(out_channels, conv_2d_out_channels, kernel_size=3, stride=1, padding=1) + + self.nonlinearity = get_activation_fn(non_linearity) + + self.use_in_shortcut = self.in_channels != conv_2d_out_channels + + self.conv_shortcut = None + if self.use_in_shortcut: + self.conv_shortcut = nn.Conv2d( + in_channels, + conv_2d_out_channels, + kernel_size=1, + stride=1, + padding=0, + bias=conv_shortcut_bias, + ) + + def forward(self, input_tensor: torch.Tensor, temb: torch.Tensor, *args, **kwargs) -> torch.Tensor: + + hidden_states = input_tensor + + hidden_states = self.norm1(hidden_states) + hidden_states = self.nonlinearity(hidden_states) + + hidden_states = self.conv1(hidden_states) + + if self.time_emb_proj is not None: + if not self.skip_time_act: + temb = self.nonlinearity(temb) + temb = self.time_emb_proj(temb)[:, :, None, None] + + if self.time_embedding_norm == "default": + if temb is not None: + hidden_states = hidden_states + temb + hidden_states = self.norm2(hidden_states) + elif self.time_embedding_norm == "scale_shift": + checkif_error(temb is None, ValueError, + f" `temb` should not be None when `time_embedding_norm` is {self.time_embedding_norm}") + + time_scale, time_shift = torch.chunk(temb, 2, dim=1) + hidden_states = self.norm2(hidden_states) + hidden_states = hidden_states * (1 + time_scale) + time_shift + else: + hidden_states = self.norm2(hidden_states) + + hidden_states = self.nonlinearity(hidden_states) + hidden_states = self.conv2(hidden_states) + + if self.conv_shortcut is not None: + input_tensor = self.conv_shortcut(input_tensor) + + output_tensor = (input_tensor + hidden_states) / self.output_scale_factor + return output_tensor \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-3/layers/sampling.py b/MindIE/MultiModal/StableDiffusion-3/layers/sampling.py new file mode 100644 index 0000000000000000000000000000000000000000..c6f7c0baac6922429955bd1acd234e881c307230 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-3/layers/sampling.py @@ -0,0 +1,162 @@ +#!/usr/bin/env python +# coding=utf-8 +# Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# This source code is licensed under the license found in the +# LICENSE file in the root directory of this source tree. + +from typing import Optional + +import torch +import torch.nn as nn +import torch.nn.functional as F +from ..utils.log import checkif_error + + +class Downsample2D(nn.Module): + + def __init__( + self, + channels: int, + use_conv: bool = False, + out_channels: Optional[int] = None, + padding: int = 1, + kernel_size=3, + norm_type=None, + eps=None, + elementwise_affine=None, + bias=True, + ): + super().__init__() + self.channels = channels + self.out_channels = out_channels or channels + self.use_conv = use_conv + self.padding = padding + stride = 2 + checkif_error(norm_type == "rms_norm", ValueError, + f"not support norm_type: {norm_type} yet") + checkif_error(norm_type and norm_type != "ln_norm", ValueError, + f"unknown norm_type: {norm_type}") + + if norm_type == "ln_norm": + self.norm = nn.LayerNorm(channels, eps, elementwise_affine) + + elif norm_type is None: + self.norm = None + + + if use_conv: + conv = nn.Conv2d( + self.channels, self.out_channels, kernel_size=kernel_size, stride=stride, padding=padding, bias=bias + ) + else: + checkif_error(self.channels != self.out_channels, ValueError, + "The parameter channels should be equal to out_channels") + conv = nn.AvgPool2d(kernel_size=stride, stride=stride) + + self.conv = conv + + def forward(self, hidden_states: torch.Tensor, *args, **kwargs) -> torch.Tensor: + checkif_error(hidden_states.shape[1] != self.channels, ValueError, + "The parameter hidden_states.shape[1] should equal to self.channels" + ) + + if self.norm is not None: + hidden_states = self.norm(hidden_states.permute(0, 2, 3, 1)).permute(0, 3, 1, 2) + + if self.use_conv and self.padding == 0: + pad = (0, 1, 0, 1) + hidden_states = F.pad(hidden_states, pad, mode="constant", value=0) + + hidden_states = self.conv(hidden_states) + + return hidden_states + + +class Upsample2D(nn.Module): + + def __init__( + self, + channels: int, + use_conv: bool = False, + use_conv_transpose: bool = False, + out_channels: Optional[int] = None, + kernel_size: Optional[int] = None, + padding=1, + norm_type=None, + eps=None, + elementwise_affine=None, + bias=True, + interpolate=True, + ): + super().__init__() + self.channels = channels + self.out_channels = out_channels or channels + self.use_conv = use_conv + self.use_conv_transpose = use_conv_transpose + self.interpolate = interpolate + checkif_error(norm_type == "rms_norm", ValueError, + f"not support norm_type: {norm_type} yet") + checkif_error(norm_type and norm_type != "ln_norm", ValueError, + f"unknown norm_type: {norm_type}") + + if norm_type == "ln_norm": + self.norm = nn.LayerNorm(channels, eps, elementwise_affine) + elif norm_type is None: + self.norm = None + + conv = None + if use_conv_transpose: + if kernel_size is None: + kernel_size = 4 + conv = nn.ConvTranspose2d( + channels, self.out_channels, kernel_size=kernel_size, stride=2, padding=padding, bias=bias + ) + elif use_conv: + if kernel_size is None: + kernel_size = 3 + conv = nn.Conv2d(self.channels, self.out_channels, kernel_size=kernel_size, padding=padding, bias=bias) + + self.conv = conv + + def forward(self, hidden_states: torch.Tensor, output_size: Optional[int] = None, *args, **kwargs) -> torch.Tensor: + + checkif_error(hidden_states.shape[1] != self.channels, ValueError, + "The parameter hidden_states.shape[1] should equal to self.channels" + ) + + if self.norm is not None: + hidden_states = self.norm(hidden_states.permute(0, 2, 3, 1)).permute(0, 3, 1, 2) + + if self.use_conv_transpose: + return self.conv(hidden_states) + + # Do not need cast to float32 as 'upsample_nearest2d_out_frame' op support bfloat16 after PyTorch 2.1 + # upsample_nearest_nhwc fails with large batch sizes. see https://github.com/huggingface/diffusers/issues/984 + if hidden_states.shape[0] >= 64: + hidden_states = hidden_states.contiguous() + + # if `output_size` is passed we force the interpolation output + # size and do not make use of `scale_factor=2` + if self.interpolate: + if output_size is None: + hidden_states = F.interpolate(hidden_states, scale_factor=2.0, mode="nearest") + else: + hidden_states = F.interpolate(hidden_states, size=output_size, mode="nearest") + + if self.use_conv: + hidden_states = self.conv(hidden_states) + + return hidden_states \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-3/layers/unet_2d_blocks.py b/MindIE/MultiModal/StableDiffusion-3/layers/unet_2d_blocks.py new file mode 100644 index 0000000000000000000000000000000000000000..6ddb8d25b165ba5fc3d6aeff7a229b081f0a0f5c --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-3/layers/unet_2d_blocks.py @@ -0,0 +1,290 @@ +#!/usr/bin/env python +# coding=utf-8 +# Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# This source code is licensed under the license found in the +# LICENSE file in the root directory of this source tree. + +from typing import Optional +from dataclasses import dataclass +import logging + +import torch +from torch import nn + +from .attention import ReconstitutionAttention +from .resnet import Downsample2D, ResnetBlock2D, Upsample2D +from ..utils.log import checkif_error + +logging.basicConfig(level=logging.INFO) +logger = logging.getLogger(__name__) + + +@dataclass +class Blockconfig: + num_layers: int + in_channels: int + out_channels: int + temb_channels: int + resnet_eps: float + resnet_act_fn: str + add_upsample: bool = None + add_downsample: bool = None + num_attention_heads: Optional[int] = None + resnet_groups: Optional[int] = None + resnet_time_scale_shift: str = "default" + attention_head_dim: Optional[int] = None + downsample_padding: Optional[int] = None + + def __post_init__(self): + # If attn head dim is not defined, we default it to the number of heads + if self.attention_head_dim is None: + self.attention_head_dim = self.num_attention_heads + + +def get_down_block( + down_block_type: str, + blockconfig: Blockconfig, +): + checkif_error(not isinstance(blockconfig, Blockconfig), ValueError, + "Please use Blockconfig to pass the parameters") + + down_block_type = down_block_type[7:] if down_block_type.startswith("UNetRes") else down_block_type + if down_block_type == "DownEncoderBlock2D": + return DownEncoderBlock2D( + num_layers=blockconfig.num_layers, + in_channels=blockconfig.in_channels, + out_channels=blockconfig.out_channels, + add_downsample=blockconfig.add_downsample, + resnet_eps=blockconfig.resnet_eps, + resnet_act_fn=blockconfig.resnet_act_fn, + resnet_groups=blockconfig.resnet_groups, + downsample_padding=blockconfig.downsample_padding, + resnet_time_scale_shift=blockconfig.resnet_time_scale_shift, + ) + else: + msg = f"{down_block_type} does not support yet." + logger.error(msg) + raise ValueError(msg) + + +def get_up_block( + up_block_type: str, + blockconfig: dataclass, +) -> nn.Module: + + checkif_error(not isinstance(blockconfig, Blockconfig), ValueError, + "Please use Blockconfig to pass the parameters") + + if up_block_type == "UpDecoderBlock2D": + return UpDecoderBlock2D( + num_layers=blockconfig.num_layers, + in_channels=blockconfig.in_channels, + out_channels=blockconfig.out_channels, + add_upsample=blockconfig.add_upsample, + resnet_eps=blockconfig.resnet_eps, + resnet_act_fn=blockconfig.resnet_act_fn, + resnet_groups=blockconfig.resnet_groups, + resnet_time_scale_shift=blockconfig.resnet_time_scale_shift, + temb_channels=blockconfig.temb_channels, + ) + else: + msg = f"{up_block_type} does not support yet." + logger.error(msg) + raise ValueError(msg) + + +class UNetMidBlock2D(nn.Module): + + def __init__( + self, + in_channels: int, + temb_channels: int, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", # default, spatial # USE deflalt + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + attn_groups: Optional[int] = None, + resnet_pre_norm: bool = True, + add_attention: bool = True, + attention_head_dim: int = 1, + output_scale_factor: float = 1.0, + ): + super().__init__() + resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32) + self.add_attention = add_attention + + if attn_groups is None: + attn_groups = resnet_groups if resnet_time_scale_shift == "default" else None + + # there is always at least one resnet + resnets = [] + checkif_error(resnet_time_scale_shift == "spatial", ValueError, + "Now only support resnet_time_scale_shift = `default` (meaning group norm)") + for _ in range(num_layers + 1): + resnets.append(ResnetBlock2D( + in_channels=in_channels, + out_channels=in_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + )) + + if attention_head_dim is None: + attention_head_dim = in_channels + attentions = [] + for _ in range(num_layers): + if self.add_attention: + attentions.append( + ReconstitutionAttention( + in_channels, + num_heads=in_channels // attention_head_dim, + head_dim=attention_head_dim, + eps=resnet_eps, + num_norm_groups=attn_groups, + qkv_bias=True, + )) + else: + attentions.append(None) + + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + def forward(self, hidden_states: torch.Tensor, temb: Optional[torch.Tensor] = None) -> torch.Tensor: + hidden_states = self.resnets[0](hidden_states, temb) + input_ndim = hidden_states.ndim + if input_ndim == 4: + batch_size, channel, height, width = hidden_states.shape + for attn, resnet in zip(self.attentions, self.resnets[1:]): + if attn is not None: + if input_ndim == 4: + hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2) + hidden_states += attn(hidden_states, temb=temb) + if input_ndim == 4: + hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width) + hidden_states = resnet(hidden_states, temb) + + return hidden_states + + +class DownEncoderBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + output_scale_factor: float = 1.0, + add_downsample: bool = True, + downsample_padding: int = 1, + ): + super().__init__() + resnets = [] + checkif_error(resnet_time_scale_shift == "spatial", ValueError, + "Now only support resnet_time_scale_shift = `default` (meaning group norm)") + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=None, + eps=resnet_eps, + groups=resnet_groups, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + )) + + self.resnets = nn.ModuleList(resnets) + + if add_downsample: + self.downsamplers = nn.ModuleList([Downsample2D( + out_channels, use_conv=True, + out_channels=out_channels, padding=downsample_padding)]) + else: + self.downsamplers = None + + def forward(self, hidden_states: torch.Tensor, *args, **kwargs) -> torch.Tensor: + + for resnet in self.resnets: + hidden_states = resnet(hidden_states, temb=None) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states) + + return hidden_states + + +class UpDecoderBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", # default, spatial + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + output_scale_factor: float = 1.0, + add_upsample: bool = True, + temb_channels: Optional[int] = None, + ): + super().__init__() + resnets = [] + checkif_error(resnet_time_scale_shift == "spatial", ValueError, + "Now only support resnet_time_scale_shift = `default` (meaning group norm)") + for i in range(num_layers): + input_channels = in_channels if i == 0 else out_channels + resnets.append( + ResnetBlock2D( + in_channels=input_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + )) + self.resnets = nn.ModuleList(resnets) + + if add_upsample: + self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)]) + else: + self.upsamplers = None + + def forward(self, hidden_states: torch.Tensor, temb: Optional[torch.Tensor] = None) -> torch.Tensor: + for resnet in self.resnets: + hidden_states = resnet(hidden_states, temb=temb) + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states) + + return hidden_states \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-3/model_index.json b/MindIE/MultiModal/StableDiffusion-3/model_index.json new file mode 100644 index 0000000000000000000000000000000000000000..e0fd629134976228d070084cf5b1a9760809b3cb --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-3/model_index.json @@ -0,0 +1,39 @@ +{ + "_class_name": "StableDiffusion3Pipeline", + "scheduler": [ + "stable_diffusion3", + "FlowMatchEulerDiscreteScheduler" + ], + "text_encoder": [ + "transformers", + "CLIPTextModelWithProjection" + ], + "text_encoder_2": [ + "transformers", + "CLIPTextModelWithProjection" + ], + "text_encoder_3": [ + "transformers", + "T5EncoderModel" + ], + "tokenizer": [ + "transformers", + "CLIPTokenizer" + ], + "tokenizer_2": [ + "transformers", + "CLIPTokenizer" + ], + "tokenizer_3": [ + "transformers", + "T5TokenizerFast" + ], + "transformer": [ + "stable_diffusion3", + "SD3DiTModel" + ], + "vae": [ + "stable_diffusion3", + "AutoencoderKL" + ] + } diff --git a/MindIE/MultiModal/StableDiffusion-3/models/dit.py b/MindIE/MultiModal/StableDiffusion-3/models/dit.py new file mode 100644 index 0000000000000000000000000000000000000000..17222dec8d211c056a1c4d76de2345567c988b06 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-3/models/dit.py @@ -0,0 +1,266 @@ +#!/usr/bin/env python +# coding=utf-8 +# Copyright 2024 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from typing import Optional, Dict, List, Any + +import torch +import torch.nn as nn +from mindiesd import DiffusionModel +from mindiesd import ConfigMixin + +from ..layers.attention import ReconstitutionAttention, JointAttnProcessor +from ..layers.norm import AdaLayerNormZero, AdaLayerNormContinuous +from ..layers.embedding import PatchEmbed2D, CombinedTimestepTextProjEmbeddings +from ..layers.mlp import Mlp + + + +class JointTransformerBlock(nn.Module): + def __init__(self, dim, num_attention_heads, attention_head_dim, context_pre_only=False) -> None: + super().__init__() + + + # norm hidden_states + self.norm1 = AdaLayerNormZero(dim) + self.context_pre_only = context_pre_only + # norm encoder_hidden_states + # only at the last block, the `context_pre_only` is true and `ada_norm_continous` is used + if context_pre_only: + self.norm1_context = AdaLayerNormContinuous(dim, dim, eps=1e-6, bias=True) + else: + self.norm1_context = AdaLayerNormZero(dim, eps=1e-6, bias=True) + + self.attn = ReconstitutionAttention( + attention_dim=dim, + cross_attention_dim=None, + num_heads=num_attention_heads, + head_dim=attention_head_dim, + qkv_bias=True, + out_proj_bias=True, + add_proj_dim=dim, + add_proj_bias=True, + enable_add_out_proj=True if not context_pre_only else False, + processor=JointAttnProcessor(), + ) + + self.mlp = Mlp(features_in=dim, features_hidden=4 * dim, features_out=dim, + act_layer="gelu-approximate") + + self.norm2 = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-6) + if not context_pre_only: + self.norm2_context = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-6) + + self.mlp_context = Mlp(features_in=dim, features_hidden=4 * dim, features_out=dim, act_layer="gelu") + else: + self.norm2_context = None + self.mlp_context = None + + def forward( + self, hidden_states: torch.FloatTensor, encoder_hidden_states: torch.FloatTensor, temb: torch.FloatTensor + ): + norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(hidden_states, emb=temb) + + # norm encoder_hidden_states + if self.context_pre_only: + norm_encoder_hidden_states = self.norm1_context(encoder_hidden_states, temb) + else: + norm_encoder_hidden_states, c_gate_msa, c_shift_mlp, c_scale_mlp, c_gate_mlp = self.norm1_context( + encoder_hidden_states, emb=temb + ) + + attn_output, context_attn_output = self.attn( + hidden_states=norm_hidden_states, encoder_hidden_states=norm_encoder_hidden_states + ) + # geta + hidden_states += gate_msa.unsqueeze(1) * attn_output + + norm_hidden_states = self.norm2(hidden_states) + norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None] + + hidden_states += gate_mlp.unsqueeze(1) * self.mlp(norm_hidden_states) + + if self.context_pre_only: + # only for last layer + encoder_hidden_states = None + else: + encoder_hidden_states += c_gate_msa.unsqueeze(1) * context_attn_output + + norm_encoder_hidden_states = self.norm2_context(encoder_hidden_states) + norm_encoder_hidden_states = norm_encoder_hidden_states * (1 + c_scale_mlp[:, None]) + c_shift_mlp[:, None] + + context_mlp_output = self.mlp_context(norm_encoder_hidden_states) + encoder_hidden_states += c_gate_mlp.unsqueeze(1) * context_mlp_output + return encoder_hidden_states, hidden_states + + +class SD3DiTConfig(ConfigMixin): + config_name = 'config.json' + + def __init__( + self, + sample_size: int = 128, + patch_size: int = 2, + in_channels: int = 16, + num_layers: int = 18, + attention_head_dim: int = 64, + num_attention_heads: int = 18, + joint_attention_dim: int = 4096, + caption_projection_dim: int = 1152, + pooled_projection_dim: int = 2048, + out_channels: int = 16, + pos_embed_max_size: int = 96, + ) -> None: + super().__init__() + self.sample_size = sample_size + self.patch_size = patch_size + self.in_channels = in_channels + self.num_layers = num_layers + self.attention_head_dim = attention_head_dim + self.num_attention_heads = num_attention_heads + self.joint_attention_dim = joint_attention_dim + self.caption_projection_dim = caption_projection_dim + self.pooled_projection_dim = pooled_projection_dim + self.out_channels = out_channels + self.pos_embed_max_size = pos_embed_max_size + + +class SD3DiTModel(DiffusionModel): + config_class = SD3DiTConfig + + def __init__(self, config): + super().__init__(config) + self.config = config + self.hidden_dims = config.num_attention_heads * config.attention_head_dim + + self.pos_embed = PatchEmbed2D( + height=self.config.sample_size, + width=self.config.sample_size, + patch_size=self.config.patch_size, + in_channels=self.config.in_channels, + embed_dim=self.hidden_dims, + pos_embed_max_size=self.config.pos_embed_max_size, # hard-code for now. + ) + + self.time_text_embed = CombinedTimestepTextProjEmbeddings( + embedding_dim=self.hidden_dims, pooled_projection_dim=self.config.pooled_projection_dim + ) + + self.context_embedder = nn.Linear(self.config.joint_attention_dim, self.config.caption_projection_dim) + self.transformer_blocks = nn.ModuleList( + [ + JointTransformerBlock( + dim=self.hidden_dims, + num_attention_heads=self.config.num_attention_heads, + attention_head_dim=self.config.attention_head_dim, + context_pre_only=(i == self.config.num_layers - 1), + ) + for i in range(self.config.num_layers) + ] + ) + + self.norm_out = AdaLayerNormContinuous(self.hidden_dims, self.hidden_dims, eps=1e-6) + + self.out_channels = config.out_channels if config.out_channels is not None else config.in_channels + self.proj_out = nn.Linear(self.hidden_dims, + self.config.patch_size * self.config.patch_size * self.out_channels, + bias=True) + + def forward( + self, + hidden_states: torch.FloatTensor, + encoder_hidden_states: torch.FloatTensor = None, + pooled_projections: torch.FloatTensor = None, + timestep: torch.LongTensor = None, + ) -> torch.FloatTensor: + + height, width = hidden_states.shape[-2:] + + hidden_states = self.pos_embed(hidden_states) # takes care of adding positional embeddings too. + temb = self.time_text_embed(timestep, pooled_projections) + encoder_hidden_states = self.context_embedder(encoder_hidden_states) + + for block in self.transformer_blocks: + + encoder_hidden_states, hidden_states = block( + hidden_states=hidden_states, encoder_hidden_states=encoder_hidden_states, temb=temb + ) + + hidden_states = self.norm_out(hidden_states, temb) + hidden_states = self.proj_out(hidden_states) + # unpatchify + patch_size = self.config.patch_size + height = height // patch_size + width = width // patch_size + + hidden_states = hidden_states.reshape( + shape=(hidden_states.shape[0], height, width, patch_size, patch_size, self.out_channels) + ) + hidden_states = torch.einsum("nhwpqc->nchpwq", hidden_states) + output = hidden_states.reshape( + shape=(hidden_states.shape[0], self.out_channels, height * patch_size, width * patch_size) + ) + return (output,) + + def _load_weights(self, state_dict): + weights = state_dict + + def weight_switch(prefix_key, new_weight, old_weight): + weights[prefix_key + new_weight] = weights.pop(prefix_key + old_weight) + + # attention_block: + num_layers = self.config.num_layers + for i in range(num_layers): + prefix_key = 'transformer_blocks.' + str(i) + '.' + to_q_weight = weights.pop(prefix_key + 'attn.to_q.weight') + to_q_bias = weights.pop(prefix_key + 'attn.to_q.bias') + to_k_weight = weights.pop(prefix_key + 'attn.to_k.weight') + to_k_bias = weights.pop(prefix_key + 'attn.to_k.bias') + to_v_weight = weights.pop(prefix_key + 'attn.to_v.weight') + to_v_bias = weights.pop(prefix_key + 'attn.to_v.bias') + weights[prefix_key + 'attn.qkv_proj.weight'] = torch.cat([to_q_weight, to_k_weight, to_v_weight], dim=0) + weights[prefix_key + 'attn.qkv_proj.bias'] = torch.cat([to_q_bias, to_k_bias, to_v_bias], dim=0) + to_q_weight = weights.pop(prefix_key + 'attn.add_q_proj.weight') + to_q_bias = weights.pop(prefix_key + 'attn.add_q_proj.bias') + to_k_weight = weights.pop(prefix_key + 'attn.add_k_proj.weight') + to_k_bias = weights.pop(prefix_key + 'attn.add_k_proj.bias') + to_v_weight = weights.pop(prefix_key + 'attn.add_v_proj.weight') + to_v_bias = weights.pop(prefix_key + 'attn.add_v_proj.bias') + weights[prefix_key + 'attn.add_qkv_proj.weight'] = torch.cat([to_q_weight, to_k_weight, to_v_weight], dim=0) + weights[prefix_key + 'attn.add_qkv_proj.bias'] = torch.cat([to_q_bias, to_k_bias, to_v_bias], dim=0) + + weight_switch(prefix_key, 'attn.out_proj.weight', 'attn.to_out.0.weight') + weight_switch(prefix_key, 'attn.out_proj.bias', 'attn.to_out.0.bias') + if not i == (self.config.num_layers - 1): + weight_switch(prefix_key, 'attn.add_out_proj.weight', 'attn.to_add_out.weight') + weight_switch(prefix_key, 'attn.add_out_proj.bias', 'attn.to_add_out.bias') + + weight_switch(prefix_key, 'mlp_context.fc1.weight', 'ff_context.net.0.proj.weight') + weight_switch(prefix_key, 'mlp_context.fc1.bias', 'ff_context.net.0.proj.bias') + weight_switch(prefix_key, 'mlp_context.fc2.weight', 'ff_context.net.2.weight') + weight_switch(prefix_key, 'mlp_context.fc2.bias', 'ff_context.net.2.bias') + + weight_switch(prefix_key, 'mlp.fc1.weight', 'ff.net.0.proj.weight') + weight_switch(prefix_key, 'mlp.fc1.bias', 'ff.net.0.proj.bias') + weight_switch(prefix_key, 'mlp.fc2.weight', 'ff.net.2.weight') + weight_switch(prefix_key, 'mlp.fc2.bias', 'ff.net.2.bias') + + for prefix_key in ['time_text_embed.timestep_embedder.', 'time_text_embed.text_embedder.']: + weight_switch(prefix_key, 'fc1.weight', 'linear_1.weight') + weight_switch(prefix_key, 'fc1.bias', 'linear_1.bias') + weight_switch(prefix_key, 'fc2.weight', 'linear_2.weight') + weight_switch(prefix_key, 'fc2.bias', 'linear_2.bias') + + self.load_state_dict(weights) \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-3/models/vae.py b/MindIE/MultiModal/StableDiffusion-3/models/vae.py new file mode 100644 index 0000000000000000000000000000000000000000..4de835cebb14a81e811d7a4a4d9155eb1d668437 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-3/models/vae.py @@ -0,0 +1,385 @@ +#!/usr/bin/env python +# coding=utf-8 +# Copyright 2024 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# This source code is licensed under the license found in the +# LICENSE file in the root directory of this source tree. +# -------------------------------------------------------- +# References: +# Open-Sora : https://github.com/hpcaitech/Open-Sora/blob/main/opensora/models/vae/vae.py +# -------------------------------------------------------- + +from typing import Dict, Optional, Tuple + +import logging +import os +import inspect +import torch +import torch.nn as nn + +from mindiesd import DiffusionModel +from mindiesd import ConfigMixin + +from ..layers.unet_2d_blocks import ( + UNetMidBlock2D, + get_down_block, + get_up_block, + Blockconfig, +) +from ..utils.log import checkif_error + + +class DiagonalGaussianDistribution(object): + def __init__( + self, + parameters, + deterministic=False, + ): + self.parameters = parameters + self.mean, self.logvar = torch.chunk(parameters, 2, dim=1) + self.logvar = torch.clamp(self.logvar, -30.0, 20.0) + self.deterministic = deterministic + self.std = torch.exp(0.5 * self.logvar) + self.var = torch.exp(self.logvar) + if self.deterministic: + self.var = self.std = torch.zeros_like(self.mean).to(device=self.parameters.device, dtype=self.mean.dtype) + + def sample(self): + # torch.randn: standard normal distribution + x = self.mean + self.std * torch.randn(self.mean.shape).to(device=self.parameters.device, dtype=self.mean.dtype) + return x + + def mode(self): + return self.mean + + +class AutoencoderKLConfig(ConfigMixin): + config_name = 'config.json' + + def __init__( + self, + in_channels: int = 3, + out_channels: int = 3, + down_block_types: Tuple[str] = ("DownEncoderBlock2D",), + up_block_types: Tuple[str] = ("UpDecoderBlock2D",), + block_out_channels: Tuple[int] = (64,), + layers_per_block: int = 1, + act_fn: str = "silu", + latent_channels: int = 4, + norm_num_groups: int = 32, + sample_size: int = 32, + scaling_factor: float = 0.18215, + shift_factor: Optional[float] = None, + latents_mean: Optional[Tuple[float]] = None, + latents_std: Optional[Tuple[float]] = None, + force_upcast: float = True, + use_quant_conv: bool = True, + use_post_quant_conv: bool = True, + ): + super().__init__() + + self.in_channels = in_channels + self.out_channels = out_channels + self.down_block_types = down_block_types + self.up_block_types = up_block_types + self.block_out_channels = block_out_channels + self.layers_per_block = layers_per_block + self.act_fn = act_fn + self.latent_channels = latent_channels + self.norm_num_groups = norm_num_groups + self.sample_size = sample_size + self.scaling_factor = scaling_factor + self.shift_factor = shift_factor + self.latents_mean = latents_mean + self.latents_std = latents_std + self.force_upcast = force_upcast + self.use_quant_conv = use_quant_conv + self.use_post_quant_conv = use_post_quant_conv + + +class AutoencoderKL(DiffusionModel): + + config_class = AutoencoderKLConfig + + def __init__(self, config: AutoencoderKLConfig): + super().__init__(config) + + # Initialize Encoder with config parameters + self.encoder = Encoder2D( + in_channels=config.in_channels, + out_channels=config.latent_channels, + down_block_types=config.down_block_types, + block_out_channels=config.block_out_channels, + layers_per_block=config.layers_per_block, + act_fn=config.act_fn, + norm_num_groups=config.norm_num_groups, + double_z=True, + ) + + # Initialize Decoder with config parameters + self.decoder = Decoder2D( + in_channels=config.latent_channels, + out_channels=config.out_channels, + up_block_types=config.up_block_types, + block_out_channels=config.block_out_channels, + layers_per_block=config.layers_per_block, + norm_num_groups=config.norm_num_groups, + act_fn=config.act_fn, + ) + + self.quant_conv = None + self.post_quant_conv = None + if config.use_quant_conv: + self.quant_conv = nn.Conv2d(2 * config.latent_channels, 2 * config.latent_channels, 1) + if config.use_post_quant_conv: + self.post_quant_conv = nn.Conv2d(config.latent_channels, config.latent_channels, 1) + + # Tiling configuration + self.tile_sample_min_size = config.sample_size + sample_size = config.sample_size[0] if isinstance(config.sample_size, (list, tuple)) else config.sample_size + self.tile_latent_min_size = int(sample_size / (2 ** (len(config.block_out_channels) - 1))) + self.tile_overlap_factor = 0.25 + + def encode(self, x: torch.Tensor): + + h = self.encoder(x) + + if self.quant_conv is not None: + moments = self.quant_conv(h) + else: + moments = h + + posterior = DiagonalGaussianDistribution(moments) + + return (posterior,) + + def decode(self, z: torch.FloatTensor): + + if self.post_quant_conv is not None: + z = self.post_quant_conv(z) + + decoded = self.decoder(z) + return (decoded,) + + def _load_weights(self, state_dict): + weights = state_dict + # attention_block: + for i in ["encoder", "decoder"]: + prefix_key = i + '.mid_block.attentions.0.' + to_q_weight = weights.pop(prefix_key + 'to_q.weight') + to_q_bias = weights.pop(prefix_key + 'to_q.bias') + to_k_weight = weights.pop(prefix_key + 'to_k.weight') + to_k_bias = weights.pop(prefix_key + 'to_k.bias') + to_v_weight = weights.pop(prefix_key + 'to_v.weight') + to_v_bias = weights.pop(prefix_key + 'to_v.bias') + weights[prefix_key + 'qkv_proj.weight'] = torch.cat([to_q_weight, to_k_weight, to_v_weight], dim=0) + weights[prefix_key + 'qkv_proj.bias'] = torch.cat([to_q_bias, to_k_bias, to_v_bias], dim=0) + weights[prefix_key + 'out_proj.weight'] = weights.pop(prefix_key + 'to_out.0.weight') + weights[prefix_key + 'out_proj.bias'] = weights.pop(prefix_key + 'to_out.0.bias') + self.load_state_dict(weights) + + +class Decoder2D(nn.Module): + + def __init__( + self, + in_channels: int = 3, + out_channels: int = 3, + up_block_types: Tuple[str, ...] = ("UpDecoderBlock2D",), + block_out_channels: Tuple[int, ...] = (64,), + layers_per_block: int = 2, + norm_num_groups: int = 32, + act_fn: str = "silu", + norm_type: str = "group", # group, spatial + mid_block_add_attention=True, + ): + super().__init__() + self.layers_per_block = layers_per_block + + self.conv_in = nn.Conv2d( + in_channels, + block_out_channels[-1], + kernel_size=3, + stride=1, + padding=1, + ) + + self.up_blocks = nn.ModuleList([]) + + temb_channels = in_channels if norm_type == "spatial" else None + + # mid + self.mid_block = UNetMidBlock2D( + in_channels=block_out_channels[-1], + resnet_eps=1e-6, + resnet_act_fn=act_fn, + output_scale_factor=1, + resnet_time_scale_shift="default" if norm_type == "group" else norm_type, + attention_head_dim=block_out_channels[-1], + resnet_groups=norm_num_groups, + temb_channels=temb_channels, + add_attention=mid_block_add_attention, + ) + + # up + reversed_block_out_channels = list(reversed(block_out_channels)) + output_channel = reversed_block_out_channels[0] + for i, up_block_type in enumerate(up_block_types): + prev_output_channel = output_channel + output_channel = reversed_block_out_channels[i] + + is_final_block = i == len(block_out_channels) - 1 + blockconfig = Blockconfig( + num_layers=self.layers_per_block + 1, + in_channels=prev_output_channel, + out_channels=output_channel, + add_upsample=not is_final_block, + resnet_eps=1e-6, + resnet_act_fn=act_fn, + resnet_groups=norm_num_groups, + attention_head_dim=output_channel, + temb_channels=temb_channels, + resnet_time_scale_shift=norm_type, + ) + up_block = get_up_block(up_block_type, blockconfig) + self.up_blocks.append(up_block) + prev_output_channel = output_channel + + # out + checkif_error(norm_type == "spatial", ValueError, "Now not support norm type==`spatial`") + self.conv_norm_out = nn.GroupNorm(num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=1e-6) + self.conv_act = nn.SiLU() + self.conv_out = nn.Conv2d(block_out_channels[0], out_channels, 3, padding=1) + + self.gradient_checkpointing = False + + def forward( + self, + sample: torch.Tensor, + latent_embeds: Optional[torch.Tensor] = None, + ) -> torch.Tensor: + r"""The forward method of the `Decoder` class.""" + + sample = self.conv_in(sample) + + upscale_dtype = next(iter(self.up_blocks.parameters())).dtype + + # middle + sample = self.mid_block(sample, latent_embeds) + sample = sample.to(upscale_dtype) + + # up + for up_block in self.up_blocks: + sample = up_block(sample, latent_embeds) + + # post-process + if latent_embeds is None: + sample = self.conv_norm_out(sample) + else: + sample = self.conv_norm_out(sample, latent_embeds) + sample = self.conv_act(sample) + sample = self.conv_out(sample) + + return sample + + +class Encoder2D(nn.Module): + + def __init__( + self, + in_channels: int = 3, + out_channels: int = 3, + down_block_types: Tuple[str, ...] = ("DownEncoderBlock2D",), + block_out_channels: Tuple[int, ...] = (64,), + layers_per_block: int = 2, + norm_num_groups: int = 32, + act_fn: str = "silu", + double_z: bool = True, + mid_block_add_attention=True, + ): + super().__init__() + self.layers_per_block = layers_per_block + + self.conv_in = nn.Conv2d( + in_channels, + block_out_channels[0], + kernel_size=3, + stride=1, + padding=1, + ) + + self.down_blocks = nn.ModuleList([]) + + # down + output_channel = block_out_channels[0] + for i, down_block_type in enumerate(down_block_types): + input_channel = output_channel + output_channel = block_out_channels[i] + is_final_block = i == len(block_out_channels) - 1 + blockconfig = Blockconfig( + num_layers=self.layers_per_block, + in_channels=input_channel, + out_channels=output_channel, + add_downsample=not is_final_block, + resnet_eps=1e-6, + downsample_padding=0, + resnet_act_fn=act_fn, + resnet_groups=norm_num_groups, + attention_head_dim=output_channel, + temb_channels=None, + ) + down_block = get_down_block(down_block_type, blockconfig) + self.down_blocks.append(down_block) + + # mid + self.mid_block = UNetMidBlock2D( + in_channels=block_out_channels[-1], + resnet_eps=1e-6, + resnet_act_fn=act_fn, + output_scale_factor=1, + resnet_time_scale_shift="default", + attention_head_dim=block_out_channels[-1], + resnet_groups=norm_num_groups, + temb_channels=None, + add_attention=mid_block_add_attention, + ) + + # out + self.conv_norm_out = nn.GroupNorm(num_channels=block_out_channels[-1], num_groups=norm_num_groups, eps=1e-6) + self.conv_act = nn.SiLU() + + conv_out_channels = 2 * out_channels if double_z else out_channels + self.conv_out = nn.Conv2d(block_out_channels[-1], conv_out_channels, 3, padding=1) + + self.gradient_checkpointing = False + + def forward(self, sample: torch.Tensor) -> torch.Tensor: + r"""The forward method of the `Encoder` class.""" + + sample = self.conv_in(sample) + + # down + for down_block in self.down_blocks: + sample = down_block(sample) + + # middle + sample = self.mid_block(sample) + + # post-process + sample = self.conv_norm_out(sample) + sample = self.conv_act(sample) + sample = self.conv_out(sample) + + return sample \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-3/pipepline/pipeline_utils.py b/MindIE/MultiModal/StableDiffusion-3/pipepline/pipeline_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..0447763952a430c630694653fafb62737d3aa071 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-3/pipepline/pipeline_utils.py @@ -0,0 +1,77 @@ +#!/usr/bin/env python +# coding=utf-8 +# Copyright 2024 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +import os +import inspect +import importlib + +from tqdm import tqdm +from mindiesd import ConfigMixin + + +PIPELINE_CONFIG_NAME = "model_index.json" +VAE = 'vae' +TEXT_ENCODER = 'text_encoder' +TOKENIZER = 'tokenizer' +TRANSFORMER = 'transformer' +SCHEDULER = 'scheduler' + + +class DiffusionPipeline(ConfigMixin): + config_name = PIPELINE_CONFIG_NAME + + def __init__(self): + super().__init__() + + @classmethod + def from_pretrained(cls, model_path, **kwargs): + dtype = kwargs.pop('dtype', None) + real_path = os.path.abspath(model_path) + if not (os.path.exists(real_path) and os.path.isdir(real_path)): + raise ValueError("model path is invalid!") + + init_dict, config_dict = cls.load_config(real_path, **kwargs) + + all_parameters = inspect.signature(cls.__init__).parameters + required_param = {k: v for k, v in all_parameters.items() if v.default is inspect.Parameter.empty} + + # init the module from kwargs + passed_module = {k: kwargs.pop(k) for k in required_param if k in kwargs} + from_diffusers = None if '_diffusers_version' not in config_dict else config_dict['_diffusers_version'] + for key, item in tqdm(init_dict.items(), desc="Loading pipeline components..."): + if key in passed_module: + init_dict[key] = passed_module.pop(key) + else: + modules, cls_name = item + if from_diffusers is not None: + try: + library = importlib.import_module("mindiesd") + except ImportError: + logger.warning(f"cannot import {cls_name} from mindiesd. Use diffuser.") + library = importlib.import_module(modules) + else: + library = importlib.import_module(modules) + class_obj = getattr(library, cls_name) + sub_folder = os.path.join(real_path, key) + if key.startswith(TOKENIZER): + init_dict[key] = class_obj.from_pretrained(sub_folder, **kwargs) + elif key.startswith(SCHEDULER): + init_dict[key] = class_obj.from_config(sub_folder, **kwargs) + else: + init_dict[key] = class_obj.from_pretrained(sub_folder, **kwargs).to(dtype) + + return cls(**init_dict) \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-3/pipepline/stable_diffusion_pipeline.py b/MindIE/MultiModal/StableDiffusion-3/pipepline/stable_diffusion_pipeline.py new file mode 100644 index 0000000000000000000000000000000000000000..4fe009ebcae7a81d52a43c64404bf153322ab457 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-3/pipepline/stable_diffusion_pipeline.py @@ -0,0 +1,673 @@ +# Copyright 2024 Stability AI and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable, Dict, List, Optional, Union, Tuple +import logging +import collections +from tqdm import tqdm + +import torch +from transformers import ( + CLIPTextModelWithProjection, + CLIPTokenizer, + T5EncoderModel, + T5TokenizerFast, +) +import numpy as np +import PIL.Image +from PIL import Image +from .pipeline_utils import DiffusionPipeline + +from ..models.vae import AutoencoderKL +from ..models.dit import SD3DiTModel +from ..schedulers.flow_match_euler_discrete import FlowMatchEulerDiscreteScheduler +from ..utils.log import checkif_error + +logging.basicConfig(level=logging.INFO) +logger = logging.getLogger(__name__) + +IMAGE_TYPE = "pil" +TENSOR_TYPE = "pt" + + +def retrieve_timesteps( + scheduler, + num_inference_steps: Optional[int] = None, + device: Optional[Union[str, torch.device]] = None, + timesteps: Optional[List[int]] = None, + sigmas: Optional[List[float]] = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`List[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`List[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + checkif_error(timesteps is not None and sigmas is not None, ValueError, + "Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values" + ) + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + checkif_error(not accepts_timesteps, ValueError, + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler.") + + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + checkif_error(not accept_sigmas, ValueError, + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler.") + + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class StableDiffusion3Pipeline(DiffusionPipeline): + config_name = "model_index.json" + + model_cpu_offload_seq = "text_encoder->text_encoder_2->text_encoder_3->transformer->vae" + _optional_components = [] + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds", "negative_pooled_prompt_embeds"] + + def __init__( + self, + transformer: SD3DiTModel, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKL, + text_encoder: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + text_encoder_2: CLIPTextModelWithProjection, + tokenizer_2: CLIPTokenizer, + text_encoder_3: T5EncoderModel, + tokenizer_3: T5TokenizerFast, + ): + super().__init__() + + self.vae = vae + self.text_encoder = text_encoder + self.text_encoder_2 = text_encoder_2 + self.text_encoder_3 = text_encoder_3 + self.tokenizer = tokenizer + self.tokenizer_2 = tokenizer_2 + self.tokenizer_3 = tokenizer_3 + self.transformer = transformer + self.scheduler = scheduler + + self.vae_scale_factor = ( + 2 ** (len(self.vae.config.block_out_channels) - 1) if hasattr(self, "vae") and self.vae is not None else 8 + ) + self.tokenizer_max_length = ( + self.tokenizer.model_max_length if hasattr(self, "tokenizer") and self.tokenizer is not None else 77 + ) + self.default_sample_size = ( + self.transformer.config.sample_size + if hasattr(self, "transformer") and self.transformer is not None + else 128 + ) + self.patch_size = ( + self.transformer.config.patch_size if hasattr(self, "transformer") and self.transformer is not None else 2 + ) + self._init_property() + + @torch.no_grad() + def __call__( + self, + prompt: Union[str, List[str]] = None, + prompt_2: Optional[Union[str, List[str]]] = None, + prompt_3: Optional[Union[str, List[str]]] = None, + height: Optional[int] = None, + width: Optional[int] = None, + num_inference_steps: int = 28, + timesteps: List[int] = None, + guidance_scale: float = 7.0, + negative_prompt: Optional[Union[str, List[str]]] = None, + negative_prompt_2: Optional[Union[str, List[str]]] = None, + negative_prompt_3: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + latents: Optional[torch.FloatTensor] = None, + prompt_embeds: Optional[torch.FloatTensor] = None, + negative_prompt_embeds: Optional[torch.FloatTensor] = None, + pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + output_type: Optional[str] = "pil", + clip_skip: Optional[int] = None, + max_sequence_length: int = 77 + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + will be used instead + prompt_3 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to `tokenizer_3` and `text_encoder_3`. If not defined, `prompt` is + will be used instead + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`List[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + guidance_scale (`float`, *optional*, defaults to 7.0): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used instead + negative_prompt_3 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_3` and + `text_encoder_3`. If not defined, `negative_prompt` is used instead + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will randomly generated by sampling. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] instead + of a plain tuple. + max_sequence_length (`int` defaults to 256): Maximum sequence length to use with the `prompt`. + + + Returns: + `tuple`: a `tuple`. When returning a tuple, the first element is a list with the generated images. + """ + + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + image_size = (height, width) + prompts = (prompt, prompt_2, prompt_3) + negative_prompts = (negative_prompt, negative_prompt_2, negative_prompt_3) + all_embeds = (prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds) + + # 1. Check inputs. Raise error if not correct + self.check_inputs(prompts, image_size, negative_prompts, all_embeds, + max_sequence_length=max_sequence_length,) + + self._init_property(guidance_scale=guidance_scale, clip_skip=clip_skip, + num_images_per_prompt=num_images_per_prompt, max_sequence_length=max_sequence_length) + + + # 2. Define call parameters + if prompt is not None: + batch_size = 1 if isinstance(prompt, str) else len(prompt) + else: + batch_size = prompt_embeds.shape[0] + device = self._execution_device + + ( + prompt_embeds, negative_prompt_embeds, + pooled_prompt_embeds, negative_pooled_prompt_embeds + ) = self._encode_prompt( + prompts=prompts, negative_prompts=negative_prompts, + do_classifier_free_guidance=self.do_classifier_free_guidance, + all_embeds=all_embeds) + + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + pooled_prompt_embeds = torch.cat([negative_pooled_prompt_embeds, pooled_prompt_embeds], dim=0) + + # 4. Prepare timesteps + timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps) + self._num_timesteps = len(timesteps) + + # 5. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels + latents = self._prepare_latents(batch_size * num_images_per_prompt, num_channels_latents, + image_size, prompt_embeds.dtype, latents) + + # 6. Denoising loop + latents = self._sampling(latents, timesteps, prompt_embeds, pooled_prompt_embeds) + + # 7. Postprocessing latents data + image = self._postprocess(latents, output_type=output_type) + return (image,) + + @property + def _execution_device(self): + return next(self.transformer.parameters()).device + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def clip_skip(self): + return self._clip_skip + + @property + def num_images_per_prompt(self): + return self._num_images_per_prompt + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def max_sequence_length(self): + return self._max_sequence_length + + def check_inputs(self, prompts, image_size, negative_prompts, all_embeds, max_sequence_length=None): + + prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds = all_embeds + + checkif_error(image_size[0] % (self.vae_scale_factor * self.patch_size) != 0 + or image_size[1] % (self.vae_scale_factor * self.patch_size) != 0, + ValueError, + f"`height` and `width` have to be divisible by {self.vae_scale_factor * self.patch_size} " + f"but are {image_size[0]} and {image_size[1]}." + ) + checkif_error(prompts[0] is None and prompt_embeds is None, + ValueError, + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined.") + + for i, prompt in enumerate(prompts): + prompt_str = "prompt" if i == 0 else f"prompt_{i + 1}" + checkif_error(prompt is not None and prompt_embeds is not None, + ValueError, + f"Cannot forward both `{prompt_str}` and `prompt_embeds`. " + "Please make sure to only forward one of the two.") + + checkif_error(prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)), + ValueError, + f"`{prompt_str}` has to be of type `str` or `list` but is {type(prompt)}") + + for i, negative_prompt in enumerate(negative_prompts): + negative_prompt_str = "negative_prompt" if i == 0 else f"negative_prompt_{i + 1}" + checkif_error(negative_prompt is not None and negative_prompt_embeds is not None, + ValueError, + f"Cannot forward both `{negative_prompt_str}` and `negative_prompt_embeds`" + "Please make sure to only forward one of the two.") + + if prompt_embeds is not None and negative_prompt_embeds is not None: + checkif_error(prompt_embeds.shape != negative_prompt_embeds.shape, ValueError, + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}.") + + checkif_error(prompt_embeds is not None and pooled_prompt_embeds is None, ValueError, + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed." + "Make sure to generate `pooled_prompt_embeds` from the same text encoder " + "that was used to generate `prompt_embeds`.") + + checkif_error(negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None, ValueError, + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed." + "Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder " + "that was used to generate `negative_prompt_embeds`.") + + checkif_error(max_sequence_length is not None and max_sequence_length > 512, ValueError, + f"`max_sequence_length` cannot be greater than 512 but is {max_sequence_length}") + + def to(self, *args, **kwargs): + self.text_encoder.to(*args, **kwargs) + self.transformer.to(*args, **kwargs) + self.vae.to(*args, **kwargs) + self.text_encoder_2.to(*args, **kwargs) + self.text_encoder_3.to(*args, **kwargs) + return self + + def _init_property(self, **kwargs): + self._guidance_scale = kwargs.get("guidance_scale", None) + self._clip_skip = kwargs.get("clip_skip", None) + self._num_images_per_prompt = kwargs.get("num_images_per_prompt", None) + self._max_sequence_length = kwargs.get("max_sequence_length", None) + self._num_timesteps = kwargs.get("num_timesteps", None) + + def _get_t5_prompt_embeds( + self, + prompt: Union[str, List[str]] = None, + num_images_per_prompt: int = 1, + max_sequence_length: int = 256, + dtype: Optional[torch.dtype] = None, + ): + device = self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if self.text_encoder_3 is None: + return torch.zeros( + ( + batch_size * num_images_per_prompt, + self.tokenizer_max_length, + self.transformer.config.joint_attention_dim, + ), + device=device, + dtype=dtype, + ) + + text_inputs = self.tokenizer_3( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + add_special_tokens=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer_3(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer_3.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + msg = "The following part of your input was truncated because `max_sequence_length` is set to " \ + f" {max_sequence_length} tokens: {removed_text}" + logger.warning(msg) + + prompt_embeds = self.text_encoder_3(text_input_ids.to(device))[0] + + dtype = self.text_encoder_3.dtype + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + _, seq_len, _ = prompt_embeds.shape + + # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + return prompt_embeds + + def _get_clip_prompt_embeds( + self, + prompt: Union[str, List[str]], + num_images_per_prompt: int = 1, + clip_skip: Optional[int] = None, + clip_model_index: int = 0, + ): + device = self._execution_device + + clip_tokenizers = [self.tokenizer, self.tokenizer_2] + clip_text_encoders = [self.text_encoder, self.text_encoder_2] + + tokenizer = clip_tokenizers[clip_model_index] + text_encoder = clip_text_encoders[clip_model_index] + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + text_inputs = tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer_max_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = tokenizer.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + msg = "The following part of your input was truncated because CLIP can only handle sequences up to" + \ + f" {self.tokenizer_max_length} tokens: {removed_text}" + logger.warning(msg) + + prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True) + pooled_prompt_embeds = prompt_embeds[0] + + if clip_skip is None: + prompt_embeds = prompt_embeds.hidden_states[-2] + else: + prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)] + + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device) + + _, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt, 1) + pooled_prompt_embeds = pooled_prompt_embeds.view(batch_size * num_images_per_prompt, -1) + + return prompt_embeds, pooled_prompt_embeds + + def _get_all_prompt_embeds( + self, + prompt, + prompt_2, + prompt_3, + ): + device = self._execution_device + num_images_per_prompt = self.num_images_per_prompt + clip_skip = self.clip_skip + max_sequence_length = self._max_sequence_length + prompt_embed, pooled_prompt_embed = self._get_clip_prompt_embeds( + prompt=prompt, + num_images_per_prompt=num_images_per_prompt, + clip_skip=clip_skip, + clip_model_index=0, + ) + + prompt_2_embed, pooled_prompt_2_embed = self._get_clip_prompt_embeds( + prompt=prompt_2, + num_images_per_prompt=num_images_per_prompt, + clip_skip=clip_skip, + clip_model_index=1, + ) + clip_prompt_embeds = torch.cat([prompt_embed, prompt_2_embed], dim=-1) + + t5_prompt_embed = self._get_t5_prompt_embeds( + prompt=prompt_3, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + ) + + clip_prompt_embeds = torch.nn.functional.pad( + clip_prompt_embeds, (0, t5_prompt_embed.shape[-1] - clip_prompt_embeds.shape[-1]) + ) + + prompt_embeds = torch.cat([clip_prompt_embeds, t5_prompt_embed], dim=-2) + pooled_prompt_embeds = torch.cat([pooled_prompt_embed, pooled_prompt_2_embed], dim=-1) + + return prompt_embeds, pooled_prompt_embeds + + def _encode_prompt( + self, + prompts: Union[Tuple[str], Tuple[List[str]]], + negative_prompts: Union[Tuple[str], Tuple[List[str]]], + all_embeds: Tuple[torch.FloatTensor], + do_classifier_free_guidance: bool = True, + ): + prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds = all_embeds + prompt, prompt_2, prompt_3 = prompts + negative_prompt, negative_prompt_2, negative_prompt_3 = negative_prompts + prompt = [prompt] if isinstance(prompt, str) else prompt + + batch_size = len(prompt) if prompt is not None else prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + prompt_3 = prompt_3 or prompt + prompt_3 = [prompt_3] if isinstance(prompt_3, str) else prompt_3 + + prompt_embeds, pooled_prompt_embeds = self._get_all_prompt_embeds( + prompt, prompt_2, prompt_3) + + if do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt_2 = negative_prompt_2 or negative_prompt + negative_prompt_3 = negative_prompt_3 or negative_prompt + + # normalize str to list + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + negative_prompt_2 = ( + batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2) + negative_prompt_3 = ( + batch_size * [negative_prompt_3] if isinstance(negative_prompt_3, str) else negative_prompt_3) + + checkif_error(prompt is not None and type(prompt) is not type(negative_prompt), + TypeError, + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}.") + + checkif_error(batch_size != len(negative_prompt), ValueError, + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`.") + + negative_prompt_embeds, negative_pooled_prompt_embeds = self._get_all_prompt_embeds( + negative_prompt, negative_prompt_2, negative_prompt_3) + + PromptEmbeddingInfo = collections.namedtuple("PromptEmbeddingInfo", + ["prompt_embeds", "negative_prompt_embeds", "pooled_prompt_embeds", "negative_pooled_prompt_embeds"]) + return PromptEmbeddingInfo(prompt_embeds, negative_prompt_embeds, + pooled_prompt_embeds, negative_pooled_prompt_embeds) + + def _prepare_latents(self, batch_size, num_channels_latents, image_size, dtype, latents=None): + device = self._execution_device + if latents is not None: + return latents.to(device=device, dtype=dtype) + + height, width = image_size + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + + latents = torch.randn(shape, device=device, dtype=dtype, layout=torch.strided) + + return latents + + def _sampling(self, latents, timesteps, prompt_embeds, pooled_prompt_embeds): + for t in tqdm(timesteps): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latent_model_input.shape[0]) + + noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=prompt_embeds, + pooled_projections=pooled_prompt_embeds, + )[0] + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents)[0] + + if latents.dtype != latents_dtype: + raise ValueError("latents.dtype != latents_dtype") + return latents + + def _numpy_to_pil(self, images: np.ndarray) -> List[PIL.Image.Image]: + + if images.ndim == 3: + images = images[None, ...] + images = (images * 255).round().astype("uint8") + if images.shape[-1] == 1: + # special case for grayscale (single channel) images + pil_images = [Image.fromarray(image.squeeze(), mode="L") for image in images] + else: + pil_images = [Image.fromarray(image) for image in images] + + return pil_images + + def _postprocess( + self, + latents: torch.Tensor, + output_type: str = "pil", + ) -> Union[PIL.Image.Image, np.ndarray, torch.Tensor]: + + # IMAGE_TYPE = "pil" TENSOR_TYPE = "pt" + if output_type not in ["latent", IMAGE_TYPE, TENSOR_TYPE]: + output_type = IMAGE_TYPE + + if output_type == "latent": + image = latents + return image + else: + latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor + image = self.vae.decode(latents)[0] + image = (image / 2 + 0.5).clamp(0, 1) + if output_type == TENSOR_TYPE: + return image + + # output_type pil + else: + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return self._numpy_to_pil(image) \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-3/requirements.txt b/MindIE/MultiModal/StableDiffusion-3/requirements.txt new file mode 100644 index 0000000000000000000000000000000000000000..5a64f3d07d8631acd564348433dada4bf771578d --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-3/requirements.txt @@ -0,0 +1,16 @@ +torch==2.1.0 +diffusers==0.29.0 +transformers==4.44.2 +open_clip_torch==2.20.0 +av==12.0.0 +tqdm==4.66.1 +timm==0.9.12 +tensorboard==2.11.0 +pre-commit==3.8.0 +mmengine==0.10.4 +ftfy==6.1.3 +accelerate==0.26.1 +bs4 +torchvision==0.16.0 +einops +numpy==1.24.0 \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-3/schedulers/flow_match_euler_discrete.py b/MindIE/MultiModal/StableDiffusion-3/schedulers/flow_match_euler_discrete.py new file mode 100644 index 0000000000000000000000000000000000000000..122d6a205830bd3c7dff50535b807a3727230ef5 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-3/schedulers/flow_match_euler_discrete.py @@ -0,0 +1,140 @@ +# Copyright 2024 Stability AI, Katherine Crowson and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# This source code is licensed under the license found in the +# LICENSE file in the root directory of this source tree. + +from typing import Optional, Union + +import numpy as np +import torch +from mindiesd import DiffusionScheduler + +from ..utils.log import checkif_error + + +class FlowMatchEulerDiscreteScheduler(DiffusionScheduler): + + def __init__( + self, + num_train_timesteps: int = 1000, + shift: float = 1.0, + ): + super().__init__() + + timesteps = np.linspace(1, num_train_timesteps, num_train_timesteps, dtype=np.float32)[::-1].copy() + timesteps = torch.from_numpy(timesteps).to(dtype=torch.float32) + sigmas = timesteps / num_train_timesteps + sigmas = shift * sigmas / (1 + (shift - 1) * sigmas) + + self.timesteps = sigmas * num_train_timesteps + + self._step_index = None + self._begin_index = None + # check when parallism. to("npu") setting may cause some problem + self.sigmas = sigmas.to("npu") + self.sigma_min = self.sigmas[-1].item() + self.sigma_max = self.sigmas[0].item() + self.num_inference_steps = None + self._num_train_timesteps = num_train_timesteps + self._shift = shift + + + @property + def step_index(self): + return self._step_index + + @property + def begin_index(self): + return self._begin_index + + def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None): + self._step_index = None + self._begin_index = None + if self.num_inference_steps != num_inference_steps: + + self.num_inference_steps = num_inference_steps + + timesteps = np.linspace( + self._sigma_to_t(self.sigma_max), self._sigma_to_t(self.sigma_min), num_inference_steps + ) + + sigmas = timesteps / self._num_train_timesteps + sigmas = self._shift * sigmas / (1 + (self._shift - 1) * sigmas) + sigmas = torch.from_numpy(sigmas).to(dtype=torch.float32, device=device) + + timesteps = sigmas * self._num_train_timesteps + self.timesteps = timesteps.to(device=device) + self.sigmas = torch.cat([sigmas, torch.zeros(1, device=sigmas.device)]) + + + + def index_for_timestep(self, timestep, schedule_timesteps=None): + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + indices = (schedule_timesteps == timestep).nonzero() + + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + pos = 1 if len(indices) > 1 else 0 + + return indices[pos].item() + + def step( + self, + model_output: torch.FloatTensor, + timestep: Union[float, torch.FloatTensor], + sample: torch.FloatTensor, + ): + checkif_error((isinstance(timestep, int) + or isinstance(timestep, torch.IntTensor) + or isinstance(timestep, torch.LongTensor)), + ValueError, + "Passing integer indices (e.g. from `enumerate(timesteps)`) as timesteps to" + " `EulerDiscreteScheduler.step()` is not supported. Make sure to pass" + " one of the `scheduler.timesteps` as a timestep." + ) + + if self.step_index is None: + self._init_step_index(timestep) + + # Upcast to avoid precision issues when computing prev_sample + sample = sample.to(torch.float32) + + sigma = self.sigmas[self.step_index] + sigma_next = self.sigmas[self.step_index + 1] + + prev_sample = sample + (sigma_next - sigma) * model_output + + # Cast sample back to model compatible dtype + prev_sample = prev_sample.to(model_output.dtype) + + # upon completion increase step index by one + self._step_index += 1 + + return (prev_sample,) + + def _sigma_to_t(self, sigma): + return sigma * self._num_train_timesteps + + def _init_step_index(self, timestep): + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-3/utils/__init__.py b/MindIE/MultiModal/StableDiffusion-3/utils/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..f35da6dceac87094e7ca1c0afb182a8f119b8c36 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-3/utils/__init__.py @@ -0,0 +1 @@ +from .utils import set_random_seed \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-3/utils/log.py b/MindIE/MultiModal/StableDiffusion-3/utils/log.py new file mode 100644 index 0000000000000000000000000000000000000000..206932eff2c4f25f14d7165da7eed706e4d1d59c --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-3/utils/log.py @@ -0,0 +1,5 @@ +from typing import Type + +def checkif_error(exprs: bool, exception: Type[BaseException], info: str): + if exprs: + raise exception(info) \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-3/utils/utils.py b/MindIE/MultiModal/StableDiffusion-3/utils/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..8b702e9cc4a20e96fcb4fcf639d704030ce0ca3e --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-3/utils/utils.py @@ -0,0 +1,18 @@ + +import importlib +import random +import torch +import numpy as np + +def set_random_seed(seed): + """Set random seed. + + Args: + seed (int, optional): Seed to be used. + + """ + + random.seed(seed) + np.random.seed(seed) + torch.manual_seed(seed) + return seed \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-XL/README.md b/MindIE/MultiModal/StableDiffusion-XL/README.md new file mode 100644 index 0000000000000000000000000000000000000000..5a3866ce8d9f7386fa1918cead855d547fbec097 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-XL/README.md @@ -0,0 +1,244 @@ +--- +license: creativeml-openrail-m +library_name: openmind +frameworks: + - PyTorch +language: + - zh +hardwares: + - NPU +pipeline_tag: text-to-image +--- +# stable-diffusion-xl模型-推理指导 + +- [概述](#ZH-CN_TOPIC_0000001172161501) + + - [输入输出数据](#section540883920406) + +- [推理环境准备](#ZH-CN_TOPIC_0000001126281702) + +- [快速上手](#ZH-CN_TOPIC_0000001126281700) + + - [获取源码](#section4622531142816) + - [模型推理](#section741711594517) + + +# 概述 + +SDXL 由一组用于潜在扩散的专家管道组成： 在第一步中，使用基础模型生成（噪声）潜伏， 然后使用专门用于最终降噪步骤的细化模型[此处获得](https://huggingface.co/stabilityai/stable-diffusion-xl-refiner-1.0/) + +参考实现： +```bash +# StableDiffusionxl +https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0 +``` + +## 输入输出数据 + +- 输入数据 + + | 输入数据 | 大小 | 数据类型 | 数据排布格式 | + | -------- | -------- | ------------------------- | ------------ | + | prompt | batch x 77 | STRING | ND| + + +- 输出数据 + + | 输出数据 | 大小 | 数据类型 | 数据排布格式 | + | -------- | -------- | -------- | ------------ | + | output1 | batch x 3 x 1024 x 1024 | FLOAT32 | NCHW | + +# 推理环境准备 + +- 该模型需要以下插件与驱动 + + **表 1** 版本配套表 + | 配套 | 版本 | 环境准备指导 | + | ------------------------------------------------------------ |--------| ------------------------------------------------------------ | + | Python | 3.10.x | - | + | torch| 2.1.0 | - | + +该模型性能受CPU规格影响，建议使用64核CPU（arm）以复现性能 + +# 快速上手 + +## 获取源码 + +0. 下载仓库到本地。 + ```bash + git clone https://gitee.com/ascend/ModelZoo-PyTorch.git + ``` + +1. 安装依赖。 + ```bash + pip3 install -r requirements.txt + + # 若要使用hpsv2验证精度，则还需要按照以下步骤安装hpsv2 + git clone https://github.com/tgxs002/HPSv2.git + cd HPSv2 + pip3 install -e . + ``` + +2. 安装mindie包 + + ```bash + # 安装mindie + chmod +x ./Ascend-mindie_xxx.run + ./Ascend-mindie_xxx.run --install + source /usr/local/Ascend/mindie/set_env.sh + ``` + +## 准备数据集 + +1. 获取原始数据集。 + + 本模型输入文本信息生成图片，无需数据集。 + + +## 模型推理 + +1. 获取权重（可选） + + 可提前下载权重，放到代码同级目录下，以避免执行后面步骤时可能会出现下载失败。 + + ```bash + # 需要使用 git-lfs (https://git-lfs.com) + git lfs install + + # xl + git clone https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0 + ``` + +2. 开始推理验证。 + + 1. 开启cpu高性能模式 + ```bash + echo performance |tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor + sysctl -w vm.swappiness=0 + sysctl -w kernel.numa_balancing=0 + ``` + + 2. 设置调度流水优化 + ```bash + export TASK_QUEUE_ENABLE=2 + ``` + + 3. 执行推理脚本。 + ```bash + xl (使用上一步下载的权重) + model_base="./stable-diffusion-xl-base-1.0" + ``` + + 执行命令 + ```bash + python3 infer_pipe.py --path=${model_base} --prompt="A dog" --device_id=0 --height=1024 --width=1024 --num_images_per_prompt=1 + ``` + + 参数说明： + - --path：模型权重路径 + - --prompt: 输入的prompt + - --device_id：推理设备ID + - --height: 生成图片的高 + - --width: 生成图片的宽 + - --num_images_per_prompt: 单个prompt生成的图片数量 + +3. 性能验证。 + + ```bash + python3 ./test_performance.py --path=${model_base} --device_id=0 --height=1024 --width=1024 --num_images_per_prompt=1 --cache_method="agb_cache" --iterator=6 + ``` + + 参数说明: + - --path：模型权重路径 + - --device_id：推理设备ID + - --height: 生成图片的高 + - --width: 生成图片的宽 + - --num_images_per_prompt: 单个prompt生成的图片数量 + - --cache_method: cache策略选择,默认为空,代表不开启cache,支持配置"static_cache"和"agb_cache","abg_cache"性能更好,但会引入更大的内存占用,当前"agb_cache"只支持50步迭代场景。 + - --iterator: 测试循环次数,取从第3次开始后续的推理耗时的均值。 + + 参考性能结果: + | 设备 | 无cache | static_cache | agb_cache | + |------|--------|---------------| -------- | + |Atlas 800I A2(8*32G)| 7.87s | 6.25s | 4.36s | + + +## 精度验证 + + 由于生成的图片存在随机性，提供两种精度验证方法： + 1. CLIP-score（文图匹配度量）：评估图片和输入文本的相关性，分数的取值范围为[-1, 1]，越高越好。使用Parti数据集进行验证。 + 2. HPSv2（图片美学度量）：评估生成图片的人类偏好评分，分数的取值范围为[0, 1]，越高越好。使用HPSv2数据集进行验证 + + 注意，由于要生成的图片数量较多，进行完整的精度验证需要耗费很长的时间。 + + 1. 下载Parti数据集 + + ```bash + wget https://raw.githubusercontent.com/google-research/parti/main/PartiPrompts.tsv --no-check-certificate + ``` + + 2. 下载模型权重 + + ```bash + # Clip Score和HPSv2均需要使用的权重 + GIT_LFS_SKIP_SMUDGE=1 + git clone https://huggingface.co/laion/CLIP-ViT-H-14-laion2B-s32B-b79K + cd ./CLIP-ViT-H-14-laion2B-s32B-b79K + + # HPSv2权重 + wget https://huggingface.co/spaces/xswu/HPSv2/resolve/main/HPS_v2_compressed.pt --no-check-certificate + ``` + 也可手动下载[权重](https://huggingface.co/laion/CLIP-ViT-H-14-laion2B-s32B-b79K/blob/main/open_clip_pytorch_model.bin) + 将权重放到`CLIP-ViT-H-14-laion2B-s32B-b79K`目录下，手动下载[HPSv2权重](https://huggingface.co/spaces/xswu/HPSv2/resolve/main/HPS_v2_compressed.pt)放到当前路径 + + 3. 使用推理脚本读取Parti数据集，生成图片 + + ```bash + python3 ./test_precision.py --path=${model_base} --device_id=0 --height=1024 --width=1024 --num_images_per_prompt=4 --cache_method="agb_cache" --prompt_file=./PartiPrompts.tsv --prompt_file_type=parti + ``` + + 参数说明： + - --path：模型权重路径。 + - --height: 生成图片的高 + - --width: 生成图片的宽 + - --num_images_per_prompt: 每个prompt生成的图片数量。注意使用hpsv2时，设置num_images_per_prompt=1即可。 + - --device_id：推理设备ID。 + - --cache_method: cache策略选择,默认为空,代表不开启cache,支持配置"static_cache"和"agb_cache","abg_cache"性能更好,但会引入更大的内存占用,当前"agb_cache"只支持50步迭代场景。 + - --prompt_file：提示词文件。 + - --prompt_file_type: prompt文件类型，用于指定读取方式，可选plain，parti，hpsv2 + + 运行完成后的结果会保存到以运行时间戳命名的文件夹下,文件夹下包含结果'image_info.json'文件以及保存的图片文件夹。 + + 4. 计算精度指标 + + 1. CLIP-score + + ```bash + python3 clip_score.py \ + --image_info="image_info.json" \ + --model_name="ViT-H-14" \ + --model_weights_path="./CLIP-ViT-H-14-laion2B-s32B-b79K/open_clip_pytorch_model.bin" + ``` + + 参数说明： + - --image_info: 上一步生成的`image_info.json`文件。 + - --model_name: Clip模型名称。 + - --model_weights_path: Clip模型权重文件路径。 + + 执行完成后会在屏幕打印出精度计算结果。 + + 2. HPSv2 + + ```bash + python3 hpsv2_score.py \ + --image_info="image_info.json" \ + --HPSv2_checkpoint="./HPS_v2_compressed.pt" \ + --clip_checkpoint="./CLIP-ViT-H-14-laion2B-s32B-b79K/open_clip_pytorch_model.bin" + ``` + + 参数说明： + - --image_info: 上一步生成的`image_info.json`文件。 + - --HPSv2_checkpoint: HPSv2模型权重文件路径。 + - --clip_checkpointh: Clip模型权重文件路径。 + + 执行完成后会在屏幕打印出精度计算结果。 diff --git a/MindIE/MultiModal/StableDiffusion-XL/clip_score.py b/MindIE/MultiModal/StableDiffusion-XL/clip_score.py new file mode 100644 index 0000000000000000000000000000000000000000..5f72562eac8c72a3d7419b085dff92c83736882f --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-XL/clip_score.py @@ -0,0 +1,144 @@ +# Copyright 2023 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import os +import json +import time +import argparse + +import open_clip +import numpy as np +from PIL import Image +import torch +import torch.nn.functional as F + + +def clip_score(model_clip, tokenizer, preprocess, prompt, image_files, device): + imgs = [] + texts = [] + for image_file in image_files: + img = preprocess(Image.open(image_file)).unsqueeze(0).to(device) + imgs.append(img) + text = tokenizer([prompt]).to(device) + texts.append(text) + + img = torch.cat(imgs) # [bs, 3, 224, 224] + text = torch.cat(texts) # [bs, 77] + + with torch.no_grad(): + text_ft = model_clip.encode_text(text).float() + img_ft = model_clip.encode_image(img).float() + score = F.cosine_similarity(img_ft, text_ft).squeeze() + + return score.cpu() + + +def main(): + args = parse_arguments() + + if args.device is None: + try: + import torch_npu + device = torch.device("npu") + except: + device = torch.device("cpu") + else: + device = torch.device(args.device) + + t_b = time.time() + print(f"Load clip model...") + model_clip, _, preprocess = open_clip.create_model_and_transforms( + args.model_name, pretrained=args.model_weights_path, device=device) + model_clip.eval() + print(f">done. elapsed time: {(time.time() - t_b):.3f} s") + + tokenizer = open_clip.get_tokenizer(args.model_name) + + with os.fdopen(os.open(args.image_info, os.O_RDONLY), "r") as f: + image_info = json.load(f) + + t_b = time.time() + print(f"Calc clip score...") + all_scores = [] + cat_scores = {} + + for i, info in enumerate(image_info): + image_files = info['images'] + category = info['category'] + prompt = info['prompt'] + + print(f"[{i + 1}/{len(image_info)}] {prompt}") + + image_scores = clip_score(model_clip, + tokenizer, + preprocess, + prompt, + image_files, + device) + if len(image_files) > 1: + best_score = max(image_scores) + else: + best_score = image_scores + + print(f"image scores: {image_scores}") + print(f"best score: {best_score}") + + all_scores.append(best_score) + if category not in cat_scores: + cat_scores[category] = [] + cat_scores[category].append(best_score) + print(f">done. elapsed time: {(time.time() - t_b):.3f} s") + + average_score = np.average(all_scores) + print(f"====================================") + print(f"average score: {average_score:.3f}") + print(f"category average scores:") + cat_average_scores = {} + for category, scores in cat_scores.items(): + cat_average_scores[category] = np.average(scores) + print(f"[{category}], average score: {cat_average_scores[category]:.3f}") + + +def parse_arguments(): + parser = argparse.ArgumentParser() + parser.add_argument( + "--device", + type=str, + default="npu", + choices=["cpu", "cuda", "npu"], + help="device for torch.", + ) + parser.add_argument( + "--image_info", + type=str, + default="./image_info.json", + help="Image_info.json file.", + ) + parser.add_argument( + "--model_name", + type=str, + default="ViT-H-14", + help="open clip model name", + ) + parser.add_argument( + "--model_weights_path", + type=str, + default="./CLIP-ViT-H-14-laion2B-s32B-b79K/open_clip_pytorch_model.bin", + help="open clip model weights", + ) + return parser.parse_args() + + +if __name__ == '__main__': + main() \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-XL/hpsv2_score.py b/MindIE/MultiModal/StableDiffusion-XL/hpsv2_score.py new file mode 100644 index 0000000000000000000000000000000000000000..4cb3783a53f4878071291e0ab185cf20e7a23f84 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-XL/hpsv2_score.py @@ -0,0 +1,127 @@ +# Copyright 2024 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import argparse +import os +from typing import Union +import json + +from clint.textui import progress +import hpsv2 +from hpsv2.utils import root_path, hps_version_map +from hpsv2.src.open_clip import create_model_and_transforms, get_tokenizer +import huggingface_hub +from PIL import Image +import requests +import torch + + +def initialize_model(pretrained_path, device): + model, _, preprocess_val = create_model_and_transforms( + "ViT-H-14", pretrained=pretrained_path, precision='amp', + device=device, + jit=False, + force_quick_gelu=False, + force_custom_text=False, + force_patch_dropout=False, + force_image_size=None, + pretrained_image=False, + image_mean=None, + image_std=None, + light_augmentation=True, + aug_cfg={}, + output_dict=True, + with_score_predictor=False, + with_region_predictor=False + ) + return model, preprocess_val + + +def parse_arguments(): + parser = argparse.ArgumentParser() + parser.add_argument( + "--image_info", + type=str, + default="./image_info.json", + help="Image_info.json file.", + ) + parser.add_argument( + "--HPSv2_checkpoint", + type=str, + default="./HPS_v2_compressed.pt", + help="HPS_v2 model weights", + ) + parser.add_argument( + "--clip_checkpoint", + type=str, + default="./CLIP-ViT-H-14-laion2B-s32B-b79K/open_clip_pytorch_model.bin", + help="open clip model weights", + ) + return parser.parse_args() + + +def main(): + args = parse_arguments() + + try: + import torch_npu + device = torch.device("npu") + except: + device = torch.device("cpu") + + model, preprocess_val = initialize_model(args.clip_checkpoint, device) + + checkpoint = torch.load(args.HPSv2_checkpoint, map_location=device) + model.load_state_dict(checkpoint['state_dict']) + tokenizer = get_tokenizer('ViT-H-14') + model = model.to(device) + model.eval() + + with os.fdopen(os.open(args.image_info, os.O_RDONLY), "r") as f: + image_info = json.load(f) + + result = [] + for i, info in enumerate(image_info): + image_file = info['images'][0] + prompt = info['prompt'] + + # Load your image and prompt + with torch.no_grad(): + # Process the image + if isinstance(image_file, str): + image = preprocess_val(Image.open(image_file)) + elif isinstance(image_file, Image.Image): + image = preprocess_val(image_file) + else: + raise TypeError('The type of parameter img_path is illegal.') + image = image.unsqueeze(0).to(device=device, non_blocking=True) + # Process the prompt + text = tokenizer([prompt]).to(device=device, non_blocking=True) + # Calculate the HPS + with torch.cuda.amp.autocast(): + outputs = model(image, text) + image_features = outputs["image_features"] + text_features = outputs["text_features"] + logits_per_image = image_features @ text_features.T + + hps_score = torch.diagonal(logits_per_image).cpu().numpy() + print(f"image {i} hps_score: ", hps_score[0]) + + result.append(hps_score[0]) + + print('avg HPSv2 score:', sum(result) / len(result)) + + +if __name__ == '__main__': + main() \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-XL/infer_pipe.py b/MindIE/MultiModal/StableDiffusion-XL/infer_pipe.py new file mode 100644 index 0000000000000000000000000000000000000000..5e234133043ba3d48df7c003b881610917f4ad7c --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-XL/infer_pipe.py @@ -0,0 +1,84 @@ +import os +import argparse +import time +from typing import Union, List +import torch +import torch_npu +import numpy as np +import cv2 + +from stable_diffusion_xl.pipeline_stable_diffusion_xl import StableDiffusionXLPipeline +from stable_diffusion_xl.unet.unet_model import UNet2DConditionModel + + +def parse_arguments(): + parser = argparse.ArgumentParser() + parser.add_argument( + "--path", + type=str, + default='/stable-diffusion-xl-base-1.0', + help="The path of all model weights, suach as vae, unet, text_encoder, tokenizer, scheduler", + ) + parser.add_argument( + "--device_id", + type=int, + default=0, + help="NPU device id", + ) + parser.add_argument( + "--dtype", + type=torch.dtype, + default=torch.float16 + ) + parser.add_argument( + "--prompt", + type=str, + default="A dog, site on beach." + ) + parser.add_argument( + "--num_images_per_prompt", + type=int, + default=1 + ) + parser.add_argument( + "--height", + type=int, + default=1024 + ) + parser.add_argument( + "--width", + type=int, + default=1024 + ) + return parser.parse_args() + + +def init_env(device_id: int): + torch.npu.set_device(device_id) + + +def init_pipe(model_path: str, dtype=torch.float16): + unet = UNet2DConditionModel.from_pretrained(os.path.join(model_path, 'unet'), cache_method="agb_cahce") + pipe = StableDiffusionXLPipeline.from_pretrained(model_path, unet=unet) + pipe.to(dtype).to("npu") + return pipe + + +def infer_prompts(pipe, prompt, height=1024, width=1024, num_image=1): + images = pipe( + prompt=prompt, + height=height, + width=width, + num_images_per_prompt=num_image + ).images + for i, img in enumerate(images): + img_bgr = cv2.cvtColor(np.asarray(img), cv2.COLOR_RGB2BGR) + cv2.imwrite(f"{i}.png", img_bgr) + + +if __name__ == "__main__": + args = parse_arguments() + init_env(args.device_id) + pipe = init_pipe(args.path, args.dtype) + prompt = args.prompt + infer_prompts(pipe, prompt, args.height, args.width, args.num_images_per_prompt) diff --git a/MindIE/MultiModal/StableDiffusion-XL/prompts.txt b/MindIE/MultiModal/StableDiffusion-XL/prompts.txt new file mode 100644 index 0000000000000000000000000000000000000000..a375a0bb63931d0d5da6c6d91df1e14f870f47d0 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-XL/prompts.txt @@ -0,0 +1,16 @@ +Beautiful illustration of The ocean. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Islands in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Seaports in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of The waves. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Grassland. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Wheat. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Hut Tong. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of The boat. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Pine trees. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Bamboo. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of The temple. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Cloud in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Sun in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Spring. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Lotus. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Snow piles. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-XL/requirements.txt b/MindIE/MultiModal/StableDiffusion-XL/requirements.txt new file mode 100644 index 0000000000000000000000000000000000000000..84365fe1f6e72f38007fc976078bf7bb404822df --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-XL/requirements.txt @@ -0,0 +1,5 @@ +torch==2.1.0 +diffusers==0.30.3 +transformers==4.39.3 +opencv-python +opencv-python-headless \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-XL/stable_diffusion_xl/layers/__init__.py b/MindIE/MultiModal/StableDiffusion-XL/stable_diffusion_xl/layers/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/MindIE/MultiModal/StableDiffusion-XL/stable_diffusion_xl/layers/attention.py b/MindIE/MultiModal/StableDiffusion-XL/stable_diffusion_xl/layers/attention.py new file mode 100644 index 0000000000000000000000000000000000000000..d468563964e2ced42056e5b3cbc409130bc957b5 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-XL/stable_diffusion_xl/layers/attention.py @@ -0,0 +1,420 @@ +# Copyright 2024 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from typing import Optional +import inspect + +import torch +import torch.nn as nn +import torch.nn.functional as F +import torch_npu + + +def get_npu_device(): + # 默认获取当前设备信息 + device_name = torch_npu.npu.get_device_name() + if "3" in device_name: + return "DUO" + elif "9" in device_name: + return "A2" + else: + return "" + + +soc = get_npu_device() + + +class SpatialNorm(nn.Module): + """ + Spatially conditioned normalization as defined in https://arxiv.org/abs/2209.09002. + + Args: + f_channels (`int`): + The number of channels for input to group normalization layer, and output of the spatial norm layer. + zq_channels (`int`): + The number of channels for the quantized vector as described in the paper. + """ + + def __init__( + self, + f_channels: int, + zq_channels: int, + ): + super().__init__() + self.norm_layer = nn.GroupNorm(num_channels=f_channels, num_groups=32, eps=1e-6, affine=True) + self.conv_y = nn.Conv2d(zq_channels, f_channels, kernel_size=1, stride=1, padding=0) + self.conv_b = nn.Conv2d(zq_channels, f_channels, kernel_size=1, stride=1, padding=0) + + def forward(self, f: torch.Tensor, zq: torch.Tensor) -> torch.Tensor: + f_size = f.shape[-2:] + zq = F.interpolate(zq, size=f_size, mode="nearest") + norm_f = self.norm_layer(f) + new_f = norm_f * self.conv_y(zq) + self.conv_b(zq) + return new_f + + +class Attention(nn.Module): + r""" + A cross attention layer. + + Parameters: + query_dim (`int`): + The number of channels in the query. + cross_attention_dim (`int`, *optional*): + The number of channels in the encoder_hidden_states. If not given, defaults to `query_dim`. + heads (`int`, *optional*, defaults to 8): + The number of heads to use for multi-head attention. + dim_head (`int`, *optional*, defaults to 64): + The number of channels in each head. + dropout (`float`, *optional*, defaults to 0.0): + The dropout probability to use. + bias (`bool`, *optional*, defaults to False): + Set to `True` for the query, key, and value linear layers to contain a bias parameter. + upcast_attention (`bool`, *optional*, defaults to False): + Set to `True` to upcast the attention computation to `float32`. + upcast_softmax (`bool`, *optional*, defaults to False): + Set to `True` to upcast the softmax computation to `float32`. + cross_attention_norm (`str`, *optional*, defaults to `None`): + The type of normalization to use for the cross attention. Can be `None`, `layer_norm`, or `group_norm`. + cross_attention_norm_num_groups (`int`, *optional*, defaults to 32): + The number of groups to use for the group norm in the cross attention. + added_kv_proj_dim (`int`, *optional*, defaults to `None`): + The number of channels to use for the added key and value projections. If `None`, no projection is used. + norm_num_groups (`int`, *optional*, defaults to `None`): + The number of groups to use for the group norm in the attention. + spatial_norm_dim (`int`, *optional*, defaults to `None`): + The number of channels to use for the spatial normalization. + out_bias (`bool`, *optional*, defaults to `True`): + Set to `True` to use a bias in the output linear layer. + scale_qk (`bool`, *optional*, defaults to `True`): + Set to `True` to scale the query and key by `1 / sqrt(dim_head)`. + only_cross_attention (`bool`, *optional*, defaults to `False`): + Set to `True` to only use cross attention and not added_kv_proj_dim. Can only be set to `True` if + `added_kv_proj_dim` is not `None`. + eps (`float`, *optional*, defaults to 1e-5): + An additional value added to the denominator in group normalization that is used for numerical stability. + rescale_output_factor (`float`, *optional*, defaults to 1.0): + A factor to rescale the output by dividing it with this value. + residual_connection (`bool`, *optional*, defaults to `False`): + Set to `True` to add the residual connection to the output. + _from_deprecated_attn_block (`bool`, *optional*, defaults to `False`): + Set to `True` if the attention block is loaded from a deprecated state dict. + processor (`AttnProcessor`, *optional*, defaults to `None`): + The attention processor to use. If `None`, defaults to `AttnProcessor2_0` if `torch 2.x` is used and + `AttnProcessor` otherwise. + """ + + def __init__( + self, + query_dim: int, + cross_attention_dim: Optional[int] = None, + heads: int = 8, + dim_head: int = 64, + dropout: float = 0.0, + bias: bool = False, + upcast_attention: bool = False, + upcast_softmax: bool = False, + cross_attention_norm: Optional[str] = None, + cross_attention_norm_num_groups: int = 32, + qk_norm: Optional[str] = None, + added_kv_proj_dim: Optional[int] = None, + norm_num_groups: Optional[int] = None, + spatial_norm_dim: Optional[int] = None, + out_bias: bool = True, + scale_qk: bool = True, + only_cross_attention: bool = False, + eps: float = 1e-5, + rescale_output_factor: float = 1.0, + residual_connection: bool = False, + _from_deprecated_attn_block: bool = False, + processor: Optional["AttnProcessor"] = None, + out_dim: int = None, + context_pre_only=None, + ): + super().__init__() + self.inner_dim = out_dim if out_dim is not None else dim_head * heads + self.query_dim = query_dim + self.use_bias = bias + self.is_cross_attention = cross_attention_dim is not None + self.cross_attention_dim = cross_attention_dim if cross_attention_dim is not None else query_dim + self.upcast_attention = upcast_attention + self.upcast_softmax = upcast_softmax + self.rescale_output_factor = rescale_output_factor + self.residual_connection = residual_connection + self.dropout = dropout + self.fused_projections = False + self.out_dim = out_dim if out_dim is not None else query_dim + self.context_pre_only = context_pre_only + + # we make use of this private variable to know whether this class is loaded + # with an deprecated state dict so that we can convert it on the fly + self._from_deprecated_attn_block = _from_deprecated_attn_block + + self.scale_qk = scale_qk + self.scale = dim_head**-0.5 if self.scale_qk else 1.0 + + self.heads = out_dim // dim_head if out_dim is not None else heads + # for slice_size > 0 the attention score computation + # is split across the batch axis to save memory + # You can set slice_size with `set_attention_slice` + self.sliceable_head_dim = heads + + self.added_kv_proj_dim = added_kv_proj_dim + self.only_cross_attention = only_cross_attention + + if self.added_kv_proj_dim is None and self.only_cross_attention: + raise ValueError( + "`only_cross_attention` can only be set to True if `added_kv_proj_dim` is not None. Make sure to set either `only_cross_attention=False` or define `added_kv_proj_dim`." + ) + + if norm_num_groups is not None: + self.group_norm = nn.GroupNorm(num_channels=query_dim, num_groups=norm_num_groups, eps=eps, affine=True) + else: + self.group_norm = None + + if spatial_norm_dim is not None: + self.spatial_norm = SpatialNorm(f_channels=query_dim, zq_channels=spatial_norm_dim) + else: + self.spatial_norm = None + + if qk_norm is None: + self.norm_q = None + self.norm_k = None + elif qk_norm == "layer_norm": + self.norm_q = nn.LayerNorm(dim_head, eps=eps) + self.norm_k = nn.LayerNorm(dim_head, eps=eps) + else: + raise ValueError(f"unknown qk_norm: {qk_norm}. Should be None or 'layer_norm'") + + if cross_attention_norm is None: + self.norm_cross = None + elif cross_attention_norm == "layer_norm": + self.norm_cross = nn.LayerNorm(self.cross_attention_dim) + elif cross_attention_norm == "group_norm": + if self.added_kv_proj_dim is not None: + # The given `encoder_hidden_states` are initially of shape + # (batch_size, seq_len, added_kv_proj_dim) before being projected + # to (batch_size, seq_len, cross_attention_dim). The norm is applied + # before the projection, so we need to use `added_kv_proj_dim` as + # the number of channels for the group norm. + norm_cross_num_channels = added_kv_proj_dim + else: + norm_cross_num_channels = self.cross_attention_dim + + self.norm_cross = nn.GroupNorm( + num_channels=norm_cross_num_channels, num_groups=cross_attention_norm_num_groups, eps=1e-5, affine=True + ) + else: + raise ValueError( + f"unknown cross_attention_norm: {cross_attention_norm}. Should be None, 'layer_norm' or 'group_norm'" + ) + + self.to_q = nn.Linear(query_dim, self.inner_dim, bias=bias) + + if not self.only_cross_attention: + # only relevant for the `AddedKVProcessor` classes + self.to_k = nn.Linear(self.cross_attention_dim, self.inner_dim, bias=bias) + self.to_v = nn.Linear(self.cross_attention_dim, self.inner_dim, bias=bias) + else: + self.to_k = None + self.to_v = None + + if self.added_kv_proj_dim is not None: + self.add_k_proj = nn.Linear(added_kv_proj_dim, self.inner_dim) + self.add_v_proj = nn.Linear(added_kv_proj_dim, self.inner_dim) + if self.context_pre_only is not None: + self.add_q_proj = nn.Linear(added_kv_proj_dim, self.inner_dim) + + self.to_out = nn.ModuleList([]) + self.to_out.append(nn.Linear(self.inner_dim, self.out_dim, bias=out_bias)) + self.to_out.append(nn.Dropout(dropout)) + + if self.context_pre_only is not None and not self.context_pre_only: + self.to_add_out = nn.Linear(self.inner_dim, self.out_dim, bias=out_bias) + + # set attention processor + if processor is None: + processor = ( + AttnProcessor2_0() + ) + self.set_processor(processor) + + def set_processor(self, processor: "AttnProcessor") -> None: + r""" + Set the attention processor to use. + + Args: + processor (`AttnProcessor`): + The attention processor to use. + """ + + if ( + hasattr(self, "processor") + and isinstance(self.processor, torch.nn.Module) + and not isinstance(processor, torch.nn.Module) + ): + logger.info(f"You are removing possibly trained weights of {self.processor} with {processor}") + self._modules.pop("processor") + + self.processor = processor + + def forward( + self, + hidden_states: torch.Tensor, + encoder_hidden_states: Optional[torch.Tensor] = None, + attention_mask: Optional[torch.Tensor] = None, + **cross_attention_kwargs, + ) -> torch.Tensor: + r""" + The forward method of the `Attention` class. + + Args: + hidden_states (`torch.Tensor`): + The hidden states of the query. + encoder_hidden_states (`torch.Tensor`, *optional*): + The hidden states of the encoder. + attention_mask (`torch.Tensor`, *optional*): + The attention mask to use. If `None`, no mask is applied. + **cross_attention_kwargs: + Additional keyword arguments to pass along to the cross attention. + + Returns: + `torch.Tensor`: The output of the attention layer. + """ + # The `Attention` class can call different attention processors / attention functions + # here we simply pass along all tensors to the selected processor class + # For standard processors that are defined here, `**cross_attention_kwargs` is empty + + attn_parameters = set(inspect.signature(self.processor.__call__).parameters.keys()) + quiet_attn_parameters = {"ip_adapter_masks"} + unused_kwargs = [ + k for k, _ in cross_attention_kwargs.items() if k not in attn_parameters and k not in quiet_attn_parameters + ] + if len(unused_kwargs) > 0: + logger.warning( + f"cross_attention_kwargs {unused_kwargs} are not expected by {self.processor.__class__.__name__} and will be ignored." + ) + cross_attention_kwargs = {k: w for k, w in cross_attention_kwargs.items() if k in attn_parameters} + + return self.processor( + self, + hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=attention_mask, + **cross_attention_kwargs, + ) + + +class AttnProcessor2_0: + r""" + Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0). + """ + + def __init__(self): + if not hasattr(F, "scaled_dot_product_attention"): + raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.") + + def __call__( + self, + attn: Attention, + hidden_states: torch.Tensor, + encoder_hidden_states: Optional[torch.Tensor] = None, + attention_mask: Optional[torch.Tensor] = None, + temb: Optional[torch.Tensor] = None, + *args, + **kwargs, + ) -> torch.Tensor: + if len(args) > 0 or kwargs.get("scale", None) is not None: + deprecation_message = "The `scale` argument is deprecated and will be ignored. Please remove it, as passing it will raise an error in the future. `scale` should directly be passed while calling the underlying pipeline component i.e., via `cross_attention_kwargs`." + deprecate("scale", "1.0.0", deprecation_message) + + residual = hidden_states + if attn.spatial_norm is not None: + hidden_states = attn.spatial_norm(hidden_states, temb) + + input_ndim = hidden_states.ndim + + if input_ndim == 4: + batch_size, channel, height, width = hidden_states.shape + hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2) + + batch_size, sequence_length, _ = ( + hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape + ) + + if attention_mask is not None: + attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) + attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1]) + + if attn.group_norm is not None: + hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2) + + query = attn.to_q(hidden_states) + + if encoder_hidden_states is None: + encoder_hidden_states = hidden_states + elif attn.norm_cross: + encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states) + + key = attn.to_k(encoder_hidden_states) + value = attn.to_v(encoder_hidden_states) + + head_dim = attn.inner_dim // attn.heads + + if attention_mask: + attention_mask = ~attention_mask.to(torch.bool) + q_seqlen = query.shape[1] + kv_seqlen = key.shape[1] + + if q_seqlen == kv_seqlen: + # self attention + query = query.reshape(batch_size, -1, attn.heads, head_dim).transpose(1, 2) + key = key.reshape(batch_size, -1, attn.heads, head_dim).transpose(1, 2) + value = value.reshape(batch_size, -1, attn.heads, head_dim).transpose(1, 2) + + hidden_states = torch_npu.npu_prompt_flash_attention( + query, key, value, atten_mask=attention_mask, + input_layout='BNSD', scale_value=attn.scale, + pre_tokens=65535, next_tokens=65535, num_heads=attn.heads) + hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.inner_dim) + else: + # cross attention + query = query.reshape(batch_size, -1, attn.heads, head_dim) + key = key.reshape(batch_size, -1, attn.heads, head_dim) + value = value.reshape(batch_size, -1, attn.heads, head_dim) + if soc == "A2": + hidden_states = torch_npu.npu_fusion_attention( + query, key, value, atten_mask=attention_mask, + input_layout='BSND', scale=attn.scale, + pre_tockens=65535, next_tockens=65535, head_num=attn.heads)[0] + else: + hidden_states = torch_npu.npu_prompt_flash_attention( + query, key, value, atten_mask=attention_mask, + input_layout='BSND', scale_value=attn.scale, + pre_tokens=65535, next_tokens=65535, num_heads=attn.heads) + hidden_states = hidden_states.reshape(batch_size, -1, attn.inner_dim) + + hidden_states = hidden_states.to(query.dtype) + + # linear proj + hidden_states = attn.to_out[0](hidden_states) + + if input_ndim == 4: + hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width) + + if attn.residual_connection: + hidden_states = hidden_states + residual + + hidden_states = hidden_states / attn.rescale_output_factor + + return hidden_states \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-XL/stable_diffusion_xl/pipeline_stable_diffusion_xl.py b/MindIE/MultiModal/StableDiffusion-XL/stable_diffusion_xl/pipeline_stable_diffusion_xl.py new file mode 100644 index 0000000000000000000000000000000000000000..b186901a59abce00ba28ce5898be4c6ad777d056 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-XL/stable_diffusion_xl/pipeline_stable_diffusion_xl.py @@ -0,0 +1,1299 @@ +# Copyright 2024 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable, Dict, List, Optional, Tuple, Union + +import torch +from transformers import ( + CLIPImageProcessor, + CLIPTextModel, + CLIPTextModelWithProjection, + CLIPTokenizer, + CLIPVisionModelWithProjection, +) + +from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback +from diffusers.image_processor import PipelineImageInput, VaeImageProcessor +from diffusers.loaders import ( + FromSingleFileMixin, + IPAdapterMixin, + StableDiffusionXLLoraLoaderMixin, + TextualInversionLoaderMixin, +) +from diffusers.models import AutoencoderKL, ImageProjection, UNet2DConditionModel +from diffusers.models.attention_processor import ( + AttnProcessor2_0, + FusedAttnProcessor2_0, + XFormersAttnProcessor, +) +from diffusers.models.lora import adjust_lora_scale_text_encoder +from diffusers.schedulers import KarrasDiffusionSchedulers +from diffusers.utils import ( + USE_PEFT_BACKEND, + deprecate, + is_invisible_watermark_available, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from diffusers.utils.torch_utils import randn_tensor +from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from diffusers.pipelines.stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput + + +if is_invisible_watermark_available(): + from .watermark import StableDiffusionXLWatermarker + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import StableDiffusionXLPipeline + + >>> pipe = StableDiffusionXLPipeline.from_pretrained( + ... "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16 + ... ) + >>> pipe = pipe.to("cuda") + + >>> prompt = "a photo of an astronaut riding a horse on mars" + >>> image = pipe(prompt).images[0] + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + """ + Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4 + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: Optional[int] = None, + device: Optional[Union[str, torch.device]] = None, + timesteps: Optional[List[int]] = None, + sigmas: Optional[List[float]] = None, + **kwargs, +): + """ + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`List[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`List[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class StableDiffusionXLPipeline( + DiffusionPipeline, + StableDiffusionMixin, + FromSingleFileMixin, + StableDiffusionXLLoraLoaderMixin, + TextualInversionLoaderMixin, + IPAdapterMixin, +): + r""" + Pipeline for text-to-image generation using Stable Diffusion XL. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files + - [`~loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder. Stable Diffusion XL uses the text portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + text_encoder_2 ([` CLIPTextModelWithProjection`]): + Second frozen text-encoder. Stable Diffusion XL uses the text and pool portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection), + specifically the + [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k) + variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`CLIPTokenizer`): + Second Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + force_zeros_for_empty_prompt (`bool`, *optional*, defaults to `"True"`): + Whether the negative prompt embeddings shall be forced to always be set to 0. Also see the config of + `stabilityai/stable-diffusion-xl-base-1-0`. + add_watermarker (`bool`, *optional*): + Whether to use the [invisible_watermark library](https://github.com/ShieldMnt/invisible-watermark/) to + watermark output images. If not defined, it will default to True if the package is installed, otherwise no + watermarker will be used. + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->unet->vae" + _optional_components = [ + "tokenizer", + "tokenizer_2", + "text_encoder", + "text_encoder_2", + "image_encoder", + "feature_extractor", + ] + _callback_tensor_inputs = [ + "latents", + "prompt_embeds", + "negative_prompt_embeds", + "add_text_embeds", + "add_time_ids", + "negative_pooled_prompt_embeds", + "negative_add_time_ids", + ] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + text_encoder_2: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + tokenizer_2: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + image_encoder: CLIPVisionModelWithProjection = None, + feature_extractor: CLIPImageProcessor = None, + force_zeros_for_empty_prompt: bool = True, + add_watermarker: Optional[bool] = None, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + unet=unet, + scheduler=scheduler, + image_encoder=image_encoder, + feature_extractor=feature_extractor, + ) + self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + + self.default_sample_size = self.unet.config.sample_size + + add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available() + + if add_watermarker: + self.watermark = StableDiffusionXLWatermarker() + else: + self.watermark = None + + def encode_prompt( + self, + prompt: str, + prompt_2: Optional[str] = None, + device: Optional[torch.device] = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt: Optional[str] = None, + negative_prompt_2: Optional[str] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + pooled_prompt_embeds: Optional[torch.Tensor] = None, + negative_pooled_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + clip_skip: Optional[int] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale) + else: + scale_lora_layers(self.text_encoder_2, lora_scale) + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # Define tokenizers and text encoders + tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2] + text_encoders = ( + [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2] + ) + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + # textual inversion: process multi-vector tokens if necessary + prompt_embeds_list = [] + prompts = [prompt, prompt_2] + for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, tokenizer) + + text_inputs = tokenizer( + prompt, + padding="max_length", + max_length=tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {tokenizer.model_max_length} tokens: {removed_text}" + ) + + prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True) + + # We are only ALWAYS interested in the pooled output of the final text encoder + pooled_prompt_embeds = prompt_embeds[0] + if clip_skip is None: + prompt_embeds = prompt_embeds.hidden_states[-2] + else: + # "2" because SDXL always indexes from the penultimate layer. + prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)] + + prompt_embeds_list.append(prompt_embeds) + + prompt_embeds = torch.concat(prompt_embeds_list, dim=-1) + + # get unconditional embeddings for classifier free guidance + zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt + if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt: + negative_prompt_embeds = torch.zeros_like(prompt_embeds) + negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds) + elif do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt_2 = negative_prompt_2 or negative_prompt + + # normalize str to list + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + negative_prompt_2 = ( + batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2 + ) + + uncond_tokens: List[str] + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = [negative_prompt, negative_prompt_2] + + negative_prompt_embeds_list = [] + for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = tokenizer( + negative_prompt, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + negative_prompt_embeds = text_encoder( + uncond_input.input_ids.to(device), + output_hidden_states=True, + ) + # We are only ALWAYS interested in the pooled output of the final text encoder + negative_pooled_prompt_embeds = negative_prompt_embeds[0] + negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2] + + negative_prompt_embeds_list.append(negative_prompt_embeds) + + negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1) + + if self.text_encoder_2 is not None: + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + else: + prompt_embeds = prompt_embeds.to(dtype=self.unet.dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + if self.text_encoder_2 is not None: + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + else: + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.unet.dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + if do_classifier_free_guidance: + negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance + ): + image_embeds = [] + if do_classifier_free_guidance: + negative_image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers + ): + output_hidden_state = not isinstance(image_proj_layer, ImageProjection) + single_image_embeds, single_negative_image_embeds = self.encode_image( + single_ip_adapter_image, device, 1, output_hidden_state + ) + + image_embeds.append(single_image_embeds[None, :]) + if do_classifier_free_guidance: + negative_image_embeds.append(single_negative_image_embeds[None, :]) + else: + for single_image_embeds in ip_adapter_image_embeds: + if do_classifier_free_guidance: + single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2) + negative_image_embeds.append(single_negative_image_embeds) + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for i, single_image_embeds in enumerate(image_embeds): + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + if do_classifier_free_guidance: + single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0) + single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0) + + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + prompt_2, + height, + width, + callback_steps, + negative_prompt=None, + negative_prompt_2=None, + prompt_embeds=None, + negative_prompt_embeds=None, + pooled_prompt_embeds=None, + negative_pooled_prompt_embeds=None, + ip_adapter_image=None, + ip_adapter_image_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + + if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + if ip_adapter_image is not None and ip_adapter_image_embeds is not None: + raise ValueError( + "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined." + ) + + if ip_adapter_image_embeds is not None: + if not isinstance(ip_adapter_image_embeds, list): + raise ValueError( + f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}" + ) + elif ip_adapter_image_embeds[0].ndim not in [3, 4]: + raise ValueError( + f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D" + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + def _get_add_time_ids( + self, original_size, crops_coords_top_left, target_size, dtype, text_encoder_projection_dim=None + ): + add_time_ids = list(original_size + crops_coords_top_left + target_size) + + passed_add_embed_dim = ( + self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim + ) + expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features + + if expected_add_embed_dim != passed_add_embed_dim: + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`." + ) + + add_time_ids = torch.tensor([add_time_ids], dtype=dtype) + return add_time_ids + + def upcast_vae(self): + dtype = self.vae.dtype + self.vae.to(dtype=torch.float32) + use_torch_2_0_or_xformers = isinstance( + self.vae.decoder.mid_block.attentions[0].processor, + ( + AttnProcessor2_0, + XFormersAttnProcessor, + FusedAttnProcessor2_0, + ), + ) + # if xformers or torch_2_0 is used attention block does not need + # to be in float32 which can save lots of memory + if use_torch_2_0_or_xformers: + self.vae.post_quant_conv.to(dtype) + self.vae.decoder.conv_in.to(dtype) + self.vae.decoder.mid_block.to(dtype) + + # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding + def get_guidance_scale_embedding( + self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32 + ) -> torch.Tensor: + """ + See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 + + Args: + w (`torch.Tensor`): + Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings. + embedding_dim (`int`, *optional*, defaults to 512): + Dimension of the embeddings to generate. + dtype (`torch.dtype`, *optional*, defaults to `torch.float32`): + Data type of the generated embeddings. + + Returns: + `torch.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`. + """ + w = w * 1000.0 + + half_dim = embedding_dim // 2 + emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) + emb = w.to(dtype)[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0, 1)) + return emb + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def guidance_rescale(self): + return self._guidance_rescale + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None + + @property + def cross_attention_kwargs(self): + return self._cross_attention_kwargs + + @property + def denoising_end(self): + return self._denoising_end + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Union[str, List[str]] = None, + prompt_2: Optional[Union[str, List[str]]] = None, + height: Optional[int] = None, + width: Optional[int] = None, + num_inference_steps: int = 50, + timesteps: List[int] = None, + sigmas: List[float] = None, + denoising_end: Optional[float] = None, + guidance_scale: float = 5.0, + negative_prompt: Optional[Union[str, List[str]]] = None, + negative_prompt_2: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + pooled_prompt_embeds: Optional[torch.Tensor] = None, + negative_pooled_prompt_embeds: Optional[torch.Tensor] = None, + ip_adapter_image: Optional[PipelineImageInput] = None, + ip_adapter_image_embeds: Optional[List[torch.Tensor]] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + guidance_rescale: float = 0.0, + original_size: Optional[Tuple[int, int]] = None, + crops_coords_top_left: Tuple[int, int] = (0, 0), + target_size: Optional[Tuple[int, int]] = None, + negative_original_size: Optional[Tuple[int, int]] = None, + negative_crops_coords_top_left: Tuple[int, int] = (0, 0), + negative_target_size: Optional[Tuple[int, int]] = None, + clip_skip: Optional[int] = None, + callback_on_step_end: Optional[ + Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks] + ] = None, + callback_on_step_end_tensor_inputs: List[str] = ["latents"], + **kwargs, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + Anything below 512 pixels won't work well for + [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0) + and checkpoints that are not specifically fine-tuned on low resolutions. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + Anything below 512 pixels won't work well for + [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0) + and checkpoints that are not specifically fine-tuned on low resolutions. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`List[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + sigmas (`List[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + denoising_end (`float`, *optional*): + When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be + completed before it is intentionally prematurely terminated. As a result, the returned sample will + still retain a substantial amount of noise as determined by the discrete timesteps selected by the + scheduler. The denoising_end parameter should ideally be utilized when this pipeline forms a part of a + "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image + Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output) + guidance_scale (`float`, *optional*, defaults to 5.0): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to + [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`List[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should + contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] instead + of a plain tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + guidance_rescale (`float`, *optional*, defaults to 0.0): + Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://arxiv.org/pdf/2305.08891.pdf) `guidance_scale` is defined as `φ` in equation 16. of + [Common Diffusion Noise Schedules and Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). + Guidance rescale factor should fix overexposure when using zero terminal SNR. + original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled. + `original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as + explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)): + `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position + `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting + `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + For most cases, `target_size` should be set to the desired height and width of the generated image. If + not specified it will default to `(height, width)`. Part of SDXL's micro-conditioning as explained in + section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + negative_original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + To negatively condition the generation process based on a specific image resolution. Part of SDXL's + micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + negative_crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)): + To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's + micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + negative_target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + To negatively condition the generation process based on a target image resolution. It should be as same + as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + + Examples: + + Returns: + [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a + `tuple`. When returning a tuple, the first element is a list with the generated images. + """ + + callback = kwargs.pop("callback", None) + callback_steps = kwargs.pop("callback_steps", None) + + if callback is not None: + deprecate( + "callback", + "1.0.0", + "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + if callback_steps is not None: + deprecate( + "callback_steps", + "1.0.0", + "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 0. Default height and width to unet + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + original_size = original_size or (height, width) + target_size = target_size or (height, width) + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + height, + width, + callback_steps, + negative_prompt, + negative_prompt_2, + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ip_adapter_image, + ip_adapter_image_embeds, + callback_on_step_end_tensor_inputs, + ) + + self._guidance_scale = guidance_scale + self._guidance_rescale = guidance_rescale + self._clip_skip = clip_skip + self._cross_attention_kwargs = cross_attention_kwargs + self._denoising_end = denoising_end + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Encode input prompt + lora_scale = ( + self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None + ) + + ( + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + lora_scale=lora_scale, + clip_skip=self.clip_skip, + ) + + # 4. Prepare timesteps + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, device, timesteps, sigmas + ) + + # 5. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7. Prepare added time ids & embeddings + add_text_embeds = pooled_prompt_embeds + if self.text_encoder_2 is None: + text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1]) + else: + text_encoder_projection_dim = self.text_encoder_2.config.projection_dim + + add_time_ids = self._get_add_time_ids( + original_size, + crops_coords_top_left, + target_size, + dtype=prompt_embeds.dtype, + text_encoder_projection_dim=text_encoder_projection_dim, + ) + if negative_original_size is not None and negative_target_size is not None: + negative_add_time_ids = self._get_add_time_ids( + negative_original_size, + negative_crops_coords_top_left, + negative_target_size, + dtype=prompt_embeds.dtype, + text_encoder_projection_dim=text_encoder_projection_dim, + ) + else: + negative_add_time_ids = add_time_ids + + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0) + add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0) + + prompt_embeds = prompt_embeds.to(device) + add_text_embeds = add_text_embeds.to(device) + add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1) + + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + ) + + # 8. Denoising loop + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + + # 8.1 Apply denoising_end + if ( + self.denoising_end is not None + and isinstance(self.denoising_end, float) + and self.denoising_end > 0 + and self.denoising_end < 1 + ): + discrete_timestep_cutoff = int( + round( + self.scheduler.config.num_train_timesteps + - (self.denoising_end * self.scheduler.config.num_train_timesteps) + ) + ) + num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps))) + timesteps = timesteps[:num_inference_steps] + + # 9. Optionally get Guidance Scale Embedding + timestep_cond = None + if self.unet.config.time_cond_proj_dim is not None: + guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt) + timestep_cond = self.get_guidance_scale_embedding( + guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim + ).to(device=device, dtype=latents.dtype) + + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids} + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + added_cond_kwargs["image_embeds"] = image_embeds + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + timestep_cond=timestep_cond, + cross_attention_kwargs=self.cross_attention_kwargs, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + step=i, + )[0] + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + + if self.do_classifier_free_guidance and self.guidance_rescale > 0.0: + # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf + noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + add_text_embeds = callback_outputs.pop("add_text_embeds", add_text_embeds) + negative_pooled_prompt_embeds = callback_outputs.pop( + "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds + ) + add_time_ids = callback_outputs.pop("add_time_ids", add_time_ids) + negative_add_time_ids = callback_outputs.pop("negative_add_time_ids", negative_add_time_ids) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + # make sure the VAE is in float32 mode, as it overflows in float16 + needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast + + if needs_upcasting: + self.upcast_vae() + latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype) + elif latents.dtype != self.vae.dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + self.vae = self.vae.to(latents.dtype) + + # unscale/denormalize the latents + # denormalize with the mean and std if available and not None + has_latents_mean = hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None + has_latents_std = hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None + if has_latents_mean and has_latents_std: + latents_mean = ( + torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1).to(latents.device, latents.dtype) + ) + latents_std = ( + torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1).to(latents.device, latents.dtype) + ) + latents = latents * latents_std / self.vae.config.scaling_factor + latents_mean + else: + latents = latents / self.vae.config.scaling_factor + + image = self.vae.decode(latents, return_dict=False)[0] + + # cast back to fp16 if needed + if needs_upcasting: + self.vae.to(dtype=torch.float16) + else: + image = latents + + if not output_type == "latent": + # apply watermark if available + if self.watermark is not None: + image = self.watermark.apply_watermark(image) + + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return StableDiffusionXLPipelineOutput(images=image) diff --git a/MindIE/MultiModal/StableDiffusion-XL/stable_diffusion_xl/unet/__init__.py b/MindIE/MultiModal/StableDiffusion-XL/stable_diffusion_xl/unet/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/MindIE/MultiModal/StableDiffusion-XL/stable_diffusion_xl/unet/transformer_2d_model.py b/MindIE/MultiModal/StableDiffusion-XL/stable_diffusion_xl/unet/transformer_2d_model.py new file mode 100644 index 0000000000000000000000000000000000000000..fdb5b1d55dea5522661fd983a3340432bc550d50 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-XL/stable_diffusion_xl/unet/transformer_2d_model.py @@ -0,0 +1,530 @@ +# Copyright 2024 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +import functools +import inspect +from typing import Optional, Dict, Any + +import torch +import torch.nn as nn + +from diffusers.models.modeling_utils import LegacyModelMixin +from diffusers.configuration_utils import LegacyConfigMixin, register_to_config +from .transformer_blocks import BasicTransformerBlock + + +class Transformer2DModel(LegacyModelMixin, LegacyConfigMixin): + """ + A 2D Transformer model for image-like data. + + Parameters: + num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention. + attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head. + in_channels (`int`, *optional*): + The number of channels in the input and output (specify if the input is **continuous**). + num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use. + dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. + cross_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use. + sample_size (`int`, *optional*): The width of the latent images (specify if the input is **discrete**). + This is fixed during training since it is used to learn a number of position embeddings. + num_vector_embeds (`int`, *optional*): + The number of classes of the vector embeddings of the latent pixels (specify if the input is **discrete**). + Includes the class for the masked latent pixel. + activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to use in feed-forward. + num_embeds_ada_norm ( `int`, *optional*): + The number of diffusion steps used during training. Pass if at least one of the norm_layers is + `AdaLayerNorm`. This is fixed during training since it is used to learn a number of embeddings that are + added to the hidden states. + + During inference, you can denoise for up to but not more steps than `num_embeds_ada_norm`. + attention_bias (`bool`, *optional*): + Configure if the `TransformerBlocks` attention should contain a bias parameter. + """ + + _supports_gradient_checkpointing = True + _no_split_modules = ["BasicTransformerBlock"] + + @register_to_config + def __init__( + self, + num_attention_heads: int = 16, + attention_head_dim: int = 88, + in_channels: Optional[int] = None, + out_channels: Optional[int] = None, + num_layers: int = 1, + dropout: float = 0.0, + norm_num_groups: int = 32, + cross_attention_dim: Optional[int] = None, + attention_bias: bool = False, + sample_size: Optional[int] = None, + num_vector_embeds: Optional[int] = None, + patch_size: Optional[int] = None, + activation_fn: str = "geglu", + num_embeds_ada_norm: Optional[int] = None, + use_linear_projection: bool = False, + only_cross_attention: bool = False, + double_self_attention: bool = False, + upcast_attention: bool = False, + norm_type: str = "layer_norm", # 'layer_norm', 'ada_norm', 'ada_norm_zero', 'ada_norm_single', 'ada_norm_continuous', 'layer_norm_i2vgen' + norm_elementwise_affine: bool = True, + norm_eps: float = 1e-5, + attention_type: str = "default", + caption_channels: int = None, + interpolation_scale: float = None, + use_additional_conditions: Optional[bool] = None, + ): + super().__init__() + # Validate inputs. + if patch_size is not None: + if norm_type not in ["ada_norm", "ada_norm_zero", "ada_norm_single"]: + raise NotImplementedError( + f"Forward pass is not implemented when `patch_size` is not None and `norm_type` is '{norm_type}'." + ) + elif norm_type in ["ada_norm", "ada_norm_zero"] and num_embeds_ada_norm is None: + raise ValueError( + f"When using a `patch_size` and this `norm_type` ({norm_type}), `num_embeds_ada_norm` cannot be None." + ) + + # 1. Transformer2DModel can process both standard continuous images of shape `(batch_size, num_channels, width, height)` as well as quantized image embeddings of shape `(batch_size, num_image_vectors)` + # Define whether input is continuous or discrete depending on configuration + self.is_input_continuous = (in_channels is not None) and (patch_size is None) + self.is_input_vectorized = num_vector_embeds is not None + self.is_input_patches = in_channels is not None and patch_size is not None + + if self.is_input_continuous and self.is_input_vectorized: + raise ValueError( + f"Cannot define both `in_channels`: {in_channels} and `num_vector_embeds`: {num_vector_embeds}. Make" + " sure that either `in_channels` or `num_vector_embeds` is None." + ) + elif self.is_input_vectorized and self.is_input_patches: + raise ValueError( + f"Cannot define both `num_vector_embeds`: {num_vector_embeds} and `patch_size`: {patch_size}. Make" + " sure that either `num_vector_embeds` or `num_patches` is None." + ) + elif not self.is_input_continuous and not self.is_input_vectorized and not self.is_input_patches: + raise ValueError( + f"Has to define `in_channels`: {in_channels}, `num_vector_embeds`: {num_vector_embeds}, or patch_size:" + f" {patch_size}. Make sure that `in_channels`, `num_vector_embeds` or `num_patches` is not None." + ) + + if norm_type == "layer_norm" and num_embeds_ada_norm is not None: + deprecation_message = ( + f"The configuration file of this model: {self.__class__} is outdated. `norm_type` is either not set or" + " incorrectly set to `'layer_norm'`. Make sure to set `norm_type` to `'ada_norm'` in the config." + " Please make sure to update the config accordingly as leaving `norm_type` might led to incorrect" + " results in future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it" + " would be very nice if you could open a Pull request for the `transformer/config.json` file" + ) + deprecate("norm_type!=num_embeds_ada_norm", "1.0.0", deprecation_message, standard_warn=False) + norm_type = "ada_norm" + + # Set some common variables used across the board. + self.use_linear_projection = use_linear_projection + self.interpolation_scale = interpolation_scale + self.caption_channels = caption_channels + self.num_attention_heads = num_attention_heads + self.attention_head_dim = attention_head_dim + self.inner_dim = self.num_attention_heads * self.attention_head_dim + self.in_channels = in_channels + self.out_channels = in_channels if out_channels is None else out_channels + self.gradient_checkpointing = False + + if use_additional_conditions is None: + if norm_type == "ada_norm_single" and sample_size == 128: + use_additional_conditions = True + else: + use_additional_conditions = False + self.use_additional_conditions = use_additional_conditions + + # 2. Initialize the right blocks. + # These functions follow a common structure: + # a. Initialize the input blocks. b. Initialize the transformer blocks. + # c. Initialize the output blocks and other projection blocks when necessary. + if self.is_input_continuous: + self._init_continuous_input(norm_type=norm_type) + elif self.is_input_vectorized: + self._init_vectorized_inputs(norm_type=norm_type) + elif self.is_input_patches: + self._init_patched_inputs(norm_type=norm_type) + + def _init_continuous_input(self, norm_type): + self.norm = torch.nn.GroupNorm( + num_groups=self.config.norm_num_groups, num_channels=self.in_channels, eps=1e-6, affine=True + ) + if self.use_linear_projection: + self.proj_in = torch.nn.Linear(self.in_channels, self.inner_dim) + else: + self.proj_in = torch.nn.Conv2d(self.in_channels, self.inner_dim, kernel_size=1, stride=1, padding=0) + + self.transformer_blocks = nn.ModuleList( + [ + BasicTransformerBlock( + self.inner_dim, + self.config.num_attention_heads, + self.config.attention_head_dim, + dropout=self.config.dropout, + cross_attention_dim=self.config.cross_attention_dim, + activation_fn=self.config.activation_fn, + num_embeds_ada_norm=self.config.num_embeds_ada_norm, + attention_bias=self.config.attention_bias, + only_cross_attention=self.config.only_cross_attention, + double_self_attention=self.config.double_self_attention, + upcast_attention=self.config.upcast_attention, + norm_type=norm_type, + norm_elementwise_affine=self.config.norm_elementwise_affine, + norm_eps=self.config.norm_eps, + attention_type=self.config.attention_type, + ) + for _ in range(self.config.num_layers) + ] + ) + + if self.use_linear_projection: + self.proj_out = torch.nn.Linear(self.inner_dim, self.out_channels) + else: + self.proj_out = torch.nn.Conv2d(self.inner_dim, self.out_channels, kernel_size=1, stride=1, padding=0) + + def _init_vectorized_inputs(self, norm_type): + self.height = self.sample_size + self.width = self.sample_size + self.num_latent_pixels = self.height * self.width + + self.latent_image_embedding = ImagePositionalEmbeddings( + num_embed=self.num_vector_embeds, embed_dim=self.inner_dim, height=self.height, width=self.width + ) + + self.transformer_blocks = nn.ModuleList( + [ + BasicTransformerBlock( + self.inner_dim, + self.config.num_attention_heads, + self.config.attention_head_dim, + dropout=self.config.dropout, + cross_attention_dim=self.config.cross_attention_dim, + activation_fn=self.config.activation_fn, + num_embeds_ada_norm=self.config.num_embeds_ada_norm, + attention_bias=self.config.attention_bias, + only_cross_attention=self.config.only_cross_attention, + double_self_attention=self.config.double_self_attention, + upcast_attention=self.config.upcast_attention, + norm_type=norm_type, + norm_elementwise_affine=self.config.norm_elementwise_affine, + norm_eps=self.config.norm_eps, + attention_type=self.config.attention_type, + ) + for _ in range(self.config.num_layers) + ] + ) + + self.norm_out = nn.LayerNorm(self.inner_dim) + self.out = nn.Linear(self.inner_dim, self.num_vector_embeds - 1) + + def _init_patched_inputs(self, norm_type): + self.height = self.sample_size + self.width = self.sample_size + + self.patch_size = self.patch_size + interpolation_scale = ( + self.interpolation_scale + if self.interpolation_scale is not None + else max(self.sample_size // 64, 1) + ) + self.pos_embed = PatchEmbed( + height=self.sample_size, + width=self.sample_size, + patch_size=self.patch_size, + in_channels=self.in_channels, + embed_dim=self.inner_dim, + interpolation_scale=interpolation_scale, + ) + + self.transformer_blocks = nn.ModuleList( + [ + BasicTransformerBlock( + self.inner_dim, + self.config.num_attention_heads, + self.config.attention_head_dim, + dropout=self.config.dropout, + cross_attention_dim=self.config.cross_attention_dim, + activation_fn=self.config.activation_fn, + num_embeds_ada_norm=self.config.num_embeds_ada_norm, + attention_bias=self.config.attention_bias, + only_cross_attention=self.config.only_cross_attention, + double_self_attention=self.config.double_self_attention, + upcast_attention=self.config.upcast_attention, + norm_type=norm_type, + norm_elementwise_affine=self.config.norm_elementwise_affine, + norm_eps=self.config.norm_eps, + attention_type=self.config.attention_type, + ) + for _ in range(self.config.num_layers) + ] + ) + + if self.norm_type != "ada_norm_single": + self.norm_out = nn.LayerNorm(self.inner_dim, elementwise_affine=False, eps=1e-6) + self.proj_out_1 = nn.Linear(self.inner_dim, 2 * self.inner_dim) + self.proj_out_2 = nn.Linear( + self.inner_dim, self.patch_size * self.patch_size * self.out_channels + ) + elif self.norm_type == "ada_norm_single": + self.norm_out = nn.LayerNorm(self.inner_dim, elementwise_affine=False, eps=1e-6) + self.scale_shift_table = nn.Parameter(torch.randn(2, self.inner_dim) / self.inner_dim**0.5) + self.proj_out = nn.Linear( + self.inner_dim, self.patch_size * self.patch_size * self.out_channels + ) + + # PixArt-Alpha blocks. + self.adaln_single = None + if self.norm_type == "ada_norm_single": + # TODO(Sayak, PVP) clean this, for now we use sample size to determine whether to use + # additional conditions until we find better name + self.adaln_single = AdaLayerNormSingle( + self.inner_dim, use_additional_conditions=self.use_additional_conditions + ) + + self.caption_projection = None + if self.caption_channels is not None: + self.caption_projection = PixArtAlphaTextProjection( + in_features=self.caption_channels, hidden_size=self.inner_dim + ) + + def _set_gradient_checkpointing(self, module, value=False): + if hasattr(module, "gradient_checkpointing"): + module.gradient_checkpointing = value + + def forward( + self, + hidden_states: torch.Tensor, + encoder_hidden_states: Optional[torch.Tensor] = None, + timestep: Optional[torch.LongTensor] = None, + added_cond_kwargs: Dict[str, torch.Tensor] = None, + class_labels: Optional[torch.LongTensor] = None, + cross_attention_kwargs: Dict[str, Any] = None, + attention_mask: Optional[torch.Tensor] = None, + encoder_attention_mask: Optional[torch.Tensor] = None, + return_dict: bool = True, + ): + """ + The [`Transformer2DModel`] forward method. + + Args: + hidden_states (`torch.LongTensor` of shape `(batch size, num latent pixels)` if discrete, `torch.Tensor` of shape `(batch size, channel, height, width)` if continuous): + Input `hidden_states`. + encoder_hidden_states ( `torch.Tensor` of shape `(batch size, sequence len, embed dims)`, *optional*): + Conditional embeddings for cross attention layer. If not given, cross-attention defaults to + self-attention. + timestep ( `torch.LongTensor`, *optional*): + Used to indicate denoising step. Optional timestep to be applied as an embedding in `AdaLayerNorm`. + class_labels ( `torch.LongTensor` of shape `(batch size, num classes)`, *optional*): + Used to indicate class labels conditioning. Optional class labels to be applied as an embedding in + `AdaLayerZeroNorm`. + cross_attention_kwargs ( `Dict[str, Any]`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + attention_mask ( `torch.Tensor`, *optional*): + An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask + is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large + negative values to the attention scores corresponding to "discard" tokens. + encoder_attention_mask ( `torch.Tensor`, *optional*): + Cross-attention mask applied to `encoder_hidden_states`. Two formats supported: + + * Mask `(batch, sequence_length)` True = keep, False = discard. + * Bias `(batch, 1, sequence_length)` 0 = keep, -10000 = discard. + + If `ndim == 2`: will be interpreted as a mask, then converted into a bias consistent with the format + above. This bias will be added to the cross-attention scores. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] instead of a plain + tuple. + + Returns: + If `return_dict` is True, an [`~models.transformers.transformer_2d.Transformer2DModelOutput`] is returned, + otherwise a `tuple` where the first element is the sample tensor. + """ + if cross_attention_kwargs is not None: + if cross_attention_kwargs.get("scale", None) is not None: + logger.warning("Passing `scale` to `cross_attention_kwargs` is deprecated. `scale` will be ignored.") + + if attention_mask is not None and attention_mask.ndim == 2: + attention_mask = (1 - attention_mask.to(hidden_states.dtype)) * -10000.0 + attention_mask = attention_mask.unsqueeze(1) + + # convert encoder_attention_mask to a bias the same way we do for attention_mask + if encoder_attention_mask is not None and encoder_attention_mask.ndim == 2: + encoder_attention_mask = (1 - encoder_attention_mask.to(hidden_states.dtype)) * -10000.0 + encoder_attention_mask = encoder_attention_mask.unsqueeze(1) + + # 1. Input + if self.is_input_continuous: + batch_size, _, height, width = hidden_states.shape + residual = hidden_states + hidden_states, inner_dim = self._operate_on_continuous_inputs(hidden_states) + elif self.is_input_vectorized: + hidden_states = self.latent_image_embedding(hidden_states) + elif self.is_input_patches: + height, width = hidden_states.shape[-2] // self.patch_size, hidden_states.shape[-1] // self.patch_size + hidden_states, encoder_hidden_states, timestep, embedded_timestep = self._operate_on_patched_inputs( + hidden_states, encoder_hidden_states, timestep, added_cond_kwargs + ) + + # 2. Blocks + for block in self.transformer_blocks: + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {} + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(block), + hidden_states, + attention_mask, + encoder_hidden_states, + encoder_attention_mask, + timestep, + cross_attention_kwargs, + class_labels, + **ckpt_kwargs, + ) + else: + hidden_states = block( + hidden_states, + attention_mask=attention_mask, + encoder_hidden_states=encoder_hidden_states, + encoder_attention_mask=encoder_attention_mask, + timestep=timestep, + cross_attention_kwargs=cross_attention_kwargs, + class_labels=class_labels, + ) + + # 3. Output + if self.is_input_continuous: + output = self._get_output_for_continuous_inputs( + hidden_states=hidden_states, + residual=residual, + batch_size=batch_size, + height=height, + width=width, + inner_dim=inner_dim, + ) + elif self.is_input_vectorized: + output = self._get_output_for_vectorized_inputs(hidden_states) + elif self.is_input_patches: + output = self._get_output_for_patched_inputs( + hidden_states=hidden_states, + timestep=timestep, + class_labels=class_labels, + embedded_timestep=embedded_timestep, + height=height, + width=width, + ) + + if not return_dict: + return (output,) + + return Transformer2DModelOutput(sample=output) + + def _operate_on_continuous_inputs(self, hidden_states): + batch, _, height, width = hidden_states.shape + hidden_states = self.norm(hidden_states) + + if not self.use_linear_projection: + hidden_states = self.proj_in(hidden_states) + inner_dim = hidden_states.shape[1] + hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * width, inner_dim) + else: + inner_dim = hidden_states.shape[1] + hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * width, inner_dim) + hidden_states = self.proj_in(hidden_states) + + return hidden_states, inner_dim + + def _operate_on_patched_inputs(self, hidden_states, encoder_hidden_states, timestep, added_cond_kwargs): + batch_size = hidden_states.shape[0] + hidden_states = self.pos_embed(hidden_states) + embedded_timestep = None + + if self.adaln_single is not None: + if self.use_additional_conditions and added_cond_kwargs is None: + raise ValueError( + "`added_cond_kwargs` cannot be None when using additional conditions for `adaln_single`." + ) + timestep, embedded_timestep = self.adaln_single( + timestep, added_cond_kwargs, batch_size=batch_size, hidden_dtype=hidden_states.dtype + ) + + if self.caption_projection is not None: + encoder_hidden_states = self.caption_projection(encoder_hidden_states) + encoder_hidden_states = encoder_hidden_states.view(batch_size, -1, hidden_states.shape[-1]) + + return hidden_states, encoder_hidden_states, timestep, embedded_timestep + + def _get_output_for_continuous_inputs(self, hidden_states, residual, batch_size, height, width, inner_dim): + if not self.use_linear_projection: + hidden_states = ( + hidden_states.reshape(batch_size, height, width, inner_dim).permute(0, 3, 1, 2).contiguous() + ) + hidden_states = self.proj_out(hidden_states) + else: + hidden_states = self.proj_out(hidden_states) + hidden_states = ( + hidden_states.reshape(batch_size, height, width, inner_dim).permute(0, 3, 1, 2).contiguous() + ) + + output = hidden_states + residual + return output + + def _get_output_for_vectorized_inputs(self, hidden_states): + hidden_states = self.norm_out(hidden_states) + logits = self.out(hidden_states) + logits = logits.permute(0, 2, 1) + output = F.log_softmax(logits.double(), dim=1).float() + return output + + def _get_output_for_patched_inputs( + self, hidden_states, timestep, class_labels, embedded_timestep, height=None, width=None + ): + if self.norm_type != "ada_norm_single": + conditioning = self.transformer_blocks[0].norm1.emb( + timestep, class_labels, hidden_dtype=hidden_states.dtype + ) + shift, scale = self.proj_out_1(F.silu(conditioning)).chunk(2, dim=1) + hidden_states = self.norm_out(hidden_states) * (1 + scale[:, None]) + shift[:, None] + hidden_states = self.proj_out_2(hidden_states) + elif self.norm_type == "ada_norm_single": + shift, scale = (self.scale_shift_table[None] + embedded_timestep[:, None]).chunk(2, dim=1) + hidden_states = self.norm_out(hidden_states) + # Modulation + hidden_states = hidden_states * (1 + scale) + shift + hidden_states = self.proj_out(hidden_states) + hidden_states = hidden_states.squeeze(1) + + # unpatchify + if self.adaln_single is None: + height = width = int(hidden_states.shape[1] ** 0.5) + hidden_states = hidden_states.reshape( + shape=(-1, height, width, self.patch_size, self.patch_size, self.out_channels) + ) + hidden_states = torch.einsum("nhwpqc->nchpwq", hidden_states) + output = hidden_states.reshape( + shape=(-1, self.out_channels, height * self.patch_size, width * self.patch_size) + ) + return output \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-XL/stable_diffusion_xl/unet/transformer_blocks.py b/MindIE/MultiModal/StableDiffusion-XL/stable_diffusion_xl/unet/transformer_blocks.py new file mode 100644 index 0000000000000000000000000000000000000000..f43e316b5486733a71edac76debfa60c2cda605b --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-XL/stable_diffusion_xl/unet/transformer_blocks.py @@ -0,0 +1,467 @@ +# Copyright 2024 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from typing import Optional, Dict, Any + +import torch +import torch.nn as nn + +from diffusers.models.normalization import AdaLayerNorm, AdaLayerNormZero, AdaLayerNormContinuous +from diffusers.models.attention import FeedForward +from ..layers.attention import Attention, AttnProcessor2_0 + + +class BasicTransformerBlock(nn.Module): + r""" + A basic Transformer block. + + Parameters: + dim (`int`): The number of channels in the input and output. + num_attention_heads (`int`): The number of heads to use for multi-head attention. + attention_head_dim (`int`): The number of channels in each head. + dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. + cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention. + activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward. + num_embeds_ada_norm (: + obj: `int`, *optional*): The number of diffusion steps used during training. See `Transformer2DModel`. + attention_bias (: + obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter. + only_cross_attention (`bool`, *optional*): + Whether to use only cross-attention layers. In this case two cross attention layers are used. + double_self_attention (`bool`, *optional*): + Whether to use two self-attention layers. In this case no cross attention layers are used. + upcast_attention (`bool`, *optional*): + Whether to upcast the attention computation to float32. This is useful for mixed precision training. + norm_elementwise_affine (`bool`, *optional*, defaults to `True`): + Whether to use learnable elementwise affine parameters for normalization. + norm_type (`str`, *optional*, defaults to `"layer_norm"`): + The normalization layer to use. Can be `"layer_norm"`, `"ada_norm"` or `"ada_norm_zero"`. + final_dropout (`bool` *optional*, defaults to False): + Whether to apply a final dropout after the last feed-forward layer. + attention_type (`str`, *optional*, defaults to `"default"`): + The type of attention to use. Can be `"default"` or `"gated"` or `"gated-text-image"`. + positional_embeddings (`str`, *optional*, defaults to `None`): + The type of positional embeddings to apply to. + num_positional_embeddings (`int`, *optional*, defaults to `None`): + The maximum number of positional embeddings to apply. + """ + + def __init__( + self, + dim: int, + num_attention_heads: int, + attention_head_dim: int, + dropout=0.0, + cross_attention_dim: Optional[int] = None, + activation_fn: str = "geglu", + num_embeds_ada_norm: Optional[int] = None, + attention_bias: bool = False, + only_cross_attention: bool = False, + double_self_attention: bool = False, + upcast_attention: bool = False, + norm_elementwise_affine: bool = True, + norm_type: str = "layer_norm", # 'layer_norm', 'ada_norm', 'ada_norm_zero', 'ada_norm_single', 'ada_norm_continuous', 'layer_norm_i2vgen' + norm_eps: float = 1e-5, + final_dropout: bool = False, + attention_type: str = "default", + positional_embeddings: Optional[str] = None, + num_positional_embeddings: Optional[int] = None, + ada_norm_continous_conditioning_embedding_dim: Optional[int] = None, + ada_norm_bias: Optional[int] = None, + ff_inner_dim: Optional[int] = None, + ff_bias: bool = True, + attention_out_bias: bool = True, + ): + super().__init__() + self.only_cross_attention = only_cross_attention + + # We keep these boolean flags for backward-compatibility. + self.use_ada_layer_norm_zero = (num_embeds_ada_norm is not None) and norm_type == "ada_norm_zero" + self.use_ada_layer_norm = (num_embeds_ada_norm is not None) and norm_type == "ada_norm" + self.use_ada_layer_norm_single = norm_type == "ada_norm_single" + self.use_layer_norm = norm_type == "layer_norm" + self.use_ada_layer_norm_continuous = norm_type == "ada_norm_continuous" + + if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None: + raise ValueError( + f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to" + f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}." + ) + + self.norm_type = norm_type + self.num_embeds_ada_norm = num_embeds_ada_norm + + if positional_embeddings and (num_positional_embeddings is None): + raise ValueError( + "If `positional_embedding` type is defined, `num_positition_embeddings` must also be defined." + ) + + if positional_embeddings == "sinusoidal": + self.pos_embed = SinusoidalPositionalEmbedding(dim, max_seq_length=num_positional_embeddings) + else: + self.pos_embed = None + + # Define 3 blocks. Each block has its own normalization layer. + # 1. Self-Attn + if norm_type == "ada_norm": + self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm) + elif norm_type == "ada_norm_zero": + self.norm1 = AdaLayerNormZero(dim, num_embeds_ada_norm) + elif norm_type == "ada_norm_continuous": + self.norm1 = AdaLayerNormContinuous( + dim, + ada_norm_continous_conditioning_embedding_dim, + norm_elementwise_affine, + norm_eps, + ada_norm_bias, + "rms_norm", + ) + else: + self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps) + + self.attn1 = Attention( + query_dim=dim, + heads=num_attention_heads, + dim_head=attention_head_dim, + dropout=dropout, + bias=attention_bias, + cross_attention_dim=cross_attention_dim if only_cross_attention else None, + upcast_attention=upcast_attention, + out_bias=attention_out_bias, + ) + + # 2. Cross-Attn + if cross_attention_dim is not None or double_self_attention: + # We currently only use AdaLayerNormZero for self attention where there will only be one attention block. + # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during + # the second cross attention block. + if norm_type == "ada_norm": + self.norm2 = AdaLayerNorm(dim, num_embeds_ada_norm) + elif norm_type == "ada_norm_continuous": + self.norm2 = AdaLayerNormContinuous( + dim, + ada_norm_continous_conditioning_embedding_dim, + norm_elementwise_affine, + norm_eps, + ada_norm_bias, + "rms_norm", + ) + else: + self.norm2 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine) + + self.attn2 = Attention( + query_dim=dim, + cross_attention_dim=cross_attention_dim if not double_self_attention else None, + heads=num_attention_heads, + dim_head=attention_head_dim, + dropout=dropout, + bias=attention_bias, + upcast_attention=upcast_attention, + out_bias=attention_out_bias, + ) # is self-attn if encoder_hidden_states is none + else: + self.norm2 = None + self.attn2 = None + + # 3. Feed-forward + if norm_type == "ada_norm_continuous": + self.norm3 = AdaLayerNormContinuous( + dim, + ada_norm_continous_conditioning_embedding_dim, + norm_elementwise_affine, + norm_eps, + ada_norm_bias, + "layer_norm", + ) + + elif norm_type in ["ada_norm_zero", "ada_norm", "layer_norm", "ada_norm_continuous"]: + self.norm3 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine) + elif norm_type == "layer_norm_i2vgen": + self.norm3 = None + + self.ff = FeedForward( + dim, + dropout=dropout, + activation_fn=activation_fn, + final_dropout=final_dropout, + inner_dim=ff_inner_dim, + bias=ff_bias, + ) + + # 4. Fuser + if attention_type == "gated" or attention_type == "gated-text-image": + raise ValueError(f"attention_type:{attention_type} is not supported!") + + # 5. Scale-shift for PixArt-Alpha. + if norm_type == "ada_norm_single": + self.scale_shift_table = nn.Parameter(torch.randn(6, dim) / dim**0.5) + + # let chunk size default to None + self._chunk_size = None + self._chunk_dim = 0 + + # agb + self.enable_agb = False + + self.attn_count = 0 + self.last_attn_x = None + self.last_attn_x_before = None + self.attn_alpha_min = float('inf') + self.attn_alpha = 0 + self.attn_alpha_max = -float('inf') + self.last_attn = None + + self.cross_count = 0 + self.last_cross_x = None + self.last_cross_x_before = None + self.cross_alpha_min = float('inf') + self.cross_alpha = 0 + self.cross_alpha_max = -float('inf') + self.last_cross = None + + self.bound = [10, 2] + self.forcefresh = 6 + self.totalstep = 50 + + def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int = 0): + # Sets chunk feed-forward + self._chunk_size = chunk_size + self._chunk_dim = dim + + def forward( + self, + hidden_states: torch.Tensor, + attention_mask: Optional[torch.Tensor] = None, + encoder_hidden_states: Optional[torch.Tensor] = None, + encoder_attention_mask: Optional[torch.Tensor] = None, + timestep: Optional[torch.LongTensor] = None, + cross_attention_kwargs: Dict[str, Any] = None, + class_labels: Optional[torch.LongTensor] = None, + added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None, + ) -> torch.Tensor: + if cross_attention_kwargs is not None: + if cross_attention_kwargs.get("scale", None) is not None: + logger.warning("Passing `scale` to `cross_attention_kwargs` is deprecated. `scale` will be ignored.") + + # Notice that normalization is always applied before the real computation in the following blocks. + batch_size = hidden_states.shape[0] + # 0. Prepare GLIGEN inputs + cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {} + gligen_kwargs = cross_attention_kwargs.pop("gligen", None) + + # 1. Self-Attention + if self.enable_agb: + if (self.bound[0] - 0.5 < self.attn_count < self.totalstep - self.bound[1] - 0.5) and \ + (self.attn_count % self.forcefresh != 0): + broadcast_attn = True + else: + broadcast_attn = False + self.attn_count = (self.attn_count + 1) % self.totalstep + self.last_attn_x = hidden_states + if broadcast_attn: + hidden_states = self.last_attn + hidden_states + if hidden_states.ndim == 4: + hidden_states = hidden_states.squeeze(1) + else: + if self.norm_type == "ada_norm": + norm_hidden_states = self.norm1(hidden_states, timestep) + elif self.norm_type == "ada_norm_zero": + norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1( + hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype + ) + elif self.norm_type in ["layer_norm", "layer_norm_i2vgen"]: + norm_hidden_states = self.norm1(hidden_states) + elif self.norm_type == "ada_norm_continuous": + norm_hidden_states = self.norm1(hidden_states, added_cond_kwargs["pooled_text_emb"]) + elif self.norm_type == "ada_norm_single": + shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = ( + self.scale_shift_table[None] + timestep.reshape(batch_size, 6, -1) + ).chunk(6, dim=1) + norm_hidden_states = self.norm1(hidden_states) + norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa + norm_hidden_states = norm_hidden_states.squeeze(1) + else: + raise ValueError("Incorrect norm used") + + if self.pos_embed is not None: + norm_hidden_states = self.pos_embed(norm_hidden_states) + + attn_output = self.attn1( + norm_hidden_states, + encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None, + attention_mask=attention_mask, + **cross_attention_kwargs, + ) + if self.norm_type == "ada_norm_zero": + attn_output = gate_msa.unsqueeze(1) * attn_output + elif self.norm_type == "ada_norm_single": + attn_output = gate_msa * attn_output + + hidden_states = attn_output + hidden_states + if hidden_states.ndim == 4: + hidden_states = hidden_states.squeeze(1) + + # 1.2 GLIGEN Control + if gligen_kwargs is not None: + raise ValueError("'gligen' is not supported now!") + + # 3. Cross-Attention + if self.attn2 is not None: + if self.norm_type == "ada_norm": + norm_hidden_states = self.norm2(hidden_states, timestep) + elif self.norm_type in ["ada_norm_zero", "layer_norm", "layer_norm_i2vgen"]: + norm_hidden_states = self.norm2(hidden_states) + elif self.norm_type == "ada_norm_single": + norm_hidden_states = hidden_states + elif self.norm_type == "ada_norm_continuous": + norm_hidden_states = self.norm2(hidden_states, added_cond_kwargs["pooled_text_emb"]) + else: + raise ValueError("Incorrect norm") + + if self.pos_embed is not None and self.norm_type != "ada_norm_single": + norm_hidden_states = self.pos_embed(norm_hidden_states) + + attn_output2 = self.attn2( + norm_hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=encoder_attention_mask, + **cross_attention_kwargs, + ) + hidden_states = attn_output2 + hidden_states + + # 4. Feed-forward + # i2vgen doesn't have this norm 🤷‍♂️ + if self.norm_type == "ada_norm_continuous": + norm_hidden_states = self.norm3(hidden_states, added_cond_kwargs["pooled_text_emb"]) + elif not self.norm_type == "ada_norm_single": + norm_hidden_states = self.norm3(hidden_states) + + if self.norm_type == "ada_norm_zero": + norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None] + + if self.norm_type == "ada_norm_single": + norm_hidden_states = self.norm2(hidden_states) + norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp + + if self._chunk_size is not None: + # "feed_forward_chunk_size" can be used to save memory + ff_output = _chunked_feed_forward(self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size) + else: + ff_output = self.ff(norm_hidden_states) + + if self.norm_type == "ada_norm_zero": + ff_output = gate_mlp.unsqueeze(1) * ff_output + elif self.norm_type == "ada_norm_single": + ff_output = gate_mlp * ff_output + + hidden_states = ff_output + hidden_states + if hidden_states.ndim == 4: + hidden_states = hidden_states.squeeze(1) + self.last_attn = attn_output + attn_output2 + ff_output + else: + if self.norm_type == "ada_norm": + norm_hidden_states = self.norm1(hidden_states, timestep) + elif self.norm_type == "ada_norm_zero": + norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1( + hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype + ) + elif self.norm_type in ["layer_norm", "layer_norm_i2vgen"]: + norm_hidden_states = self.norm1(hidden_states) + elif self.norm_type == "ada_norm_continuous": + norm_hidden_states = self.norm1(hidden_states, added_cond_kwargs["pooled_text_emb"]) + elif self.norm_type == "ada_norm_single": + shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = ( + self.scale_shift_table[None] + timestep.reshape(batch_size, 6, -1) + ).chunk(6, dim=1) + norm_hidden_states = self.norm1(hidden_states) + norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa + norm_hidden_states = norm_hidden_states.squeeze(1) + else: + raise ValueError("Incorrect norm used") + + if self.pos_embed is not None: + norm_hidden_states = self.pos_embed(norm_hidden_states) + + attn_output = self.attn1( + norm_hidden_states, + encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None, + attention_mask=attention_mask, + **cross_attention_kwargs, + ) + if self.norm_type == "ada_norm_zero": + attn_output = gate_msa.unsqueeze(1) * attn_output + elif self.norm_type == "ada_norm_single": + attn_output = gate_msa * attn_output + + hidden_states = attn_output + hidden_states + if hidden_states.ndim == 4: + hidden_states = hidden_states.squeeze(1) + + # 1.2 GLIGEN Control + if gligen_kwargs is not None: + raise ValueError("'gligen' is not supported now!") + + # 3. Cross-Attention + if self.attn2 is not None: + if self.norm_type == "ada_norm": + norm_hidden_states = self.norm2(hidden_states, timestep) + elif self.norm_type in ["ada_norm_zero", "layer_norm", "layer_norm_i2vgen"]: + norm_hidden_states = self.norm2(hidden_states) + elif self.norm_type == "ada_norm_single": + norm_hidden_states = hidden_states + elif self.norm_type == "ada_norm_continuous": + norm_hidden_states = self.norm2(hidden_states, added_cond_kwargs["pooled_text_emb"]) + else: + raise ValueError("Incorrect norm") + + if self.pos_embed is not None and self.norm_type != "ada_norm_single": + norm_hidden_states = self.pos_embed(norm_hidden_states) + + attn_output = self.attn2( + norm_hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=encoder_attention_mask, + **cross_attention_kwargs, + ) + hidden_states = attn_output + hidden_states + + # 4. Feed-forward + # i2vgen doesn't have this norm 🤷‍♂️ + if self.norm_type == "ada_norm_continuous": + norm_hidden_states = self.norm3(hidden_states, added_cond_kwargs["pooled_text_emb"]) + elif not self.norm_type == "ada_norm_single": + norm_hidden_states = self.norm3(hidden_states) + + if self.norm_type == "ada_norm_zero": + norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None] + + if self.norm_type == "ada_norm_single": + norm_hidden_states = self.norm2(hidden_states) + norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp + + if self._chunk_size is not None: + # "feed_forward_chunk_size" can be used to save memory + ff_output = _chunked_feed_forward(self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size) + else: + ff_output = self.ff(norm_hidden_states) + + if self.norm_type == "ada_norm_zero": + ff_output = gate_mlp.unsqueeze(1) * ff_output + elif self.norm_type == "ada_norm_single": + ff_output = gate_mlp * ff_output + + hidden_states = ff_output + hidden_states + if hidden_states.ndim == 4: + hidden_states = hidden_states.squeeze(1) + + return hidden_states \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-XL/stable_diffusion_xl/unet/unet_blocks.py b/MindIE/MultiModal/StableDiffusion-XL/stable_diffusion_xl/unet/unet_blocks.py new file mode 100644 index 0000000000000000000000000000000000000000..de030f72306432e4ab2ae41c851b7bdaa1e4986e --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-XL/stable_diffusion_xl/unet/unet_blocks.py @@ -0,0 +1,579 @@ +# Copyright 2024 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from typing import Optional, Union, Tuple, Dict, Any +import torch +import torch.nn.functional as F +import torch.nn as nn + +from diffusers.models.resnet import ResnetBlock2D +from diffusers.models.unets.unet_2d_blocks import UpBlock2D, DownBlock2D +from diffusers.models.downsampling import Downsample2D +from diffusers.models.upsampling import Upsample2D +from .transformer_2d_model import Transformer2DModel + + +def get_down_block( + down_block_type: str, + num_layers: int, + in_channels: int, + out_channels: int, + temb_channels: int, + add_downsample: bool, + resnet_eps: float, + resnet_act_fn: str, + transformer_layers_per_block: int = 1, + num_attention_heads: Optional[int] = None, + resnet_groups: Optional[int] = None, + cross_attention_dim: Optional[int] = None, + downsample_padding: Optional[int] = None, + dual_cross_attention: bool = False, + use_linear_projection: bool = False, + only_cross_attention: bool = False, + upcast_attention: bool = False, + resnet_time_scale_shift: str = "default", + attention_type: str = "default", + resnet_skip_time_act: bool = False, + resnet_out_scale_factor: float = 1.0, + cross_attention_norm: Optional[str] = None, + attention_head_dim: Optional[int] = None, + downsample_type: Optional[str] = None, + dropout: float = 0.0, +): + # If attn head dim is not defined, we default it to the number of heads + if attention_head_dim is None: + logger.warning( + f"It is recommended to provide `attention_head_dim` when calling `get_down_block`. Defaulting `attention_head_dim` to {num_attention_heads}." + ) + attention_head_dim = num_attention_heads + + down_block_type = down_block_type[7:] if down_block_type.startswith("UNetRes") else down_block_type + if down_block_type == "DownBlock2D": + return DownBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + downsample_padding=downsample_padding, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + elif down_block_type == "CrossAttnDownBlock2D": + if cross_attention_dim is None: + raise ValueError("cross_attention_dim must be specified for CrossAttnDownBlock2D") + return CrossAttnDownBlock2D( + num_layers=num_layers, + transformer_layers_per_block=transformer_layers_per_block, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + downsample_padding=downsample_padding, + cross_attention_dim=cross_attention_dim, + num_attention_heads=num_attention_heads, + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention, + upcast_attention=upcast_attention, + resnet_time_scale_shift=resnet_time_scale_shift, + attention_type=attention_type, + ) + raise ValueError(f"{down_block_type} does not exist.") + + +def get_mid_block( + mid_block_type: str, + temb_channels: int, + in_channels: int, + resnet_eps: float, + resnet_act_fn: str, + resnet_groups: int, + output_scale_factor: float = 1.0, + transformer_layers_per_block: int = 1, + num_attention_heads: Optional[int] = None, + cross_attention_dim: Optional[int] = None, + dual_cross_attention: bool = False, + use_linear_projection: bool = False, + mid_block_only_cross_attention: bool = False, + upcast_attention: bool = False, + resnet_time_scale_shift: str = "default", + attention_type: str = "default", + resnet_skip_time_act: bool = False, + cross_attention_norm: Optional[str] = None, + attention_head_dim: Optional[int] = 1, + dropout: float = 0.0, +): + if mid_block_type == "UNetMidBlock2DCrossAttn": + return UNetMidBlock2DCrossAttn( + transformer_layers_per_block=transformer_layers_per_block, + in_channels=in_channels, + temb_channels=temb_channels, + dropout=dropout, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + output_scale_factor=output_scale_factor, + resnet_time_scale_shift=resnet_time_scale_shift, + cross_attention_dim=cross_attention_dim, + num_attention_heads=num_attention_heads, + resnet_groups=resnet_groups, + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, + upcast_attention=upcast_attention, + attention_type=attention_type, + ) + else: + raise ValueError(f"unknown mid_block_type : {mid_block_type}") + + +def get_up_block( + up_block_type: str, + num_layers: int, + in_channels: int, + out_channels: int, + prev_output_channel: int, + temb_channels: int, + add_upsample: bool, + resnet_eps: float, + resnet_act_fn: str, + resolution_idx: Optional[int] = None, + transformer_layers_per_block: int = 1, + num_attention_heads: Optional[int] = None, + resnet_groups: Optional[int] = None, + cross_attention_dim: Optional[int] = None, + dual_cross_attention: bool = False, + use_linear_projection: bool = False, + only_cross_attention: bool = False, + upcast_attention: bool = False, + resnet_time_scale_shift: str = "default", + attention_type: str = "default", + resnet_skip_time_act: bool = False, + resnet_out_scale_factor: float = 1.0, + cross_attention_norm: Optional[str] = None, + attention_head_dim: Optional[int] = None, + upsample_type: Optional[str] = None, + dropout: float = 0.0, +) -> nn.Module: + # If attn head dim is not defined, we default it to the number of heads + if attention_head_dim is None: + logger.warning( + f"It is recommended to provide `attention_head_dim` when calling `get_up_block`. Defaulting `attention_head_dim` to {num_attention_heads}." + ) + attention_head_dim = num_attention_heads + + up_block_type = up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type + if up_block_type == "UpBlock2D": + return UpBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + prev_output_channel=prev_output_channel, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + elif up_block_type == "CrossAttnUpBlock2D": + if cross_attention_dim is None: + raise ValueError("cross_attention_dim must be specified for CrossAttnUpBlock2D") + return CrossAttnUpBlock2D( + num_layers=num_layers, + transformer_layers_per_block=transformer_layers_per_block, + in_channels=in_channels, + out_channels=out_channels, + prev_output_channel=prev_output_channel, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + cross_attention_dim=cross_attention_dim, + num_attention_heads=num_attention_heads, + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention, + upcast_attention=upcast_attention, + resnet_time_scale_shift=resnet_time_scale_shift, + attention_type=attention_type, + ) + raise ValueError(f"{up_block_type} does not exist.") + + +class CrossAttnDownBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + transformer_layers_per_block: Union[int, Tuple[int]] = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + num_attention_heads: int = 1, + cross_attention_dim: int = 1280, + output_scale_factor: float = 1.0, + downsample_padding: int = 1, + add_downsample: bool = True, + dual_cross_attention: bool = False, # now is not used. + use_linear_projection: bool = False, + only_cross_attention: bool = False, + upcast_attention: bool = False, + attention_type: str = "default", + ): + super().__init__() + resnets = [] + attentions = [] + + self.has_cross_attention = True + self.num_attention_heads = num_attention_heads + if isinstance(transformer_layers_per_block, int): + transformer_layers_per_block = [transformer_layers_per_block] * num_layers + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + if not dual_cross_attention: + attentions.append( + Transformer2DModel( + num_attention_heads, + out_channels // num_attention_heads, + in_channels=out_channels, + num_layers=transformer_layers_per_block[i], + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention, + upcast_attention=upcast_attention, + attention_type=attention_type, + ) + ) + else: + raise ValueError("DualTransformer2DModel is not supported now.") + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if add_downsample: + self.downsamplers = nn.ModuleList( + [ + Downsample2D( + out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op" + ) + ] + ) + else: + self.downsamplers = None + + def forward( + self, + hidden_states: torch.Tensor, + temb: Optional[torch.Tensor] = None, + encoder_hidden_states: Optional[torch.Tensor] = None, + attention_mask: Optional[torch.Tensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + encoder_attention_mask: Optional[torch.Tensor] = None, + additional_residuals: Optional[torch.Tensor] = None, + ) -> Tuple[torch.Tensor, Tuple[torch.Tensor, ...]]: + + output_states = () + + blocks = list(zip(self.resnets, self.attentions)) + + for i, (resnet, attn) in enumerate(blocks): + hidden_states = resnet(hidden_states, temb) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + return_dict=False, + )[0] + + # apply additional residuals to the output of the last pair of resnet and attention blocks + if i == len(blocks) - 1 and additional_residuals is not None: + hidden_states = hidden_states + additional_residuals + output_states = output_states + (hidden_states,) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states) + + output_states = output_states + (hidden_states,) + + return hidden_states, output_states + + +class UNetMidBlock2DCrossAttn(nn.Module): + def __init__( + self, + in_channels: int, + temb_channels: int, + out_channels: Optional[int] = None, + dropout: float = 0.0, + num_layers: int = 1, + transformer_layers_per_block: Union[int, Tuple[int]] = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_groups_out: Optional[int] = None, + resnet_pre_norm: bool = True, + num_attention_heads: int = 1, + output_scale_factor: float = 1.0, + cross_attention_dim: int = 1280, + dual_cross_attention: bool = False, + use_linear_projection: bool = False, + upcast_attention: bool = False, + attention_type: str = "default", + ): + super().__init__() + + out_channels = out_channels or in_channels + self.in_channels = in_channels + self.out_channels = out_channels + + self.has_cross_attention = True + self.num_attention_heads = num_attention_heads + resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32) + + # support for variable transformer layers per block + if isinstance(transformer_layers_per_block, int): + transformer_layers_per_block = [transformer_layers_per_block] * num_layers + + resnet_groups_out = resnet_groups_out or resnet_groups + + # there is always at least one resnet + resnets = [ + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + groups_out=resnet_groups_out, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ] + attentions = [] + + for i in range(num_layers): + if not dual_cross_attention: + attentions.append( + Transformer2DModel( + num_attention_heads, + out_channels // num_attention_heads, + in_channels=out_channels, + num_layers=transformer_layers_per_block[i], + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups_out, + use_linear_projection=use_linear_projection, + upcast_attention=upcast_attention, + attention_type=attention_type, + ) + ) + else: + raise ValueError("DualTransformer2DModel is not supported now.") + resnets.append( + ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups_out, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + def forward( + self, + hidden_states: torch.Tensor, + temb: Optional[torch.Tensor] = None, + encoder_hidden_states: Optional[torch.Tensor] = None, + attention_mask: Optional[torch.Tensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + encoder_attention_mask: Optional[torch.Tensor] = None, + ) -> torch.Tensor: + if cross_attention_kwargs is not None: + if cross_attention_kwargs.get("scale", None) is not None: + logger.warning("Passing `scale` to `cross_attention_kwargs` is deprecated. `scale` will be ignored.") + + hidden_states = self.resnets[0](hidden_states, temb) + for attn, resnet in zip(self.attentions, self.resnets[1:]): + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + return_dict=False, + )[0] + hidden_states = resnet(hidden_states, temb) + + return hidden_states + + +class CrossAttnUpBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + prev_output_channel: int, + temb_channels: int, + resolution_idx: Optional[int] = None, + dropout: float = 0.0, # will not used in inference mode + num_layers: int = 1, + transformer_layers_per_block: Union[int, Tuple[int]] = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + num_attention_heads: int = 1, + cross_attention_dim: int = 1280, + output_scale_factor: float = 1.0, + add_upsample: bool = True, + dual_cross_attention: bool = False, + use_linear_projection: bool = False, + only_cross_attention: bool = False, + upcast_attention: bool = False, + attention_type: str = "default", + ): + super().__init__() + resnets = [] + attentions = [] + + self.has_cross_attention = True + self.num_attention_heads = num_attention_heads + + if isinstance(transformer_layers_per_block, int): + transformer_layers_per_block = [transformer_layers_per_block] * num_layers + + for i in range(num_layers): + res_skip_channels = in_channels if (i == num_layers - 1) else out_channels + resnet_in_channels = prev_output_channel if i == 0 else out_channels + + resnets.append( + ResnetBlock2D( + in_channels=resnet_in_channels + res_skip_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + if not dual_cross_attention: + attentions.append( + Transformer2DModel( + num_attention_heads, + out_channels // num_attention_heads, + in_channels=out_channels, + num_layers=transformer_layers_per_block[i], + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention, + upcast_attention=upcast_attention, + attention_type=attention_type, + ) + ) + else: + raise ValueError("DualTransformer2DModel is not supported now.") + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if add_upsample: + self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)]) + else: + self.upsamplers = None + + + def forward( + self, + hidden_states: torch.Tensor, + res_hidden_states_tuple: Tuple[torch.Tensor, ...], + temb: Optional[torch.Tensor] = None, + encoder_hidden_states: Optional[torch.Tensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + upsample_size: Optional[int] = None, + attention_mask: Optional[torch.Tensor] = None, + encoder_attention_mask: Optional[torch.Tensor] = None, + ) -> torch.Tensor: + + for resnet, attn in zip(self.resnets, self.attentions): + # pop res hidden states + res_hidden_states = res_hidden_states_tuple[-1] + res_hidden_states_tuple = res_hidden_states_tuple[:-1] + + hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1) + + hidden_states = resnet(hidden_states, temb) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + return_dict=False, + )[0] + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states, upsample_size) + + return hidden_states \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-XL/stable_diffusion_xl/unet/unet_model.py b/MindIE/MultiModal/StableDiffusion-XL/stable_diffusion_xl/unet/unet_model.py new file mode 100644 index 0000000000000000000000000000000000000000..947c3f0319754bee6d136666ba19c3c813d443e6 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-XL/stable_diffusion_xl/unet/unet_model.py @@ -0,0 +1,1242 @@ +# Copyright 2024 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from dataclasses import dataclass +from typing import Optional, Tuple, Union, Dict, Any +import torch +import torch.nn as nn +import torch_npu + +from diffusers.models.embeddings import ( + Timesteps, + TimestepEmbedding, + TextImageProjection, + ImageProjection, + TextTimeEmbedding, + TextImageTimeEmbedding, + ImageTimeEmbedding, + ImageHintTimeEmbedding +) + +from diffusers.models.activations import get_activation +from diffusers.models.modeling_utils import ModelMixin +from diffusers.configuration_utils import ConfigMixin, register_to_config +from diffusers.loaders import PeftAdapterMixin, UNet2DConditionLoadersMixin +from diffusers.loaders.single_file_model import FromOriginalModelMixin +from diffusers.utils import USE_PEFT_BACKEND, BaseOutput, scale_lora_layers, unscale_lora_layers + +from .unet_blocks import ( + get_down_block, + get_mid_block, + get_up_block +) + + +@dataclass +class UNet2DConditionOutput(BaseOutput): + """ + The output of [`UNet2DConditionModel`]. + + Args: + sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)`): + The hidden states output conditioned on `encoder_hidden_states` input. Output of last layer of model. + """ + + sample: torch.Tensor = None + + +class UNet2DConditionModel( + ModelMixin, ConfigMixin, FromOriginalModelMixin, UNet2DConditionLoadersMixin, PeftAdapterMixin +): + r""" + A conditional 2D UNet model that takes a noisy sample, conditional state, and a timestep and returns a sample + shaped output. + + This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented + for all models (such as downloading or saving). + + Parameters: + sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`): + Height and width of input/output sample. + in_channels (`int`, *optional*, defaults to 4): Number of channels in the input sample. + out_channels (`int`, *optional*, defaults to 4): Number of channels in the output. + center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample. + flip_sin_to_cos (`bool`, *optional*, defaults to `True`): + Whether to flip the sin to cos in the time embedding. + freq_shift (`int`, *optional*, defaults to 0): The frequency shift to apply to the time embedding. + down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`): + The tuple of downsample blocks to use. + mid_block_type (`str`, *optional*, defaults to `"UNetMidBlock2DCrossAttn"`): + Block type for middle of UNet, it can be one of `UNetMidBlock2DCrossAttn`, `UNetMidBlock2D`, or + `UNetMidBlock2DSimpleCrossAttn`. If `None`, the mid block layer is skipped. + up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D")`): + The tuple of upsample blocks to use. + only_cross_attention(`bool` or `Tuple[bool]`, *optional*, default to `False`): + Whether to include self-attention in the basic transformer blocks, see + [`~models.attention.BasicTransformerBlock`]. + block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`): + The tuple of output channels for each block. + layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block. + downsample_padding (`int`, *optional*, defaults to 1): The padding to use for the downsampling convolution. + mid_block_scale_factor (`float`, *optional*, defaults to 1.0): The scale factor to use for the mid block. + dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. + act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use. + norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for the normalization. + If `None`, normalization and activation layers is skipped in post-processing. + norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization. + cross_attention_dim (`int` or `Tuple[int]`, *optional*, defaults to 1280): + The dimension of the cross attention features. + transformer_layers_per_block (`int`, `Tuple[int]`, or `Tuple[Tuple]` , *optional*, defaults to 1): + The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for + [`~models.unets.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unets.unet_2d_blocks.CrossAttnUpBlock2D`], + [`~models.unets.unet_2d_blocks.UNetMidBlock2DCrossAttn`]. + reverse_transformer_layers_per_block : (`Tuple[Tuple]`, *optional*, defaults to None): + The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`], in the upsampling + blocks of the U-Net. Only relevant if `transformer_layers_per_block` is of type `Tuple[Tuple]` and for + [`~models.unets.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unets.unet_2d_blocks.CrossAttnUpBlock2D`], + [`~models.unets.unet_2d_blocks.UNetMidBlock2DCrossAttn`]. + encoder_hid_dim (`int`, *optional*, defaults to None): + If `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim` + dimension to `cross_attention_dim`. + encoder_hid_dim_type (`str`, *optional*, defaults to `None`): + If given, the `encoder_hidden_states` and potentially other embeddings are down-projected to text + embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`. + attention_head_dim (`int`, *optional*, defaults to 8): The dimension of the attention heads. + num_attention_heads (`int`, *optional*): + The number of attention heads. If not defined, defaults to `attention_head_dim` + resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config + for ResNet blocks (see [`~models.resnet.ResnetBlock2D`]). Choose from `default` or `scale_shift`. + class_embed_type (`str`, *optional*, defaults to `None`): + The type of class embedding to use which is ultimately summed with the time embeddings. Choose from `None`, + `"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`. + addition_embed_type (`str`, *optional*, defaults to `None`): + Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or + "text". "text" will use the `TextTimeEmbedding` layer. + addition_time_embed_dim: (`int`, *optional*, defaults to `None`): + Dimension for the timestep embeddings. + num_class_embeds (`int`, *optional*, defaults to `None`): + Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing + class conditioning with `class_embed_type` equal to `None`. + time_embedding_type (`str`, *optional*, defaults to `positional`): + The type of position embedding to use for timesteps. Choose from `positional` or `fourier`. + time_embedding_dim (`int`, *optional*, defaults to `None`): + An optional override for the dimension of the projected time embedding. + time_embedding_act_fn (`str`, *optional*, defaults to `None`): + Optional activation function to use only once on the time embeddings before they are passed to the rest of + the UNet. Choose from `silu`, `mish`, `gelu`, and `swish`. + timestep_post_act (`str`, *optional*, defaults to `None`): + The second activation function to use in timestep embedding. Choose from `silu`, `mish` and `gelu`. + time_cond_proj_dim (`int`, *optional*, defaults to `None`): + The dimension of `cond_proj` layer in the timestep embedding. + conv_in_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_in` layer. + conv_out_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_out` layer. + projection_class_embeddings_input_dim (`int`, *optional*): The dimension of the `class_labels` input when + `class_embed_type="projection"`. Required when `class_embed_type="projection"`. + class_embeddings_concat (`bool`, *optional*, defaults to `False`): Whether to concatenate the time + embeddings with the class embeddings. + mid_block_only_cross_attention (`bool`, *optional*, defaults to `None`): + Whether to use cross attention with the mid block when using the `UNetMidBlock2DSimpleCrossAttn`. If + `only_cross_attention` is given as a single boolean and `mid_block_only_cross_attention` is `None`, the + `only_cross_attention` value is used as the value for `mid_block_only_cross_attention`. Default to `False` + otherwise. + """ + + _supports_gradient_checkpointing = True + _no_split_modules = ["BasicTransformerBlock", "ResnetBlock2D", "CrossAttnUpBlock2D"] + + @register_to_config + def __init__( + self, + sample_size: Optional[int] = None, + in_channels: int = 4, + out_channels: int = 4, + center_input_sample: bool = False, + flip_sin_to_cos: bool = True, + freq_shift: int = 0, + down_block_types: Tuple[str] = ( + "CrossAttnDownBlock2D", + "CrossAttnDownBlock2D", + "CrossAttnDownBlock2D", + "DownBlock2D", + ), + mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn", + up_block_types: Tuple[str] = ("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D"), + only_cross_attention: Union[bool, Tuple[bool]] = False, + block_out_channels: Tuple[int] = (320, 640, 1280, 1280), + layers_per_block: Union[int, Tuple[int]] = 2, + downsample_padding: int = 1, + mid_block_scale_factor: float = 1, + dropout: float = 0.0, + act_fn: str = "silu", + norm_num_groups: Optional[int] = 32, + norm_eps: float = 1e-5, + cross_attention_dim: Union[int, Tuple[int]] = 1280, + transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple]] = 1, + reverse_transformer_layers_per_block: Optional[Tuple[Tuple[int]]] = None, + encoder_hid_dim: Optional[int] = None, + encoder_hid_dim_type: Optional[str] = None, + attention_head_dim: Union[int, Tuple[int]] = 8, + num_attention_heads: Optional[Union[int, Tuple[int]]] = None, + dual_cross_attention: bool = False, + use_linear_projection: bool = False, + class_embed_type: Optional[str] = None, + addition_embed_type: Optional[str] = None, + addition_time_embed_dim: Optional[int] = None, + num_class_embeds: Optional[int] = None, + upcast_attention: bool = False, + resnet_time_scale_shift: str = "default", + resnet_skip_time_act: bool = False, + resnet_out_scale_factor: float = 1.0, + time_embedding_type: str = "positional", + time_embedding_dim: Optional[int] = None, + time_embedding_act_fn: Optional[str] = None, + timestep_post_act: Optional[str] = None, + time_cond_proj_dim: Optional[int] = None, + conv_in_kernel: int = 3, + conv_out_kernel: int = 3, + projection_class_embeddings_input_dim: Optional[int] = None, + attention_type: str = "default", + class_embeddings_concat: bool = False, + mid_block_only_cross_attention: Optional[bool] = None, + cross_attention_norm: Optional[str] = None, + addition_embed_type_num_heads: int = 64, + cache_method: str = None, + ): + super().__init__() + + self.sample_size = sample_size + + if num_attention_heads is not None: + raise ValueError( + "At the moment it is not possible to define the number of attention heads via `num_attention_heads` because of a naming issue as described in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131. Passing `num_attention_heads` will only be supported in diffusers v0.19." + ) + + # If `num_attention_heads` is not defined (which is the case for most models) + # it will default to `attention_head_dim`. This looks weird upon first reading it and it is. + # The reason for this behavior is to correct for incorrectly named variables that were introduced + # when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131 + # Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking + # which is why we correct for the naming here. + num_attention_heads = num_attention_heads or attention_head_dim + + # Check inputs + self._check_config( + down_block_types=down_block_types, + up_block_types=up_block_types, + only_cross_attention=only_cross_attention, + block_out_channels=block_out_channels, + layers_per_block=layers_per_block, + cross_attention_dim=cross_attention_dim, + transformer_layers_per_block=transformer_layers_per_block, + reverse_transformer_layers_per_block=reverse_transformer_layers_per_block, + attention_head_dim=attention_head_dim, + num_attention_heads=num_attention_heads, + ) + + # input + conv_in_padding = (conv_in_kernel - 1) // 2 + self.conv_in = nn.Conv2d( + in_channels, block_out_channels[0], kernel_size=conv_in_kernel, padding=conv_in_padding + ) + + # time + time_embed_dim, timestep_input_dim = self._set_time_proj( + time_embedding_type, + block_out_channels=block_out_channels, + flip_sin_to_cos=flip_sin_to_cos, + freq_shift=freq_shift, + time_embedding_dim=time_embedding_dim, + ) + + self.time_embedding = TimestepEmbedding( + timestep_input_dim, + time_embed_dim, + act_fn=act_fn, + post_act_fn=timestep_post_act, + cond_proj_dim=time_cond_proj_dim, + ) + + self._set_encoder_hid_proj( + encoder_hid_dim_type, + cross_attention_dim=cross_attention_dim, + encoder_hid_dim=encoder_hid_dim, + ) + + # class embedding + self._set_class_embedding( + class_embed_type, + act_fn=act_fn, + num_class_embeds=num_class_embeds, + projection_class_embeddings_input_dim=projection_class_embeddings_input_dim, + time_embed_dim=time_embed_dim, + timestep_input_dim=timestep_input_dim, + ) + + self._set_add_embedding( + addition_embed_type, + addition_embed_type_num_heads=addition_embed_type_num_heads, + addition_time_embed_dim=addition_time_embed_dim, + cross_attention_dim=cross_attention_dim, + encoder_hid_dim=encoder_hid_dim, + flip_sin_to_cos=flip_sin_to_cos, + freq_shift=freq_shift, + projection_class_embeddings_input_dim=projection_class_embeddings_input_dim, + time_embed_dim=time_embed_dim, + ) + + if time_embedding_act_fn is None: + self.time_embed_act = None + else: + self.time_embed_act = get_activation(time_embedding_act_fn) + + self.down_blocks = nn.ModuleList([]) + self.up_blocks = nn.ModuleList([]) + + if isinstance(only_cross_attention, bool): + if mid_block_only_cross_attention is None: + mid_block_only_cross_attention = only_cross_attention + + only_cross_attention = [only_cross_attention] * len(down_block_types) + + if mid_block_only_cross_attention is None: + mid_block_only_cross_attention = False + + if isinstance(num_attention_heads, int): + num_attention_heads = (num_attention_heads,) * len(down_block_types) + + if isinstance(attention_head_dim, int): + attention_head_dim = (attention_head_dim,) * len(down_block_types) + + if isinstance(cross_attention_dim, int): + cross_attention_dim = (cross_attention_dim,) * len(down_block_types) + + if isinstance(layers_per_block, int): + layers_per_block = [layers_per_block] * len(down_block_types) + + if isinstance(transformer_layers_per_block, int): + transformer_layers_per_block = [transformer_layers_per_block] * len(down_block_types) + + if class_embeddings_concat: + # The time embeddings are concatenated with the class embeddings. The dimension of the + # time embeddings passed to the down, middle, and up blocks is twice the dimension of the + # regular time embeddings + blocks_time_embed_dim = time_embed_dim * 2 + else: + blocks_time_embed_dim = time_embed_dim + + # down + output_channel = block_out_channels[0] + for i, down_block_type in enumerate(down_block_types): + input_channel = output_channel + output_channel = block_out_channels[i] + is_final_block = i == len(block_out_channels) - 1 + + down_block = get_down_block( + down_block_type, + num_layers=layers_per_block[i], + transformer_layers_per_block=transformer_layers_per_block[i], + in_channels=input_channel, + out_channels=output_channel, + temb_channels=blocks_time_embed_dim, + add_downsample=not is_final_block, + resnet_eps=norm_eps, + resnet_act_fn=act_fn, + resnet_groups=norm_num_groups, + cross_attention_dim=cross_attention_dim[i], + num_attention_heads=num_attention_heads[i], + downsample_padding=downsample_padding, + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention[i], + upcast_attention=upcast_attention, + resnet_time_scale_shift=resnet_time_scale_shift, + attention_type=attention_type, + resnet_skip_time_act=resnet_skip_time_act, + resnet_out_scale_factor=resnet_out_scale_factor, + cross_attention_norm=cross_attention_norm, + attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel, + dropout=dropout, + ) + self.down_blocks.append(down_block) + + # mid + self.mid_block = get_mid_block( + mid_block_type, + temb_channels=blocks_time_embed_dim, + in_channels=block_out_channels[-1], + resnet_eps=norm_eps, + resnet_act_fn=act_fn, + resnet_groups=norm_num_groups, + output_scale_factor=mid_block_scale_factor, + transformer_layers_per_block=transformer_layers_per_block[-1], + num_attention_heads=num_attention_heads[-1], + cross_attention_dim=cross_attention_dim[-1], + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, + mid_block_only_cross_attention=mid_block_only_cross_attention, + upcast_attention=upcast_attention, + resnet_time_scale_shift=resnet_time_scale_shift, + attention_type=attention_type, + resnet_skip_time_act=resnet_skip_time_act, + cross_attention_norm=cross_attention_norm, + attention_head_dim=attention_head_dim[-1], + dropout=dropout, + ) + + # count how many layers upsample the images + self.num_upsamplers = 0 + + # up + reversed_block_out_channels = list(reversed(block_out_channels)) + reversed_num_attention_heads = list(reversed(num_attention_heads)) + reversed_layers_per_block = list(reversed(layers_per_block)) + reversed_cross_attention_dim = list(reversed(cross_attention_dim)) + reversed_transformer_layers_per_block = ( + list(reversed(transformer_layers_per_block)) + if reverse_transformer_layers_per_block is None + else reverse_transformer_layers_per_block + ) + only_cross_attention = list(reversed(only_cross_attention)) + + output_channel = reversed_block_out_channels[0] + for i, up_block_type in enumerate(up_block_types): + is_final_block = i == len(block_out_channels) - 1 + + prev_output_channel = output_channel + output_channel = reversed_block_out_channels[i] + input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)] + + # add upsample block for all BUT final layer + if not is_final_block: + add_upsample = True + self.num_upsamplers += 1 + else: + add_upsample = False + + up_block = get_up_block( + up_block_type, + num_layers=reversed_layers_per_block[i] + 1, + transformer_layers_per_block=reversed_transformer_layers_per_block[i], + in_channels=input_channel, + out_channels=output_channel, + prev_output_channel=prev_output_channel, + temb_channels=blocks_time_embed_dim, + add_upsample=add_upsample, + resnet_eps=norm_eps, + resnet_act_fn=act_fn, + resolution_idx=i, + resnet_groups=norm_num_groups, + cross_attention_dim=reversed_cross_attention_dim[i], + num_attention_heads=reversed_num_attention_heads[i], + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention[i], + upcast_attention=upcast_attention, + resnet_time_scale_shift=resnet_time_scale_shift, + attention_type=attention_type, + resnet_skip_time_act=resnet_skip_time_act, + resnet_out_scale_factor=resnet_out_scale_factor, + cross_attention_norm=cross_attention_norm, + attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel, + dropout=dropout, + ) + self.up_blocks.append(up_block) + prev_output_channel = output_channel + + # out + if norm_num_groups is not None: + self.conv_norm_out = nn.GroupNorm( + num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=norm_eps + ) + + self.conv_act = get_activation(act_fn) + + else: + self.conv_norm_out = None + self.conv_act = None + + conv_out_padding = (conv_out_kernel - 1) // 2 + self.conv_out = nn.Conv2d( + block_out_channels[0], out_channels, kernel_size=conv_out_kernel, padding=conv_out_padding + ) + + self._set_pos_net_if_use_gligen(attention_type=attention_type, cross_attention_dim=cross_attention_dim) + + def enable_agb_cache(module): + for child in module.children(): + if hasattr(child, "enable_agb"): + child.enable_agb = True + if len(list(child.children())) > 0: + enable_agb_cache(child) + + if cache_method == "agb_cache": + enable_agb_cache(self.down_blocks) + enable_agb_cache(self.mid_block) + enable_agb_cache(self.up_blocks) + self.enable_unet_cache = cache_method == "static_cache" + self.cache = None + self.cache_step = [1, 2, 4, 6, 7, 9, 10, 12, 13, 14, 16, 18, 19, 21, 23, 24, 26, 27, 29, \ + 30, 31, 33, 34, 36, 37, 39, 40, 42, 43, 45, 47, 48, 49] + + def _check_config( + self, + down_block_types: Tuple[str], + up_block_types: Tuple[str], + only_cross_attention: Union[bool, Tuple[bool]], + block_out_channels: Tuple[int], + layers_per_block: Union[int, Tuple[int]], + cross_attention_dim: Union[int, Tuple[int]], + transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple[int]]], + reverse_transformer_layers_per_block: bool, + attention_head_dim: int, + num_attention_heads: Optional[Union[int, Tuple[int]]], + ): + if len(down_block_types) != len(up_block_types): + raise ValueError( + f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}." + ) + + if len(block_out_channels) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}." + ) + + if not isinstance(only_cross_attention, bool) and len(only_cross_attention) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `only_cross_attention` as `down_block_types`. `only_cross_attention`: {only_cross_attention}. `down_block_types`: {down_block_types}." + ) + + if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}." + ) + + if not isinstance(attention_head_dim, int) and len(attention_head_dim) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `attention_head_dim` as `down_block_types`. `attention_head_dim`: {attention_head_dim}. `down_block_types`: {down_block_types}." + ) + + if isinstance(cross_attention_dim, list) and len(cross_attention_dim) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `cross_attention_dim` as `down_block_types`. `cross_attention_dim`: {cross_attention_dim}. `down_block_types`: {down_block_types}." + ) + + if not isinstance(layers_per_block, int) and len(layers_per_block) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `layers_per_block` as `down_block_types`. `layers_per_block`: {layers_per_block}. `down_block_types`: {down_block_types}." + ) + if isinstance(transformer_layers_per_block, list) and reverse_transformer_layers_per_block is None: + for layer_number_per_block in transformer_layers_per_block: + if isinstance(layer_number_per_block, list): + raise ValueError("Must provide 'reverse_transformer_layers_per_block` if using asymmetrical UNet.") + + def _set_time_proj( + self, + time_embedding_type: str, + block_out_channels: int, + flip_sin_to_cos: bool, + freq_shift: float, + time_embedding_dim: int, + ) -> Tuple[int, int]: + if time_embedding_type == "fourier": + time_embed_dim = time_embedding_dim or block_out_channels[0] * 2 + if time_embed_dim % 2 != 0: + raise ValueError(f"`time_embed_dim` should be divisible by 2, but is {time_embed_dim}.") + self.time_proj = GaussianFourierProjection( + time_embed_dim // 2, set_W_to_weight=False, log=False, flip_sin_to_cos=flip_sin_to_cos + ) + timestep_input_dim = time_embed_dim + elif time_embedding_type == "positional": + time_embed_dim = time_embedding_dim or block_out_channels[0] * 4 + + self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift) + timestep_input_dim = block_out_channels[0] + else: + raise ValueError( + f"{time_embedding_type} does not exist. Please make sure to use one of `fourier` or `positional`." + ) + + return time_embed_dim, timestep_input_dim + + def _set_encoder_hid_proj( + self, + encoder_hid_dim_type: Optional[str], + cross_attention_dim: Union[int, Tuple[int]], + encoder_hid_dim: Optional[int], + ): + if encoder_hid_dim_type is None and encoder_hid_dim is not None: + encoder_hid_dim_type = "text_proj" + self.register_to_config(encoder_hid_dim_type=encoder_hid_dim_type) + logger.info("encoder_hid_dim_type defaults to 'text_proj' as `encoder_hid_dim` is defined.") + + if encoder_hid_dim is None and encoder_hid_dim_type is not None: + raise ValueError( + f"`encoder_hid_dim` has to be defined when `encoder_hid_dim_type` is set to {encoder_hid_dim_type}." + ) + + if encoder_hid_dim_type == "text_proj": + self.encoder_hid_proj = nn.Linear(encoder_hid_dim, cross_attention_dim) + elif encoder_hid_dim_type == "text_image_proj": + # image_embed_dim DOESN'T have to be `cross_attention_dim`. To not clutter the __init__ too much + # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use + # case when `addition_embed_type == "text_image_proj"` (Kandinsky 2.1)` + self.encoder_hid_proj = TextImageProjection( + text_embed_dim=encoder_hid_dim, + image_embed_dim=cross_attention_dim, + cross_attention_dim=cross_attention_dim, + ) + elif encoder_hid_dim_type == "image_proj": + # Kandinsky 2.2 + self.encoder_hid_proj = ImageProjection( + image_embed_dim=encoder_hid_dim, + cross_attention_dim=cross_attention_dim, + ) + elif encoder_hid_dim_type is not None: + raise ValueError( + f"encoder_hid_dim_type: {encoder_hid_dim_type} must be None, 'text_proj' or 'text_image_proj'." + ) + else: + self.encoder_hid_proj = None + + def _set_class_embedding( + self, + class_embed_type: Optional[str], + act_fn: str, + num_class_embeds: Optional[int], + projection_class_embeddings_input_dim: Optional[int], + time_embed_dim: int, + timestep_input_dim: int, + ): + if class_embed_type is None and num_class_embeds is not None: + self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim) + elif class_embed_type == "timestep": + self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim, act_fn=act_fn) + elif class_embed_type == "identity": + self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim) + elif class_embed_type == "projection": + if projection_class_embeddings_input_dim is None: + raise ValueError( + "`class_embed_type`: 'projection' requires `projection_class_embeddings_input_dim` be set" + ) + # The projection `class_embed_type` is the same as the timestep `class_embed_type` except + # 1. the `class_labels` inputs are not first converted to sinusoidal embeddings + # 2. it projects from an arbitrary input dimension. + # + # Note that `TimestepEmbedding` is quite general, being mainly linear layers and activations. + # When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings. + # As a result, `TimestepEmbedding` can be passed arbitrary vectors. + self.class_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim) + elif class_embed_type == "simple_projection": + if projection_class_embeddings_input_dim is None: + raise ValueError( + "`class_embed_type`: 'simple_projection' requires `projection_class_embeddings_input_dim` be set" + ) + self.class_embedding = nn.Linear(projection_class_embeddings_input_dim, time_embed_dim) + else: + self.class_embedding = None + + def _set_add_embedding( + self, + addition_embed_type: str, + addition_embed_type_num_heads: int, + addition_time_embed_dim: Optional[int], + flip_sin_to_cos: bool, + freq_shift: float, + cross_attention_dim: Optional[int], + encoder_hid_dim: Optional[int], + projection_class_embeddings_input_dim: Optional[int], + time_embed_dim: int, + ): + if addition_embed_type == "text": + if encoder_hid_dim is not None: + text_time_embedding_from_dim = encoder_hid_dim + else: + text_time_embedding_from_dim = cross_attention_dim + + self.add_embedding = TextTimeEmbedding( + text_time_embedding_from_dim, time_embed_dim, num_heads=addition_embed_type_num_heads + ) + elif addition_embed_type == "text_image": + # text_embed_dim and image_embed_dim DON'T have to be `cross_attention_dim`. To not clutter the __init__ too much + # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use + # case when `addition_embed_type == "text_image"` (Kandinsky 2.1)` + self.add_embedding = TextImageTimeEmbedding( + text_embed_dim=cross_attention_dim, image_embed_dim=cross_attention_dim, time_embed_dim=time_embed_dim + ) + elif addition_embed_type == "text_time": + self.add_time_proj = Timesteps(addition_time_embed_dim, flip_sin_to_cos, freq_shift) + self.add_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim) + elif addition_embed_type == "image": + # Kandinsky 2.2 + self.add_embedding = ImageTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim) + elif addition_embed_type == "image_hint": + # Kandinsky 2.2 ControlNet + self.add_embedding = ImageHintTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim) + elif addition_embed_type is not None: + raise ValueError(f"addition_embed_type: {addition_embed_type} must be None, 'text' or 'text_image'.") + + def _set_pos_net_if_use_gligen(self, attention_type: str, cross_attention_dim: int): + if attention_type in ["gated", "gated-text-image"]: + positive_len = 768 + if isinstance(cross_attention_dim, int): + positive_len = cross_attention_dim + elif isinstance(cross_attention_dim, (list, tuple)): + positive_len = cross_attention_dim[0] + + feature_type = "text-only" if attention_type == "gated" else "text-image" + self.position_net = GLIGENTextBoundingboxProjection( + positive_len=positive_len, out_dim=cross_attention_dim, feature_type=feature_type + ) + + def get_time_embed( + self, sample: torch.Tensor, timestep: Union[torch.Tensor, float, int] + ) -> Optional[torch.Tensor]: + timesteps = timestep + if not torch.is_tensor(timesteps): + is_mps = sample.device.type == "mps" + if isinstance(timestep, float): + dtype = torch.float32 if is_mps else torch.float64 + else: + dtype = torch.int32 if is_mps else torch.int64 + timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device) + elif len(timesteps.shape) == 0: + timesteps = timesteps[None].to(sample.device) + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timesteps = timesteps.expand(sample.shape[0]) + + t_emb = self.time_proj(timesteps) + # `Timesteps` does not contain any weights and will always return f32 tensors + # but time_embedding might actually be running in fp16. so we need to cast here. + # there might be better ways to encapsulate this. + t_emb = t_emb.to(dtype=sample.dtype) + return t_emb + + def get_class_embed(self, sample: torch.Tensor, class_labels: Optional[torch.Tensor]) -> Optional[torch.Tensor]: + class_emb = None + if self.class_embedding is not None: + if class_labels is None: + raise ValueError("class_labels should be provided when num_class_embeds > 0") + + if self.config.class_embed_type == "timestep": + class_labels = self.time_proj(class_labels) + + # `Timesteps` does not contain any weights and will always return f32 tensors + # there might be better ways to encapsulate this. + class_labels = class_labels.to(dtype=sample.dtype) + + class_emb = self.class_embedding(class_labels).to(dtype=sample.dtype) + return class_emb + + def get_aug_embed( + self, emb: torch.Tensor, encoder_hidden_states: torch.Tensor, added_cond_kwargs: Dict[str, Any] + ) -> Optional[torch.Tensor]: + aug_emb = None + if self.config.addition_embed_type == "text": + aug_emb = self.add_embedding(encoder_hidden_states) + elif self.config.addition_embed_type == "text_image": + # Kandinsky 2.1 - style + if "image_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `addition_embed_type` set to 'text_image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`" + ) + + image_embs = added_cond_kwargs.get("image_embeds") + text_embs = added_cond_kwargs.get("text_embeds", encoder_hidden_states) + aug_emb = self.add_embedding(text_embs, image_embs) + elif self.config.addition_embed_type == "text_time": + # SDXL - style + if "text_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`" + ) + text_embeds = added_cond_kwargs.get("text_embeds") + if "time_ids" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`" + ) + time_ids = added_cond_kwargs.get("time_ids") + time_embeds = self.add_time_proj(time_ids.flatten()) + time_embeds = time_embeds.reshape((text_embeds.shape[0], -1)) + add_embeds = torch.concat([text_embeds, time_embeds], dim=-1) + add_embeds = add_embeds.to(emb.dtype) + aug_emb = self.add_embedding(add_embeds) + elif self.config.addition_embed_type == "image": + # Kandinsky 2.2 - style + if "image_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `addition_embed_type` set to 'image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`" + ) + image_embs = added_cond_kwargs.get("image_embeds") + aug_emb = self.add_embedding(image_embs) + elif self.config.addition_embed_type == "image_hint": + # Kandinsky 2.2 - style + if "image_embeds" not in added_cond_kwargs or "hint" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `addition_embed_type` set to 'image_hint' which requires the keyword arguments `image_embeds` and `hint` to be passed in `added_cond_kwargs`" + ) + image_embs = added_cond_kwargs.get("image_embeds") + hint = added_cond_kwargs.get("hint") + aug_emb = self.add_embedding(image_embs, hint) + return aug_emb + + def process_encoder_hidden_states( + self, encoder_hidden_states: torch.Tensor, added_cond_kwargs: Dict[str, Any] + ) -> torch.Tensor: + if self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_proj": + encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states) + elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_image_proj": + # Kandinsky 2.1 - style + if "image_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'text_image_proj' which requires the keyword argument `image_embeds` to be passed in `added_conditions`" + ) + + image_embeds = added_cond_kwargs.get("image_embeds") + encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states, image_embeds) + elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "image_proj": + # Kandinsky 2.2 - style + if "image_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'image_proj' which requires the keyword argument `image_embeds` to be passed in `added_conditions`" + ) + image_embeds = added_cond_kwargs.get("image_embeds") + encoder_hidden_states = self.encoder_hid_proj(image_embeds) + elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "ip_image_proj": + if "image_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'ip_image_proj' which requires the keyword argument `image_embeds` to be passed in `added_conditions`" + ) + image_embeds = added_cond_kwargs.get("image_embeds") + image_embeds = self.encoder_hid_proj(image_embeds) + encoder_hidden_states = (encoder_hidden_states, image_embeds) + return encoder_hidden_states + + def forward( + self, + sample: torch.Tensor, + timestep: Union[torch.Tensor, float, int], + encoder_hidden_states: torch.Tensor, + class_labels: Optional[torch.Tensor] = None, + timestep_cond: Optional[torch.Tensor] = None, + attention_mask: Optional[torch.Tensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None, + down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None, + mid_block_additional_residual: Optional[torch.Tensor] = None, + down_intrablock_additional_residuals: Optional[Tuple[torch.Tensor]] = None, + encoder_attention_mask: Optional[torch.Tensor] = None, + return_dict: bool = True, + **kwargs + ) -> Union[UNet2DConditionOutput, Tuple]: + r""" + The [`UNet2DConditionModel`] forward method. + + Args: + sample (`torch.Tensor`): + The noisy input tensor with the following shape `(batch, channel, height, width)`. + timestep (`torch.Tensor` or `float` or `int`): The number of timesteps to denoise an input. + encoder_hidden_states (`torch.Tensor`): + The encoder hidden states with shape `(batch, sequence_length, feature_dim)`. + class_labels (`torch.Tensor`, *optional*, defaults to `None`): + Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings. + timestep_cond: (`torch.Tensor`, *optional*, defaults to `None`): + Conditional embeddings for timestep. If provided, the embeddings will be summed with the samples passed + through the `self.time_embedding` layer to obtain the timestep embeddings. + attention_mask (`torch.Tensor`, *optional*, defaults to `None`): + An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask + is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large + negative values to the attention scores corresponding to "discard" tokens. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + added_cond_kwargs: (`dict`, *optional*): + A kwargs dictionary containing additional embeddings that if specified are added to the embeddings that + are passed along to the UNet blocks. + down_block_additional_residuals: (`tuple` of `torch.Tensor`, *optional*): + A tuple of tensors that if specified are added to the residuals of down unet blocks. + mid_block_additional_residual: (`torch.Tensor`, *optional*): + A tensor that if specified is added to the residual of the middle unet block. + down_intrablock_additional_residuals (`tuple` of `torch.Tensor`, *optional*): + additional residuals to be added within UNet down blocks, for example from T2I-Adapter side model(s) + encoder_attention_mask (`torch.Tensor`): + A cross-attention mask of shape `(batch, sequence_length)` is applied to `encoder_hidden_states`. If + `True` the mask is kept, otherwise if `False` it is discarded. Mask will be converted into a bias, + which adds large negative values to the attention scores corresponding to "discard" tokens. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] instead of a plain + tuple. + + Returns: + [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] or `tuple`: + If `return_dict` is True, an [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] is returned, + otherwise a `tuple` is returned where the first element is the sample tensor. + """ + # By default samples have to be AT least a multiple of the overall upsampling factor. + # The overall upsampling factor is equal to 2 ** (# num of upsampling layers). + # However, the upsampling interpolation output size can be forced to fit any upsampling size + # on the fly if necessary. + default_overall_up_factor = 2**self.num_upsamplers + + # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor` + forward_upsample_size = False + upsample_size = None + + for dim in sample.shape[-2:]: + if dim % default_overall_up_factor != 0: + # Forward upsample size to force interpolation output size. + forward_upsample_size = True + break + if attention_mask is not None: + attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0 + attention_mask = attention_mask.unsqueeze(1) + + # convert encoder_attention_mask to a bias the same way we do for attention_mask + if encoder_attention_mask is not None: + encoder_attention_mask = (1 - encoder_attention_mask.to(sample.dtype)) * -10000.0 + encoder_attention_mask = encoder_attention_mask.unsqueeze(1) + + # 0. center input if necessary + if self.config.center_input_sample: + sample = 2 * sample - 1.0 + + # 1. time + t_emb = self.get_time_embed(sample=sample, timestep=timestep) + emb = self.time_embedding(t_emb, timestep_cond) + aug_emb = None + + class_emb = self.get_class_embed(sample=sample, class_labels=class_labels) + if class_emb is not None: + if self.config.class_embeddings_concat: + emb = torch.cat([emb, class_emb], dim=-1) + else: + emb = emb + class_emb + + aug_emb = self.get_aug_embed( + emb=emb, encoder_hidden_states=encoder_hidden_states, added_cond_kwargs=added_cond_kwargs + ) + if self.config.addition_embed_type == "image_hint": + aug_emb, hint = aug_emb + sample = torch.cat([sample, hint], dim=1) + + emb = emb + aug_emb if aug_emb is not None else emb + + if self.time_embed_act is not None: + emb = self.time_embed_act(emb) + + encoder_hidden_states = self.process_encoder_hidden_states( + encoder_hidden_states=encoder_hidden_states, added_cond_kwargs=added_cond_kwargs + ) + + # 2. pre-process + sample = self.conv_in(sample) + + # 2.5 GLIGEN position net + if cross_attention_kwargs is not None and cross_attention_kwargs.get("gligen", None) is not None: + cross_attention_kwargs = cross_attention_kwargs.copy() + gligen_args = cross_attention_kwargs.pop("gligen") + cross_attention_kwargs["gligen"] = {"objs": self.position_net(**gligen_args)} + + # 3. down + # we're popping the `scale` instead of getting it because otherwise `scale` will be propagated + # to the internal blocks and will raise deprecation warnings. this will be confusing for our users. + if cross_attention_kwargs is not None: + cross_attention_kwargs = cross_attention_kwargs.copy() + lora_scale = cross_attention_kwargs.pop("scale", 1.0) + else: + lora_scale = 1.0 + + if USE_PEFT_BACKEND: + # weight the lora layers by setting `lora_scale` for each PEFT layer + scale_lora_layers(self, lora_scale) + + is_controlnet = mid_block_additional_residual is not None and down_block_additional_residuals is not None + # using new arg down_intrablock_additional_residuals for T2I-Adapters, to distinguish from controlnets + is_adapter = down_intrablock_additional_residuals is not None + # maintain backward compatibility for legacy usage, where + # T2I-Adapter and ControlNet both use down_block_additional_residuals arg + # but can only use one or the other + if not is_adapter and mid_block_additional_residual is None and down_block_additional_residuals is not None: + deprecate( + "T2I should not use down_block_additional_residuals", + "1.3.0", + "Passing intrablock residual connections with `down_block_additional_residuals` is deprecated \ + and will be removed in diffusers 1.3.0. `down_block_additional_residuals` should only be used \ + for ControlNet. Please make sure use `down_intrablock_additional_residuals` instead. ", + standard_warn=False, + ) + down_intrablock_additional_residuals = down_block_additional_residuals + is_adapter = True + + down_block_res_samples = (sample,) + if self.enable_unet_cache: + step = kwargs.get("step", 0) + if len(self.cache_step) > 0 and (step + 1) not in self.cache_step: + for block_id, downsample_block in enumerate(self.down_blocks): + if block_id >= 2: # skip last block + break + if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention: + # For t2i-adapter CrossAttnDownBlock2D + additional_residuals = {} + if is_adapter and len(down_intrablock_additional_residuals) > 0: + additional_residuals["additional_residuals"] = down_intrablock_additional_residuals.pop(0) + + sample, res_samples = downsample_block( + hidden_states=sample, + temb=emb, + encoder_hidden_states=encoder_hidden_states, + attention_mask=attention_mask, + cross_attention_kwargs=cross_attention_kwargs, + encoder_attention_mask=encoder_attention_mask, + **additional_residuals, + ) + else: + sample, res_samples = downsample_block(hidden_states=sample, temb=emb) + if is_adapter and len(down_intrablock_additional_residuals) > 0: + sample += down_intrablock_additional_residuals.pop(0) + + down_block_res_samples += res_samples + + if is_controlnet: + new_down_block_res_samples = () + + for down_block_res_sample, down_block_additional_residual in zip( + down_block_res_samples, down_block_additional_residuals + ): + down_block_res_sample = down_block_res_sample + down_block_additional_residual + new_down_block_res_samples = new_down_block_res_samples + (down_block_res_sample,) + + down_block_res_samples = new_down_block_res_samples + sample = self.cache.detach() + # 5. up + for i, upsample_block in enumerate(self.up_blocks): + is_final_block = i == len(self.up_blocks) - 1 + if i == 1: + res_samples = down_block_res_samples[-4:-1] + if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention: + sample = upsample_block( + hidden_states=sample, + temb=emb, + res_hidden_states_tuple=res_samples, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + upsample_size=upsample_size, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + ) + if i == 2: + res_samples = down_block_res_samples[:3] + sample = upsample_block( + hidden_states=sample, + temb=emb, + res_hidden_states_tuple=res_samples, + upsample_size=upsample_size, + ) + else: + for downsample_block in self.down_blocks: + if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention: + # For t2i-adapter CrossAttnDownBlock2D + additional_residuals = {} + if is_adapter and len(down_intrablock_additional_residuals) > 0: + additional_residuals["additional_residuals"] = down_intrablock_additional_residuals.pop(0) + + sample, res_samples = downsample_block( + hidden_states=sample, + temb=emb, + encoder_hidden_states=encoder_hidden_states, + attention_mask=attention_mask, + cross_attention_kwargs=cross_attention_kwargs, + encoder_attention_mask=encoder_attention_mask, + **additional_residuals, + ) + else: + sample, res_samples = downsample_block(hidden_states=sample, temb=emb) + if is_adapter and len(down_intrablock_additional_residuals) > 0: + sample += down_intrablock_additional_residuals.pop(0) + + down_block_res_samples += res_samples + + if is_controlnet: + new_down_block_res_samples = () + + for down_block_res_sample, down_block_additional_residual in zip( + down_block_res_samples, down_block_additional_residuals + ): + down_block_res_sample = down_block_res_sample + down_block_additional_residual + new_down_block_res_samples = new_down_block_res_samples + (down_block_res_sample,) + + down_block_res_samples = new_down_block_res_samples + + # 4. mid + if self.mid_block is not None: + if hasattr(self.mid_block, "has_cross_attention") and self.mid_block.has_cross_attention: + sample = self.mid_block( + sample, + emb, + encoder_hidden_states=encoder_hidden_states, + attention_mask=attention_mask, + cross_attention_kwargs=cross_attention_kwargs, + encoder_attention_mask=encoder_attention_mask, + ) + else: + sample = self.mid_block(sample, emb) + + # To support T2I-Adapter-XL + if ( + is_adapter + and len(down_intrablock_additional_residuals) > 0 + and sample.shape == down_intrablock_additional_residuals[0].shape + ): + sample += down_intrablock_additional_residuals.pop(0) + + if is_controlnet: + sample = sample + mid_block_additional_residual + + # 5. up + for i, upsample_block in enumerate(self.up_blocks): + is_final_block = i == len(self.up_blocks) - 1 + + res_samples = down_block_res_samples[-len(upsample_block.resnets) :] + down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)] + + # if we have not reached the final block and need to forward the + # upsample size, we do it here + if not is_final_block and forward_upsample_size: + upsample_size = down_block_res_samples[-1].shape[2:] + + if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention: + sample = upsample_block( + hidden_states=sample, + temb=emb, + res_hidden_states_tuple=res_samples, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + upsample_size=upsample_size, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + ) + else: + sample = upsample_block( + hidden_states=sample, + temb=emb, + res_hidden_states_tuple=res_samples, + upsample_size=upsample_size, + ) + if i == 0: + self.cache = sample + else: + for downsample_block in self.down_blocks: + if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention: + # For t2i-adapter CrossAttnDownBlock2D + additional_residuals = {} + if is_adapter and len(down_intrablock_additional_residuals) > 0: + additional_residuals["additional_residuals"] = down_intrablock_additional_residuals.pop(0) + + sample, res_samples = downsample_block( + hidden_states=sample, + temb=emb, + encoder_hidden_states=encoder_hidden_states, + attention_mask=attention_mask, + cross_attention_kwargs=cross_attention_kwargs, + encoder_attention_mask=encoder_attention_mask, + **additional_residuals, + ) + else: + sample, res_samples = downsample_block(hidden_states=sample, temb=emb) + if is_adapter and len(down_intrablock_additional_residuals) > 0: + sample += down_intrablock_additional_residuals.pop(0) + + down_block_res_samples += res_samples + + if is_controlnet: + new_down_block_res_samples = () + + for down_block_res_sample, down_block_additional_residual in zip( + down_block_res_samples, down_block_additional_residuals + ): + down_block_res_sample = down_block_res_sample + down_block_additional_residual + new_down_block_res_samples = new_down_block_res_samples + (down_block_res_sample,) + + down_block_res_samples = new_down_block_res_samples + + # 4. mid + if self.mid_block is not None: + if hasattr(self.mid_block, "has_cross_attention") and self.mid_block.has_cross_attention: + sample = self.mid_block( + sample, + emb, + encoder_hidden_states=encoder_hidden_states, + attention_mask=attention_mask, + cross_attention_kwargs=cross_attention_kwargs, + encoder_attention_mask=encoder_attention_mask, + ) + else: + sample = self.mid_block(sample, emb) + + # To support T2I-Adapter-XL + if ( + is_adapter + and len(down_intrablock_additional_residuals) > 0 + and sample.shape == down_intrablock_additional_residuals[0].shape + ): + sample += down_intrablock_additional_residuals.pop(0) + + if is_controlnet: + sample = sample + mid_block_additional_residual + + # 5. up + for i, upsample_block in enumerate(self.up_blocks): + is_final_block = i == len(self.up_blocks) - 1 + + res_samples = down_block_res_samples[-len(upsample_block.resnets) :] + down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)] + + # if we have not reached the final block and need to forward the + # upsample size, we do it here + if not is_final_block and forward_upsample_size: + upsample_size = down_block_res_samples[-1].shape[2:] + + if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention: + sample = upsample_block( + hidden_states=sample, + temb=emb, + res_hidden_states_tuple=res_samples, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + upsample_size=upsample_size, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + ) + else: + sample = upsample_block( + hidden_states=sample, + temb=emb, + res_hidden_states_tuple=res_samples, + upsample_size=upsample_size, + ) + + # 6. post-process + if self.conv_norm_out: + sample = self.conv_norm_out(sample) + sample = self.conv_act(sample) + sample = self.conv_out(sample) + + if USE_PEFT_BACKEND: + # remove `lora_scale` from each PEFT layer + unscale_lora_layers(self, lora_scale) + + if not return_dict: + return (sample,) + + return UNet2DConditionOutput(sample=sample) \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-XL/test_performance.py b/MindIE/MultiModal/StableDiffusion-XL/test_performance.py new file mode 100644 index 0000000000000000000000000000000000000000..c6d9e23f234fb898eba4b3e612008315673f094d --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-XL/test_performance.py @@ -0,0 +1,107 @@ +# Copyright 2024 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import os +import argparse +import time +import logging + +import torch +import torch_npu + +from stable_diffusion_xl.pipeline_stable_diffusion_xl import StableDiffusionXLPipeline +from stable_diffusion_xl.unet.unet_model import UNet2DConditionModel + + +logging.basicConfig(level=logging.INFO) +logger = logging.getLogger() + + +def parse_arguments(): + parser = argparse.ArgumentParser() + parser.add_argument( + "--path", + type=str, + default='/stable-diffusion-xl-base-1.0', + help="The path of all model weights, suach as vae, unet, text_encoder, tokenizer, scheduler", + ) + parser.add_argument( + "--device_id", + type=int, + default=0, + help="NPU device id", + ) + parser.add_argument( + "--dtype", + type=torch.dtype, + default=torch.float16 + ) + parser.add_argument( + "--num_images_per_prompt", + type=int, + default=1 + ) + parser.add_argument( + "--height", + type=int, + default=1024 + ) + parser.add_argument( + "--width", + type=int, + default=1024 + ) + parser.add_argument( + "--cache_method", + type=str, + default="", + help="static_cache or agb_cache", + ) + parser.add_argument( + "--iterator", + type=int, + default=6, + help="The number of times the test was iterated.", + ) + return parser.parse_args() + + +def test_performance(): + args = parse_arguments() + + torch.npu.set_device(args.device_id) + + unet = UNet2DConditionModel.from_pretrained(os.path.join(args.path, 'unet'), cache_method=args.cache_method) + pipe = StableDiffusionXLPipeline.from_pretrained(args.path, unet=unet) + pipe.to(args.dtype).to("npu") + + prompts = "A beautiful girl with flowers." + + all_time = 0 + for i in range(args.iterator): + begin_time = time.time() + image = pipe( + prompt=prompts, + height=args.height, + width=args.width, + num_images_per_prompt=args.num_images_per_prompt + )[0] + if i > 2: # skip the first 3 infer. + end_time = time.time() + all_time += (end_time - begin_time) + logger.info(f"Time interval is {all_time / (args.iterator - 3)}") # skip the first 3 infer. + + +if __name__ == "__main__": + test_performance() \ No newline at end of file diff --git a/MindIE/MultiModal/StableDiffusion-XL/test_precision.py b/MindIE/MultiModal/StableDiffusion-XL/test_precision.py new file mode 100644 index 0000000000000000000000000000000000000000..db945c21ba470850e9ee6f943f04e466f42b1a46 --- /dev/null +++ b/MindIE/MultiModal/StableDiffusion-XL/test_precision.py @@ -0,0 +1,255 @@ +# Copyright 2024 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import os +import argparse +import time +import logging +import csv +import json + +from typing import Callable, List, Optional, Union +import numpy as np + +import torch +import torch_npu + +from stable_diffusion_xl.pipeline_stable_diffusion_xl import StableDiffusionXLPipeline +from stable_diffusion_xl.unet.unet_model import UNet2DConditionModel + + +logging.basicConfig(level=logging.INFO) +logger = logging.getLogger() + + +class PromptLoader: + def __init__( + self, + prompt_file: str, + prompt_file_type: str, + batch_size: int, + num_images_per_prompt: int = 1, + max_num_prompts: int = 0 + ): + self.prompts = [] + self.catagories = ['Not_specified'] + self.batch_size = batch_size + self.num_images_per_prompt = num_images_per_prompt + + if prompt_file_type == 'plain': + self.load_prompts_plain(prompt_file, max_num_prompts) + elif prompt_file_type == 'parti': + self.load_prompts_parti(prompt_file, max_num_prompts) + elif prompt_file_type == 'hpsv2': + self.load_prompts_hpsv2(max_num_prompts) + else: + print("This operation is not supported!") + + self.current_id = 0 + self.inner_id = 0 + + def __len__(self): + return len(self.prompts) * self.num_images_per_prompt + + def __iter__(self): + return self + + def __next__(self): + if self.current_id == len(self.prompts): + raise StopIteration + + ret = { + 'prompts': [], + 'catagories': [], + 'save_names': [], + 'n_prompts': self.batch_size, + } + for _ in range(self.batch_size): + if self.current_id == len(self.prompts): + ret['prompts'].append('') + ret['save_names'].append('') + ret['catagories'].append('') + ret['n_prompts'] -= 1 + + else: + prompt, catagory_id = self.prompts[self.current_id] + ret['prompts'].append(prompt) + ret['catagories'].append(self.catagories[catagory_id]) + ret['save_names'].append(f'{self.current_id}_{self.inner_id}') + + self.inner_id += 1 + if self.inner_id == self.num_images_per_prompt: + self.inner_id = 0 + self.current_id += 1 + + return ret + + def load_prompts_plain(self, file_path: str, max_num_prompts: int): + with os.fdopen(os.open(file_path, os.O_RDONLY), "r") as f: + for i, line in enumerate(f): + if max_num_prompts and i == max_num_prompts: + break + + prompt = line.strip() + self.prompts.append((prompt, 0)) + + def load_prompts_parti(self, file_path: str, max_num_prompts: int): + with os.fdopen(os.open(file_path, os.O_RDONLY), "r") as f: + # Skip the first line + next(f) + tsv_file = csv.reader(f, delimiter="\t") + for i, line in enumerate(tsv_file): + if max_num_prompts and i == max_num_prompts: + break + + prompt = line[0] + catagory = line[1] + if catagory not in self.catagories: + self.catagories.append(catagory) + + catagory_id = self.catagories.index(catagory) + self.prompts.append((prompt, catagory_id)) + + def load_prompts_hpsv2(self, max_num_prompts: int): + with open('hpsv2_benchmark_prompts.json', 'r') as file: + all_prompts = json.load(file) + count = 0 + for style, prompts in all_prompts.items(): + for prompt in prompts: + count += 1 + if max_num_prompts and count >= max_num_prompts: + break + + if style not in self.catagories: + self.catagories.append(style) + + catagory_id = self.catagories.index(style) + self.prompts.append((prompt, catagory_id)) + + +def parse_arguments(): + parser = argparse.ArgumentParser() + parser.add_argument( + "--path", + type=str, + default='/stable-diffusion-xl-base-1.0', + help="The path of all model weights, suach as vae, unet, text_encoder, tokenizer, scheduler", + ) + parser.add_argument( + "--device_id", + type=int, + default=0, + help="NPU device id", + ) + parser.add_argument( + "--dtype", + type=torch.dtype, + default=torch.float16 + ) + parser.add_argument( + "--height", + type=int, + default=1024 + ) + parser.add_argument( + "--width", + type=int, + default=1024 + ) + parser.add_argument( + "--num_images_per_prompt", + type=int, + default=1 + ) + parser.add_argument( + "--cache_method", + type=str, + default="", + help="static_cache or agb_cache", + ) + parser.add_argument( + "--prompt_file", + type=str, + default="./prompts.txt", + help="A text file of prompts for generating images.", + ) + parser.add_argument( + "--prompt_file_type", + choices=["plain", "parti", "hpsv2"], + default="plain", + help="Type of prompt file.", + ) + return parser.parse_args() + + +def test_precision(): + args = parse_arguments() + + torch.npu.set_device(args.device_id) + + unet = UNet2DConditionModel.from_pretrained(os.path.join(args.path, 'unet'), cache_method=args.cache_method) + pipe = StableDiffusionXLPipeline.from_pretrained(args.path, unet=unet) + pipe.to(args.dtype).to("npu") + + prompt_loader = PromptLoader( + args.prompt_file, + args.prompt_file_type, + batch_size=1, + num_images_per_prompt=args.num_images_per_prompt + ) + prompt_2 = "" + image_info = [] + current_prompt = None + infer_num = 0 + + current_time = time.strftime("%Y%m%d_%H%M%S", time.localtime()) + img_dir = f"{current_time}/images" + os.makedirs(img_dir) + + for i, input_info in enumerate(prompt_loader): + prompts = input_info['prompts'] + catagories = input_info['catagories'] + save_names = input_info['save_names'] + n_prompts = input_info['n_prompts'] + + logger.info(f"[{infer_num + n_prompts}/{len(prompt_loader)}]: {prompts}") + + infer_num += 1 + + images = pipe( + prompt=prompts, + prompt_2=prompt_2, + height=args.height, + width=args.width, + ) + + for j in range(n_prompts): + image_save_path = os.path.join(img_dir, f"{save_names[j]}.png") + image = images[0][j] + image.save(image_save_path) + + if current_prompt != prompts[j]: + current_prompt = prompts[j] + image_info.append({'images': [], 'prompt': current_prompt, 'category': catagories[j]}) + + image_info[-1]['images'].append(image_save_path) + + img_json = f"{current_time}/image_info.json" + + with os.fdopen(os.open(img_json, os.O_RDWR | os.O_CREAT, 0o640), "w") as f: + json.dump(image_info, f) + + +if __name__ == "__main__": + test_precision() \ No newline at end of file