# 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 Callable, Optional import torch import torch.nn.functional as F from torch import nn from diffusers.utils.import_utils import is_xformers_available # from diffusers.models.attention_processor import Attention # from t2v_enhanced.model.diffusers_conditional.models.controlnet.attention import Attention from t2v_enhanced.model.diffusers_conditional.models.controlnet.attention_processor import Attention from diffusers.models.embeddings import CombinedTimestepLabelEmbeddings # from t2v_enhanced.model.diffusers_conditional.models.controlnet.attention_processor import Attention if is_xformers_available(): import xformers import xformers.ops else: xformers = None 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. 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. 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. """ def __init__( self, dim: int, num_attention_heads: int, attention_head_dim: int, is_spatial_attention: bool = False, 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", final_dropout: bool = False, use_image_embedding: bool = False, unet_params=None, ): 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" 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}." ) # Define 3 blocks. Each block has its own normalization layer. # 1. Self-Attn if self.use_ada_layer_norm: self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm) elif self.use_ada_layer_norm_zero: self.norm1 = AdaLayerNormZero(dim, num_embeds_ada_norm) else: self.norm1 = nn.LayerNorm( dim, elementwise_affine=norm_elementwise_affine) 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, is_spatial_attention=is_spatial_attention, use_image_embedding=use_image_embedding, ) # 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. self.norm2 = ( AdaLayerNorm(dim, num_embeds_ada_norm) if self.use_ada_layer_norm else nn.LayerNorm(dim, elementwise_affine=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, is_spatial_attention=is_spatial_attention, use_image_embedding=use_image_embedding, unet_params=unet_params, ) # is self-attn if encoder_hidden_states is none else: self.norm2 = None self.attn2 = None # 3. Feed-forward self.norm3 = nn.LayerNorm( dim, elementwise_affine=norm_elementwise_affine) self.ff = FeedForward( dim, dropout=dropout, activation_fn=activation_fn, final_dropout=final_dropout) def forward( self, hidden_states, attention_mask=None, encoder_hidden_states=None, encoder_attention_mask=None, timestep=None, cross_attention_kwargs=None, class_labels=None, ): # Notice that normalization is always applied before the real computation in the following blocks. # 1. Self-Attention if self.use_ada_layer_norm: norm_hidden_states = self.norm1(hidden_states, timestep) elif self.use_ada_layer_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 ) else: norm_hidden_states = self.norm1(hidden_states) cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {} 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.use_ada_layer_norm_zero: attn_output = gate_msa.unsqueeze(1) * attn_output hidden_states = attn_output + hidden_states # 2. Cross-Attention if self.attn2 is not None: norm_hidden_states = ( self.norm2(hidden_states, timestep) if self.use_ada_layer_norm else self.norm2( hidden_states) ) # TODO (Birch-San): Here we should prepare the encoder_attention mask correctly # prepare attention mask here 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 # 3. Feed-forward norm_hidden_states = self.norm3(hidden_states) if self.use_ada_layer_norm_zero: norm_hidden_states = norm_hidden_states * \ (1 + scale_mlp[:, None]) + shift_mlp[:, None] ff_output = self.ff(norm_hidden_states) if self.use_ada_layer_norm_zero: ff_output = gate_mlp.unsqueeze(1) * ff_output hidden_states = ff_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. """ 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, ): super().__init__() 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) if activation_fn == "gelu-approximate": act_fn = GELU(dim, inner_dim, approximate="tanh") elif activation_fn == "geglu": act_fn = GEGLU(dim, inner_dim) elif activation_fn == "geglu-approximate": act_fn = ApproximateGELU(dim, inner_dim) 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)) # 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): for module in self.net: hidden_states = module(hidden_states) return hidden_states class GEGLU(nn.Module): r""" A variant of the gated linear unit activation function from https://arxiv.org/abs/2002.05202. Parameters: dim_in (`int`): The number of channels in the input. dim_out (`int`): The number of channels in the output. """ def __init__(self, dim_in: int, dim_out: int): super().__init__() self.proj = nn.Linear(dim_in, dim_out * 2) def gelu(self, gate): 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): hidden_states, gate = self.proj(hidden_states).chunk(2, dim=-1) return hidden_states * self.gelu(gate)