# Adapted from Open-Sora-Plan # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. # -------------------------------------------------------- # References: # Open-Sora-Plan: https://github.com/PKU-YuanGroup/Open-Sora-Plan # -------------------------------------------------------- from importlib import import_module from typing import Any, Callable, Dict, Optional, Tuple import numpy as np import torch import collections import torch.nn.functional as F from torch.nn.attention import SDPBackend, sdpa_kernel from diffusers.models.activations import GEGLU, GELU, ApproximateGELU from diffusers.models.attention_processor import ( AttnAddedKVProcessor, AttnAddedKVProcessor2_0, AttnProcessor, CustomDiffusionAttnProcessor, CustomDiffusionAttnProcessor2_0, CustomDiffusionXFormersAttnProcessor, LoRAAttnAddedKVProcessor, LoRAAttnProcessor, LoRAAttnProcessor2_0, LoRAXFormersAttnProcessor, SlicedAttnAddedKVProcessor, SlicedAttnProcessor, SpatialNorm, XFormersAttnAddedKVProcessor, XFormersAttnProcessor, ) from diffusers.models.embeddings import SinusoidalPositionalEmbedding from diffusers.models.normalization import AdaLayerNorm, AdaLayerNormZero from diffusers.utils import USE_PEFT_BACKEND, deprecate, is_xformers_available from diffusers.utils.torch_utils import maybe_allow_in_graph from torch import nn from allegro.models.transformers.rope import RoPE3D, PositionGetter3D from allegro.models.transformers.embedding import CombinedTimestepSizeEmbeddings if is_xformers_available(): import xformers import xformers.ops else: xformers = None from diffusers.utils import logging logger = logging.get_logger(__name__) def to_2tuple(x): if isinstance(x, collections.abc.Iterable): return x return (x, x) @maybe_allow_in_graph 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, attention_mode: str = "xformers", use_rope: bool = False, interpolation_scale_thw=None, ): super().__init__() self.inner_dim = dim_head * heads 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.use_rope = use_rope # 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 = 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'" ) linear_cls = nn.Linear 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, query_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( attention_mode, use_rope, interpolation_scale_thw=interpolation_scale_thw, ) 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") 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", _remove_lora: bool = False) -> None: r""" Set the attention processor to use. Args: processor (`AttnProcessor`): The attention processor to use. _remove_lora (`bool`, *optional*, defaults to `False`): Set to `True` to remove LoRA layers from the model. """ if not USE_PEFT_BACKEND and hasattr(self, "processor") and _remove_lora and self.to_q.lora_layer is not None: deprecate( "set_processor to offload LoRA", "0.26.0", "In detail, removing LoRA layers via calling `set_default_attn_processor` is deprecated. Please make sure to call `pipe.unload_lora_weights()` instead.", ) # TODO(Patrick, Sayak) - this can be deprecated once PEFT LoRA integration is complete # We need to remove all LoRA layers # Don't forget to remove ALL `_remove_lora` from the codebase for module in self.modules(): if hasattr(module, "set_lora_layer"): module.set_lora_layer(None) # 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): 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, head_size = None, ) -> 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 = head_size if head_size is not None else 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 def _init_compress(self): self.sr.bias.data.zero_() self.norm = nn.LayerNorm(self.inner_dim) class AttnProcessor2_0(nn.Module): r""" Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0). """ def __init__(self, attention_mode="xformers", use_rope=False, interpolation_scale_thw=None): super().__init__() self.attention_mode = attention_mode self.use_rope = use_rope self.interpolation_scale_thw = interpolation_scale_thw if self.use_rope: self._init_rope(interpolation_scale_thw) 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 _init_rope(self, interpolation_scale_thw): self.rope = RoPE3D(interpolation_scale_thw=interpolation_scale_thw) self.position_getter = PositionGetter3D() 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, frame: int = 8, height: int = 16, width: int = 16, ) -> 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 and self.attention_mode == 'xformers': attention_heads = attn.heads attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size, head_size=attention_heads) attention_mask = attention_mask.view(batch_size, attention_heads, -1, attention_mask.shape[-1]) else: 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) 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) attn_heads = attn.heads 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) if self.use_rope: # require the shape of (batch_size x nheads x ntokens x dim) pos_thw = self.position_getter(batch_size, t=frame, h=height, w=width, device=query.device) query = self.rope(query, pos_thw) key = self.rope(key, pos_thw) # the output of sdp = (batch, num_heads, seq_len, head_dim) # TODO: add support for attn.scale when we move to Torch 2.1 if self.attention_mode == 'flash': # assert attention_mask is None, 'flash-attn do not support attention_mask' with sdpa_kernel(SDPBackend.FLASH_ATTENTION): hidden_states = F.scaled_dot_product_attention( query, key, value, dropout_p=0.0, is_causal=False ) elif self.attention_mode == 'xformers': with sdpa_kernel(SDPBackend.EFFICIENT_ATTENTION): 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) # 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 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 linear_cls = nn.Linear 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(linear_cls(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: torch.Tensor) -> torch.Tensor: for module in self.net: hidden_states = module(hidden_states) return hidden_states @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. 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' norm_eps: float = 1e-5, final_dropout: bool = False, positional_embeddings: Optional[str] = None, num_positional_embeddings: Optional[int] = None, sa_attention_mode: str = "flash", ca_attention_mode: str = "xformers", use_rope: bool = False, interpolation_scale_thw: Tuple[int] = (1, 1, 1), block_idx: Optional[int] = 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" self.use_ada_layer_norm_single = norm_type == "ada_norm_single" self.use_layer_norm = norm_type == "layer_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}." ) 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 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, 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, attention_mode=sa_attention_mode, use_rope=use_rope, interpolation_scale_thw=interpolation_scale_thw, ) # 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, eps=norm_eps) ) 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, attention_mode=ca_attention_mode, # only xformers support attention_mask use_rope=False, # do not position in cross attention interpolation_scale_thw=interpolation_scale_thw, ) # is self-attn if encoder_hidden_states is none else: self.norm2 = None self.attn2 = None # 3. Feed-forward if not self.use_ada_layer_norm_single: self.norm3 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps) self.ff = FeedForward( dim, dropout=dropout, activation_fn=activation_fn, final_dropout=final_dropout, ) # 5. Scale-shift for PixArt-Alpha. if self.use_ada_layer_norm_single: self.scale_shift_table = nn.Parameter(torch.randn(6, dim) / dim**0.5) 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, frame: int = None, height: int = None, width: int = None, ) -> torch.FloatTensor: # Notice that normalization is always applied before the real computation in the following blocks. cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {} # 0. Self-Attention batch_size = hidden_states.shape[0] 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 ) elif self.use_layer_norm: norm_hidden_states = self.norm1(hidden_states) elif self.use_ada_layer_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, frame=frame, height=height, width=width, **cross_attention_kwargs, ) if self.use_ada_layer_norm_zero: attn_output = gate_msa.unsqueeze(1) * attn_output elif self.use_ada_layer_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. Cross-Attention if self.attn2 is not None: if self.use_ada_layer_norm: norm_hidden_states = self.norm2(hidden_states, timestep) elif self.use_ada_layer_norm_zero or self.use_layer_norm: norm_hidden_states = self.norm2(hidden_states) elif self.use_ada_layer_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 else: raise ValueError("Incorrect norm") if self.pos_embed is not None and self.use_ada_layer_norm_single is False: 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 # 2. Feed-forward if not self.use_ada_layer_norm_single: 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] if self.use_ada_layer_norm_single: norm_hidden_states = self.norm2(hidden_states) norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp ff_output = self.ff(norm_hidden_states) if self.use_ada_layer_norm_zero: ff_output = gate_mlp.unsqueeze(1) * ff_output elif self.use_ada_layer_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 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 = CombinedTimestepSizeEmbeddings( 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: Dict[str, torch.Tensor] = None, batch_size: 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, batch_size=batch_size, hidden_dtype=hidden_dtype, resolution=None, aspect_ratio=None ) return self.linear(self.silu(embedded_timestep)), embedded_timestep