from dataclasses import dataclass from typing import Any, Dict, Optional import torch from diffusers.configuration_utils import ConfigMixin, register_to_config from diffusers.models.lora import LoRACompatibleConv, LoRACompatibleLinear from diffusers.models.modeling_utils import ModelMixin from diffusers.models.normalization import AdaLayerNormSingle from diffusers.utils import USE_PEFT_BACKEND, BaseOutput, deprecate, is_torch_version from torch import nn from memo.models.attention import BasicTransformerBlock @dataclass class Transformer2DModelOutput(BaseOutput): sample: torch.FloatTensor ref_feature_list: list[torch.FloatTensor] class Transformer2DModel(ModelMixin, ConfigMixin): _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, 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", is_final_block: bool = False, ): super().__init__() self.use_linear_projection = use_linear_projection self.num_attention_heads = num_attention_heads self.attention_head_dim = attention_head_dim self.is_final_block = is_final_block 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." ) if 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." ) if 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 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) # 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, is_final_block=(is_final_block and d == num_layers - 1), ) for d in range(num_layers) ] ) # 4. Define output layers self.out_channels = in_channels if out_channels is None else out_channels # TODO: should use out_channels for continuous projections if not is_final_block: 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) # 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 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, 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, ): 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) # Retrieve lora scale. lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0 # 1. Input 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) ) # 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]) ref_feature_list = [] 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) return module(*inputs) return custom_forward ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {} hidden_states, ref_feature = 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, ref_feature = block( hidden_states, # shape [5, 4096, 320] attention_mask=attention_mask, encoder_hidden_states=encoder_hidden_states, # shape [1,4,768] encoder_attention_mask=encoder_attention_mask, timestep=timestep, cross_attention_kwargs=cross_attention_kwargs, class_labels=class_labels, ) ref_feature_list.append(ref_feature) # 3. Output output = None if self.is_final_block: if not return_dict: return (output, ref_feature_list) return Transformer2DModelOutput(sample=output, ref_feature_list=ref_feature_list) 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 if not return_dict: return (output, ref_feature_list) return Transformer2DModelOutput(sample=output, ref_feature_list=ref_feature_list)