src/transformer_wan_nag.py

from typing import Any, Dict, Optional, Tuple, Union

import torch

from diffusers.utils import USE_PEFT_BACKEND, logging, scale_lora_layers, unscale_lora_layers
from diffusers.models.modeling_outputs import Transformer2DModelOutput
from diffusers.models.transformers.transformer_wan import WanTransformer3DModel
from diffusers.models.attention_processor import AttentionProcessor


logger = logging.get_logger(__name__)  # pylint: disable=invalid-name


class NagWanTransformer3DModel(WanTransformer3DModel):
    @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()

            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)

    def forward(
        self,
        hidden_states: torch.Tensor,
        timestep: torch.LongTensor,
        encoder_hidden_states: torch.Tensor,
        encoder_hidden_states_image: Optional[torch.Tensor] = None,
        return_dict: bool = True,
        attention_kwargs: Optional[Dict[str, Any]] = None,
    ) -> Union[torch.Tensor, Dict[str, torch.Tensor]]:
        if attention_kwargs is not None:
            attention_kwargs = attention_kwargs.copy()
            lora_scale = 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)
        else:
            if attention_kwargs is not None and attention_kwargs.get("scale", None) is not None:
                logger.warning(
                    "Passing `scale` via `attention_kwargs` when not using the PEFT backend is ineffective."
                )

        batch_size, num_channels, num_frames, height, width = hidden_states.shape
        p_t, p_h, p_w = self.config.patch_size
        post_patch_num_frames = num_frames // p_t
        post_patch_height = height // p_h
        post_patch_width = width // p_w

        rotary_emb = self.rope(hidden_states)

        hidden_states = self.patch_embedding(hidden_states)
        hidden_states = hidden_states.flatten(2).transpose(1, 2)

        temb, timestep_proj, encoder_hidden_states, encoder_hidden_states_image = self.condition_embedder(
            timestep, encoder_hidden_states, encoder_hidden_states_image
        )
        timestep_proj = timestep_proj.unflatten(1, (6, -1))

        if encoder_hidden_states_image is not None:
            bs_encoder_hidden_states = len(encoder_hidden_states)
            bs_encoder_hidden_states_image = len(encoder_hidden_states_image)
            bs_scale = bs_encoder_hidden_states / bs_encoder_hidden_states_image
            assert bs_scale in [1, 2, 3]
            if bs_scale != 1:
                encoder_hidden_states_image = encoder_hidden_states_image.tile(int(bs_scale), 1, 1)
            encoder_hidden_states = torch.concat([encoder_hidden_states_image, encoder_hidden_states], dim=1)

        # 4. Transformer blocks
        if torch.is_grad_enabled() and self.gradient_checkpointing:
            for block in self.blocks:
                hidden_states = self._gradient_checkpointing_func(
                    block, hidden_states, encoder_hidden_states, timestep_proj, rotary_emb
                )
        else:
            for block in self.blocks:
                hidden_states = block(hidden_states, encoder_hidden_states, timestep_proj, rotary_emb)

        # 5. Output norm, projection & unpatchify
        shift, scale = (self.scale_shift_table + temb.unsqueeze(1)).chunk(2, dim=1)

        # Move the shift and scale tensors to the same device as hidden_states.
        # When using multi-GPU inference via accelerate these will be on the
        # first device rather than the last device, which hidden_states ends up
        # on.
        shift = shift.to(hidden_states.device)
        scale = scale.to(hidden_states.device)

        hidden_states = (self.norm_out(hidden_states.float()) * (1 + scale) + shift).type_as(hidden_states)
        hidden_states = self.proj_out(hidden_states)

        hidden_states = hidden_states.reshape(
            batch_size, post_patch_num_frames, post_patch_height, post_patch_width, p_t, p_h, p_w, -1
        )
        hidden_states = hidden_states.permute(0, 7, 1, 4, 2, 5, 3, 6)
        output = hidden_states.flatten(6, 7).flatten(4, 5).flatten(2, 3)

        if USE_PEFT_BACKEND:
            # remove `lora_scale` from each PEFT layer
            unscale_lora_layers(self, lora_scale)

        if not return_dict:
            return (output,)

        return Transformer2DModelOutput(sample=output)