src/transformer_flux.py

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

import numpy as np
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_flux import FluxTransformer2DModel


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


class NAGFluxTransformer2DModel(FluxTransformer2DModel):
    def forward(
        self,
        hidden_states: torch.Tensor,
        encoder_hidden_states: torch.Tensor = None,
        pooled_projections: torch.Tensor = None,
        timestep: torch.LongTensor = None,
        img_ids: torch.Tensor = None,
        txt_ids: torch.Tensor = None,
        guidance: torch.Tensor = None,
        joint_attention_kwargs: Optional[Dict[str, Any]] = None,
        controlnet_block_samples=None,
        controlnet_single_block_samples=None,
        return_dict: bool = True,
        controlnet_blocks_repeat: bool = False,
    ) -> Union[torch.Tensor, Transformer2DModelOutput]:
        """
        The [`FluxTransformer2DModel`] forward method.

        Args:
            hidden_states (`torch.Tensor` of shape `(batch_size, image_sequence_length, in_channels)`):
                Input `hidden_states`.
            encoder_hidden_states (`torch.Tensor` of shape `(batch_size, text_sequence_length, joint_attention_dim)`):
                Conditional embeddings (embeddings computed from the input conditions such as prompts) to use.
            pooled_projections (`torch.Tensor` of shape `(batch_size, projection_dim)`): Embeddings projected
                from the embeddings of input conditions.
            timestep ( `torch.LongTensor`):
                Used to indicate denoising step.
            block_controlnet_hidden_states: (`list` of `torch.Tensor`):
                A list of tensors that if specified are added to the residuals of transformer blocks.
            joint_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.transformer_2d.Transformer2DModelOutput`] 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.
        """
        if joint_attention_kwargs is not None:
            joint_attention_kwargs = joint_attention_kwargs.copy()
            lora_scale = joint_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 joint_attention_kwargs is not None and joint_attention_kwargs.get("scale", None) is not None:
                logger.warning(
                    "Passing `scale` via `joint_attention_kwargs` when not using the PEFT backend is ineffective."
                )

        do_nag = hidden_states.shape[0] != encoder_hidden_states.shape[0]

        hidden_states = self.x_embedder(hidden_states)

        timestep = timestep.to(hidden_states.dtype) * 1000
        if guidance is not None:
            guidance = guidance.to(hidden_states.dtype) * 1000
        else:
            guidance = None

        temb = (
            self.time_text_embed(timestep, pooled_projections)
            if guidance is None
            else self.time_text_embed(timestep, guidance, pooled_projections)
        )
        encoder_hidden_states = self.context_embedder(encoder_hidden_states)

        if txt_ids.ndim == 3:
            logger.warning(
                "Passing `txt_ids` 3d torch.Tensor is deprecated."
                "Please remove the batch dimension and pass it as a 2d torch Tensor"
            )
            txt_ids = txt_ids[0]
        if img_ids.ndim == 3:
            logger.warning(
                "Passing `img_ids` 3d torch.Tensor is deprecated."
                "Please remove the batch dimension and pass it as a 2d torch Tensor"
            )
            img_ids = img_ids[0]

        ids = torch.cat((txt_ids, img_ids), dim=0)
        image_rotary_emb = self.pos_embed(ids)

        if joint_attention_kwargs is not None and "ip_adapter_image_embeds" in joint_attention_kwargs:
            ip_adapter_image_embeds = joint_attention_kwargs.pop("ip_adapter_image_embeds")
            ip_hidden_states = self.encoder_hid_proj(ip_adapter_image_embeds)
            joint_attention_kwargs.update({"ip_hidden_states": ip_hidden_states})

        for index_block, block in enumerate(self.transformer_blocks):
            if torch.is_grad_enabled() and self.gradient_checkpointing:
                encoder_hidden_states, hidden_states = self._gradient_checkpointing_func(
                    block,
                    hidden_states,
                    encoder_hidden_states,
                    temb,
                    image_rotary_emb,
                )

            else:
                encoder_hidden_states, hidden_states = block(
                    hidden_states=hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
                    temb=temb,
                    image_rotary_emb=image_rotary_emb,
                    joint_attention_kwargs=joint_attention_kwargs,
                )

            # controlnet residual
            if controlnet_block_samples is not None:
                interval_control = len(self.transformer_blocks) / len(controlnet_block_samples)
                interval_control = int(np.ceil(interval_control))
                # For Xlabs ControlNet.
                if controlnet_blocks_repeat:
                    hidden_states = (
                        hidden_states + controlnet_block_samples[index_block % len(controlnet_block_samples)]
                    )
                else:
                    hidden_states = hidden_states + controlnet_block_samples[index_block // interval_control]

        if do_nag:
            hidden_states = hidden_states.tile(2, 1, 1)
        hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)

        for index_block, block in enumerate(self.single_transformer_blocks):
            if torch.is_grad_enabled() and self.gradient_checkpointing:
                hidden_states = self._gradient_checkpointing_func(
                    block,
                    hidden_states,
                    temb,
                    image_rotary_emb,
                )

            else:
                hidden_states = block(
                    hidden_states=hidden_states,
                    temb=temb,
                    image_rotary_emb=image_rotary_emb,
                    joint_attention_kwargs=joint_attention_kwargs,
                )

            # controlnet residual
            if controlnet_single_block_samples is not None:
                interval_control = len(self.single_transformer_blocks) / len(controlnet_single_block_samples)
                interval_control = int(np.ceil(interval_control))
                controlnet_single_block_sample = controlnet_single_block_samples[index_block // interval_control]
                if do_nag:
                    controlnet_single_block_sample = controlnet_single_block_sample.tile(2, 1, 1)
                hidden_states[:, encoder_hidden_states.shape[1] :, ...] = (
                    hidden_states[:, encoder_hidden_states.shape[1] :, ...] + controlnet_single_block_sample
                )

        hidden_states = hidden_states[:, encoder_hidden_states.shape[1] :, ...]

        if do_nag:
            hidden_states = torch.chunk(hidden_states, 2, dim=0)[0]

        hidden_states = self.norm_out(hidden_states, temb)
        output = self.proj_out(hidden_states)

        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)