flux_model.py

# Copyright 2024 Black Forest Labs, The HuggingFace Team and The InstantX 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.

from diffusers.configuration_utils import ConfigMixin, register_to_config
from diffusers.models.activations import FP32SiLU, get_activation
from diffusers.models.embeddings import Timesteps, PixArtAlphaTextProjection
from diffusers.models.modeling_utils import ModelMixin
from diffusers.models.transformers.transformer_flux import AdaLayerNormContinuous, CombinedTimestepGuidanceTextProjEmbeddings, CombinedTimestepTextProjEmbeddings, EmbedND, FluxSingleTransformerBlock, FluxTransformerBlock
from diffusers.models.modeling_outputs import Transformer2DModelOutput
from diffusers.utils import logging
import torch
from torch import nn

logger = logging.get_logger(__name__)

class FluxTransformer2DModel(ModelMixin, ConfigMixin):
    """
    The Transformer model introduced in Flux.

    Reference: https://blackforestlabs.ai/announcing-black-forest-labs/

    Parameters:
        patch_size (`int`): Patch size to turn the input data into small patches.
        in_channels (`int`, *optional*, defaults to 16): The number of channels in the input.
        num_layers (`int`, *optional*, defaults to 18): The number of layers of MMDiT blocks to use.
        num_single_layers (`int`, *optional*, defaults to 18): The number of layers of single DiT blocks to use.
        attention_head_dim (`int`, *optional*, defaults to 64): The number of channels in each head.
        num_attention_heads (`int`, *optional*, defaults to 18): The number of heads to use for multi-head attention.
        joint_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use.
        pooled_projection_dim (`int`): Number of dimensions to use when projecting the `pooled_projections`.
        guidance_embeds (`bool`, defaults to False): Whether to use guidance embeddings.
    """

    _supports_gradient_checkpointing = True
    _no_split_modules = ["FluxTransformerBlock", "FluxSingleTransformerBlock"]

    @register_to_config
    def __init__(
        self,
        patch_size: int = 1,
        in_channels: int = 64,
        num_layers: int = 19,
        num_single_layers: int = 38,
        attention_head_dim: int = 128,
        num_attention_heads: int = 24,
        joint_attention_dim: int = 4096,
        pooled_projection_dim: int = 768,
        guidance_embeds: bool = False,
        axes_dims_rope=(16, 56, 56),
        device=None
    ):
        super().__init__()
        self.out_channels = in_channels
        self.inner_dim = self.config.num_attention_heads * self.config.attention_head_dim

        # self.pos_embed = FluxPosEmbed(theta=10000, axes_dim=axes_dims_rope)
        self.pos_embed = EmbedND(dim=self.inner_dim, theta=10000, axes_dim=axes_dims_rope).to(device)

        text_time_guidance_cls = (
            CombinedTimestepGuidanceTextProjEmbeddings if guidance_embeds else CombinedTimestepTextProjEmbeddings
        )
        self.time_text_embed = text_time_guidance_cls(
            embedding_dim=self.inner_dim, pooled_projection_dim=self.config.pooled_projection_dim
        ).to(device)

        self.context_embedder = nn.Linear(self.config.joint_attention_dim, self.inner_dim).to(device)
        self.x_embedder = nn.Linear(self.config.in_channels, self.inner_dim).to(device)

        self.transformer_blocks = nn.ModuleList(
            [
                FluxTransformerBlock(
                    dim=self.inner_dim,
                    num_attention_heads=self.config.num_attention_heads,
                    attention_head_dim=self.config.attention_head_dim,
                ).to(device)
                for i in range(self.config.num_layers)
            ]
        )

        self.single_transformer_blocks = nn.ModuleList(
            [
                FluxSingleTransformerBlock(
                    dim=self.inner_dim,
                    num_attention_heads=self.config.num_attention_heads,
                    attention_head_dim=self.config.attention_head_dim,
                ).to(device)
                for i in range(self.config.num_single_layers)
            ]
        )

        self.norm_out = AdaLayerNormContinuous(self.inner_dim, self.inner_dim, elementwise_affine=False, eps=1e-6).to(device)
        self.proj_out = nn.Linear(self.inner_dim, patch_size * patch_size * self.out_channels, bias=True).to(device)

        self.pul_id = None
        self.pul_id_weight = 1.0

        self.gradient_checkpointing = False

    @property
    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
    def attn_processors(self):
        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: nn.Module, processors):
            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):
        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: 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 _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: 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 = None,
        controlnet_block_samples=None,
        controlnet_single_block_samples=None,
        return_dict: bool = True
    ):
        """
        The [`FluxTransformer2DModel`] forward method.

        Args:
            hidden_states (`torch.FloatTensor` of shape `(batch size, channel, height, width)`):
                Input `hidden_states`.
            encoder_hidden_states (`torch.FloatTensor` of shape `(batch size, sequence_len, embed_dims)`):
                Conditional embeddings (embeddings computed from the input conditions such as prompts) to use.
            pooled_projections (`torch.FloatTensor` 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.
        """
        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)

        ###
        # Modified by huggingface/twodgirl.
        # Code from diffusers and PuLID.
        
        ids = torch.cat((txt_ids, img_ids), dim=1)
        image_rotary_emb = self.pos_embed(ids)
        ca_index = 0

        for index_block, block in enumerate(self.transformer_blocks):
            encoder_hidden_states, hidden_states = block(
                hidden_states=hidden_states,
                encoder_hidden_states=encoder_hidden_states,
                temb=temb,
                image_rotary_emb=image_rotary_emb,
            )

            if index_block % self.pulid_double_interval == 0 and self.pul_id is not None:
                weighted = self.pul_id_weight * self.pulid_ca[ca_index](self.pul_id, hidden_states.to(self.pul_id.dtype))
                hidden_states = hidden_states + weighted.to(hidden_states.dtype)
                ca_index += 1

        hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)

        for index_block, block in enumerate(self.single_transformer_blocks):
            hidden_states = block(
                hidden_states=hidden_states,
                temb=temb,
                image_rotary_emb=image_rotary_emb,
            )
            if index_block % self.pulid_single_interval == 0 and self.pul_id is not None:
                encoder_hidden_states, real_ = hidden_states[:, :encoder_hidden_states.shape[1], ...], hidden_states[:, encoder_hidden_states.shape[1]:, ...]
                weighted = self.pul_id_weight * self.pulid_ca[ca_index](self.pul_id, real_.to(self.pul_id.dtype))
                real_ = real_ + weighted.to(real_.dtype)
                hidden_states = torch.cat([encoder_hidden_states, real_], dim=1)
                ca_index += 1

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

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

        if not return_dict:
            return (output,)

        return Transformer2DModelOutput(sample=output)