diffusers/models/transformers/transformer_sd3.py

# Copyright 2025 Stability AI, 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 typing import Any

import torch
import torch.nn as nn

from ...configuration_utils import ConfigMixin, register_to_config
from ...loaders import FromOriginalModelMixin, PeftAdapterMixin, SD3Transformer2DLoadersMixin
from ...utils import apply_lora_scale, logging
from ...utils.torch_utils import maybe_allow_in_graph
from ..attention import AttentionMixin, FeedForward, JointTransformerBlock
from ..attention_processor import (
    Attention,
    FusedJointAttnProcessor2_0,
    JointAttnProcessor2_0,
)
from ..embeddings import CombinedTimestepTextProjEmbeddings, PatchEmbed
from ..modeling_outputs import Transformer2DModelOutput
from ..modeling_utils import ModelMixin
from ..normalization import AdaLayerNormContinuous, AdaLayerNormZero


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


@maybe_allow_in_graph
class SD3SingleTransformerBlock(nn.Module):
    def __init__(
        self,
        dim: int,
        num_attention_heads: int,
        attention_head_dim: int,
    ):
        super().__init__()

        self.norm1 = AdaLayerNormZero(dim)
        self.attn = Attention(
            query_dim=dim,
            dim_head=attention_head_dim,
            heads=num_attention_heads,
            out_dim=dim,
            bias=True,
            processor=JointAttnProcessor2_0(),
            eps=1e-6,
        )

        self.norm2 = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-6)
        self.ff = FeedForward(dim=dim, dim_out=dim, activation_fn="gelu-approximate")

    def forward(self, hidden_states: torch.Tensor, temb: torch.Tensor):
        # 1. Attention
        norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(hidden_states, emb=temb)
        attn_output = self.attn(hidden_states=norm_hidden_states, encoder_hidden_states=None)
        attn_output = gate_msa.unsqueeze(1) * attn_output
        hidden_states = hidden_states + attn_output

        # 2. Feed Forward
        norm_hidden_states = self.norm2(hidden_states)
        norm_hidden_states = norm_hidden_states * (1 + scale_mlp.unsqueeze(1)) + shift_mlp.unsqueeze(1)
        ff_output = self.ff(norm_hidden_states)
        ff_output = gate_mlp.unsqueeze(1) * ff_output
        hidden_states = hidden_states + ff_output

        return hidden_states


class SD3Transformer2DModel(
    ModelMixin, AttentionMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin, SD3Transformer2DLoadersMixin
):
    """
    The Transformer model introduced in [Stable Diffusion 3](https://huggingface.co/papers/2403.03206).

    Parameters:
        sample_size (`int`, defaults to `128`):
            The width/height of the latents. This is fixed during training since it is used to learn a number of
            position embeddings.
        patch_size (`int`, defaults to `2`):
            Patch size to turn the input data into small patches.
        in_channels (`int`, defaults to `16`):
            The number of latent channels in the input.
        num_layers (`int`, defaults to `18`):
            The number of layers of transformer blocks to use.
        attention_head_dim (`int`, defaults to `64`):
            The number of channels in each head.
        num_attention_heads (`int`, defaults to `18`):
            The number of heads to use for multi-head attention.
        joint_attention_dim (`int`, defaults to `4096`):
            The embedding dimension to use for joint text-image attention.
        caption_projection_dim (`int`, defaults to `1152`):
            The embedding dimension of caption embeddings.
        pooled_projection_dim (`int`, defaults to `2048`):
            The embedding dimension of pooled text projections.
        out_channels (`int`, defaults to `16`):
            The number of latent channels in the output.
        pos_embed_max_size (`int`, defaults to `96`):
            The maximum latent height/width of positional embeddings.
        dual_attention_layers (`tuple[int, ...]`, defaults to `()`):
            The number of dual-stream transformer blocks to use.
        qk_norm (`str`, *optional*, defaults to `None`):
            The normalization to use for query and key in the attention layer. If `None`, no normalization is used.
    """

    _supports_gradient_checkpointing = True
    _no_split_modules = ["JointTransformerBlock"]
    _skip_layerwise_casting_patterns = ["pos_embed", "norm"]

    @register_to_config
    def __init__(
        self,
        sample_size: int = 128,
        patch_size: int = 2,
        in_channels: int = 16,
        num_layers: int = 18,
        attention_head_dim: int = 64,
        num_attention_heads: int = 18,
        joint_attention_dim: int = 4096,
        caption_projection_dim: int = 1152,
        pooled_projection_dim: int = 2048,
        out_channels: int = 16,
        pos_embed_max_size: int = 96,
        dual_attention_layers: tuple[
            int, ...
        ] = (),  # () for sd3.0; (0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12) for sd3.5
        qk_norm: str | None = None,
    ):
        super().__init__()
        self.out_channels = out_channels if out_channels is not None else in_channels
        self.inner_dim = num_attention_heads * attention_head_dim

        self.pos_embed = PatchEmbed(
            height=sample_size,
            width=sample_size,
            patch_size=patch_size,
            in_channels=in_channels,
            embed_dim=self.inner_dim,
            pos_embed_max_size=pos_embed_max_size,  # hard-code for now.
        )
        self.time_text_embed = CombinedTimestepTextProjEmbeddings(
            embedding_dim=self.inner_dim, pooled_projection_dim=pooled_projection_dim
        )
        self.context_embedder = nn.Linear(joint_attention_dim, caption_projection_dim)

        self.transformer_blocks = nn.ModuleList(
            [
                JointTransformerBlock(
                    dim=self.inner_dim,
                    num_attention_heads=num_attention_heads,
                    attention_head_dim=attention_head_dim,
                    context_pre_only=i == num_layers - 1,
                    qk_norm=qk_norm,
                    use_dual_attention=True if i in dual_attention_layers else False,
                )
                for i in range(num_layers)
            ]
        )

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

        self.gradient_checkpointing = False

    # Copied from diffusers.models.unets.unet_3d_condition.UNet3DConditionModel.enable_forward_chunking
    def enable_forward_chunking(self, chunk_size: int | None = None, dim: int = 0) -> None:
        """
        Sets the attention processor to use [feed forward
        chunking](https://huggingface.co/blog/reformer#2-chunked-feed-forward-layers).

        Parameters:
            chunk_size (`int`, *optional*):
                The chunk size of the feed-forward layers. If not specified, will run feed-forward layer individually
                over each tensor of dim=`dim`.
            dim (`int`, *optional*, defaults to `0`):
                The dimension over which the feed-forward computation should be chunked. Choose between dim=0 (batch)
                or dim=1 (sequence length).
        """
        if dim not in [0, 1]:
            raise ValueError(f"Make sure to set `dim` to either 0 or 1, not {dim}")

        # By default chunk size is 1
        chunk_size = chunk_size or 1

        def fn_recursive_feed_forward(module: torch.nn.Module, chunk_size: int, dim: int):
            if hasattr(module, "set_chunk_feed_forward"):
                module.set_chunk_feed_forward(chunk_size=chunk_size, dim=dim)

            for child in module.children():
                fn_recursive_feed_forward(child, chunk_size, dim)

        for module in self.children():
            fn_recursive_feed_forward(module, chunk_size, dim)

    # Copied from diffusers.models.unets.unet_3d_condition.UNet3DConditionModel.disable_forward_chunking
    def disable_forward_chunking(self):
        def fn_recursive_feed_forward(module: torch.nn.Module, chunk_size: int, dim: int):
            if hasattr(module, "set_chunk_feed_forward"):
                module.set_chunk_feed_forward(chunk_size=chunk_size, dim=dim)

            for child in module.children():
                fn_recursive_feed_forward(child, chunk_size, dim)

        for module in self.children():
            fn_recursive_feed_forward(module, None, 0)

    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.fuse_qkv_projections with FusedAttnProcessor2_0->FusedJointAttnProcessor2_0
    def fuse_qkv_projections(self):
        """
        Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, key, value)
        are fused. For cross-attention modules, key and value projection matrices are fused.

        > [!WARNING] > This API is 🧪 experimental.
        """
        self.original_attn_processors = None

        for _, attn_processor in self.attn_processors.items():
            if "Added" in str(attn_processor.__class__.__name__):
                raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")

        self.original_attn_processors = self.attn_processors

        for module in self.modules():
            if isinstance(module, Attention):
                module.fuse_projections(fuse=True)

        self.set_attn_processor(FusedJointAttnProcessor2_0())

    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.unfuse_qkv_projections
    def unfuse_qkv_projections(self):
        """Disables the fused QKV projection if enabled.

        > [!WARNING] > This API is 🧪 experimental.

        """
        if self.original_attn_processors is not None:
            self.set_attn_processor(self.original_attn_processors)

    @apply_lora_scale("joint_attention_kwargs")
    def forward(
        self,
        hidden_states: torch.Tensor,
        encoder_hidden_states: torch.Tensor = None,
        pooled_projections: torch.Tensor = None,
        timestep: torch.LongTensor = None,
        block_controlnet_hidden_states: list = None,
        joint_attention_kwargs: dict[str, Any] | None = None,
        return_dict: bool = True,
        skip_layers: list[int] | None = None,
    ) -> torch.Tensor | Transformer2DModelOutput:
        """
        The [`SD3Transformer2DModel`] forward method.

        Args:
            hidden_states (`torch.Tensor` of shape `(batch size, channel, height, width)`):
                Input `hidden_states`.
            encoder_hidden_states (`torch.Tensor` of shape `(batch size, sequence_len, embed_dims)`):
                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.
            skip_layers (`list` of `int`, *optional*):
                A list of layer indices to skip during the forward pass.

        Returns:
            If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
            `tuple` where the first element is the sample tensor.
        """

        height, width = hidden_states.shape[-2:]

        hidden_states = self.pos_embed(hidden_states)  # takes care of adding positional embeddings too.
        temb = self.time_text_embed(timestep, pooled_projections)
        encoder_hidden_states = self.context_embedder(encoder_hidden_states)

        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, ip_temb = self.image_proj(ip_adapter_image_embeds, timestep)

            joint_attention_kwargs.update(ip_hidden_states=ip_hidden_states, temb=ip_temb)

        for index_block, block in enumerate(self.transformer_blocks):
            # Skip specified layers
            is_skip = True if skip_layers is not None and index_block in skip_layers else False

            if torch.is_grad_enabled() and self.gradient_checkpointing and not is_skip:
                encoder_hidden_states, hidden_states = self._gradient_checkpointing_func(
                    block,
                    hidden_states,
                    encoder_hidden_states,
                    temb,
                    joint_attention_kwargs,
                )
            elif not is_skip:
                encoder_hidden_states, hidden_states = block(
                    hidden_states=hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
                    temb=temb,
                    joint_attention_kwargs=joint_attention_kwargs,
                )

            # controlnet residual
            if block_controlnet_hidden_states is not None and block.context_pre_only is False:
                interval_control = len(self.transformer_blocks) / len(block_controlnet_hidden_states)
                hidden_states = hidden_states + block_controlnet_hidden_states[int(index_block / interval_control)]

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

        # unpatchify
        patch_size = self.config.patch_size
        height = height // patch_size
        width = width // patch_size

        hidden_states = hidden_states.reshape(
            shape=(hidden_states.shape[0], height, width, patch_size, patch_size, self.out_channels)
        )
        hidden_states = torch.einsum("nhwpqc->nchpwq", hidden_states)
        output = hidden_states.reshape(
            shape=(hidden_states.shape[0], self.out_channels, height * patch_size, width * patch_size)
        )

        if not return_dict:
            return (output,)

        return Transformer2DModelOutput(sample=output)