diffusion/sana_transformer.py

# Copyright 2024 The HuggingFace 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, Dict, Optional, Tuple, Union

import torch
import torch.nn.functional as F
from torch import nn

# from ...configuration_utils import ConfigMixin, register_to_config
# from ...loaders import FromOriginalModelMixin, PeftAdapterMixin
# from ...utils import USE_PEFT_BACKEND, logging, scale_lora_layers, unscale_lora_layers
# from ..attention_processor import (
#     Attention,
#     AttentionProcessor,
#     SanaLinearAttnProcessor2_0,
# )
# from ..embeddings import PatchEmbed, PixArtAlphaTextProjection, TimestepEmbedding, Timesteps
# from ..modeling_outputs import Transformer2DModelOutput
# from ..modeling_utils import ModelMixin
# from ..normalization import AdaLayerNormSingle, RMSNorm
from diffusers.configuration_utils import ConfigMixin, register_to_config
from diffusers.loaders import FromOriginalModelMixin, PeftAdapterMixin
from diffusers.utils import USE_PEFT_BACKEND, logging, scale_lora_layers, unscale_lora_layers
from diffusers.models.attention_processor import (
    Attention,
    AttentionProcessor,
    SanaLinearAttnProcessor2_0,
)
from diffusers.models.embeddings import PatchEmbed, PixArtAlphaTextProjection, TimestepEmbedding, Timesteps
from diffusers.models.modeling_outputs import Transformer2DModelOutput
from diffusers.models.modeling_utils import ModelMixin
from diffusers.models.normalization import AdaLayerNormSingle, RMSNorm


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


class GLUMBConv(nn.Module):
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        expand_ratio: float = 4,
        norm_type: Optional[str] = None,
        residual_connection: bool = True,
    ) -> None:
        super().__init__()

        hidden_channels = int(expand_ratio * in_channels)
        self.norm_type = norm_type
        self.residual_connection = residual_connection

        self.nonlinearity = nn.SiLU()
        self.conv_inverted = nn.Conv2d(in_channels, hidden_channels * 2, 1, 1, 0)
        self.conv_depth = nn.Conv2d(hidden_channels * 2, hidden_channels * 2, 3, 1, 1, groups=hidden_channels * 2)
        self.conv_point = nn.Conv2d(hidden_channels, out_channels, 1, 1, 0, bias=False)

        self.norm = None
        if norm_type == "rms_norm":
            self.norm = RMSNorm(out_channels, eps=1e-5, elementwise_affine=True, bias=True)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        if self.residual_connection:
            residual = hidden_states

        hidden_states = self.conv_inverted(hidden_states)
        hidden_states = self.nonlinearity(hidden_states)

        hidden_states = self.conv_depth(hidden_states)
        hidden_states, gate = torch.chunk(hidden_states, 2, dim=1)
        hidden_states = hidden_states * self.nonlinearity(gate)

        hidden_states = self.conv_point(hidden_states)

        if self.norm_type == "rms_norm":
            # move channel to the last dimension so we apply RMSnorm across channel dimension
            hidden_states = self.norm(hidden_states.movedim(1, -1)).movedim(-1, 1)

        if self.residual_connection:
            hidden_states = hidden_states + residual

        return hidden_states


class SanaModulatedNorm(nn.Module):
    def __init__(self, dim: int, elementwise_affine: bool = False, eps: float = 1e-6):
        super().__init__()
        self.norm = nn.LayerNorm(dim, elementwise_affine=elementwise_affine, eps=eps)

    def forward(
        self, hidden_states: torch.Tensor, temb: torch.Tensor, scale_shift_table: torch.Tensor
    ) -> torch.Tensor:
        hidden_states = self.norm(hidden_states)
        shift, scale = (scale_shift_table[None] + temb[:, None].to(scale_shift_table.device)).chunk(2, dim=1)
        hidden_states = hidden_states * (1 + scale) + shift
        return hidden_states


class SanaCombinedTimestepGuidanceEmbeddings(nn.Module):
    def __init__(self, embedding_dim):
        super().__init__()
        self.time_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0)
        self.timestep_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=embedding_dim)

        self.guidance_condition_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0)
        self.guidance_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=embedding_dim)

        self.silu = nn.SiLU()
        self.linear = nn.Linear(embedding_dim, 6 * embedding_dim, bias=True)

    def forward(self, timestep: torch.Tensor, guidance: torch.Tensor = None, hidden_dtype: torch.dtype = None):
        timesteps_proj = self.time_proj(timestep)
        timesteps_emb = self.timestep_embedder(timesteps_proj.to(dtype=hidden_dtype))  # (N, D)

        guidance_proj = self.guidance_condition_proj(guidance)
        guidance_emb = self.guidance_embedder(guidance_proj.to(dtype=hidden_dtype))
        conditioning = timesteps_emb + guidance_emb

        return self.linear(self.silu(conditioning)), conditioning


class SanaAttnProcessor2_0:
    r"""
    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
    """

    def __init__(self):
        if not hasattr(F, "scaled_dot_product_attention"):
            raise ImportError("SanaAttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")

    def __call__(
        self,
        attn: Attention,
        hidden_states: torch.Tensor,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        batch_size, sequence_length, _ = (
            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
        )

        if attention_mask is not None:
            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
            # scaled_dot_product_attention expects attention_mask shape to be
            # (batch, heads, source_length, target_length)
            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])

        query = attn.to_q(hidden_states)

        if encoder_hidden_states is None:
            encoder_hidden_states = hidden_states

        key = attn.to_k(encoder_hidden_states)
        value = attn.to_v(encoder_hidden_states)

        if attn.norm_q is not None:
            query = attn.norm_q(query)
        if attn.norm_k is not None:
            key = attn.norm_k(key)

        inner_dim = key.shape[-1]
        head_dim = inner_dim // attn.heads

        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)

        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)

        # the output of sdp = (batch, num_heads, seq_len, head_dim)
        # TODO: add support for attn.scale when we move to Torch 2.1
        hidden_states = F.scaled_dot_product_attention(
            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
        )

        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
        hidden_states = hidden_states.to(query.dtype)

        # linear proj
        hidden_states = attn.to_out[0](hidden_states)
        # dropout
        hidden_states = attn.to_out[1](hidden_states)

        hidden_states = hidden_states / attn.rescale_output_factor

        return hidden_states


class SanaTransformerBlock(nn.Module):
    r"""
    Transformer block introduced in [Sana](https://huggingface.co/papers/2410.10629).
    """

    def __init__(
        self,
        dim: int = 2240,
        num_attention_heads: int = 70,
        attention_head_dim: int = 32,
        dropout: float = 0.0,
        num_cross_attention_heads: Optional[int] = 20,
        cross_attention_head_dim: Optional[int] = 112,
        cross_attention_dim: Optional[int] = 2240,
        attention_bias: bool = True,
        norm_elementwise_affine: bool = False,
        norm_eps: float = 1e-6,
        attention_out_bias: bool = True,
        mlp_ratio: float = 2.5,
        qk_norm: Optional[str] = None,
    ) -> None:
        super().__init__()

        # 1. Self Attention
        self.norm1 = nn.LayerNorm(dim, elementwise_affine=False, eps=norm_eps)
        self.attn1 = Attention(
            query_dim=dim,
            heads=num_attention_heads,
            dim_head=attention_head_dim,
            kv_heads=num_attention_heads if qk_norm is not None else None,
            qk_norm=qk_norm,
            dropout=dropout,
            bias=attention_bias,
            cross_attention_dim=None,
            processor=SanaLinearAttnProcessor2_0(),
        )

        # 2. Cross Attention
        if cross_attention_dim is not None:
            self.norm2 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
            self.attn2 = Attention(
                query_dim=dim,
                qk_norm=qk_norm,
                kv_heads=num_cross_attention_heads if qk_norm is not None else None,
                cross_attention_dim=cross_attention_dim,
                heads=num_cross_attention_heads,
                dim_head=cross_attention_head_dim,
                dropout=dropout,
                bias=True,
                out_bias=attention_out_bias,
                processor=SanaAttnProcessor2_0(),
            )

        # 3. Feed-forward
        self.ff = GLUMBConv(dim, dim, mlp_ratio, norm_type=None, residual_connection=False)

        self.scale_shift_table = nn.Parameter(torch.randn(6, dim) / dim**0.5)

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        encoder_attention_mask: Optional[torch.Tensor] = None,
        timestep: Optional[torch.LongTensor] = None,
        height: int = None,
        width: int = None,
    ) -> torch.Tensor:
        batch_size = hidden_states.shape[0]

        # 1. Modulation
        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
            self.scale_shift_table[None] + timestep.reshape(batch_size, 6, -1)
        ).chunk(6, dim=1)

        # 2. Self Attention
        norm_hidden_states = self.norm1(hidden_states)
        norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa
        norm_hidden_states = norm_hidden_states.to(hidden_states.dtype)

        attn_output = self.attn1(norm_hidden_states)
        hidden_states = hidden_states + gate_msa * attn_output

        # 3. Cross Attention
        if self.attn2 is not None:
            attn_output = self.attn2(
                hidden_states,
                encoder_hidden_states=encoder_hidden_states,
                attention_mask=encoder_attention_mask,
            )
            hidden_states = attn_output + hidden_states

        # 4. Feed-forward
        norm_hidden_states = self.norm2(hidden_states)
        norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp

        norm_hidden_states = norm_hidden_states.unflatten(1, (height, width)).permute(0, 3, 1, 2)
        ff_output = self.ff(norm_hidden_states)
        ff_output = ff_output.flatten(2, 3).permute(0, 2, 1)
        hidden_states = hidden_states + gate_mlp * ff_output

        return hidden_states


class SanaTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin):
    r"""
    A 2D Transformer model introduced in [Sana](https://huggingface.co/papers/2410.10629) family of models.

    Args:
        in_channels (`int`, defaults to `32`):
            The number of channels in the input.
        out_channels (`int`, *optional*, defaults to `32`):
            The number of channels in the output.
        num_attention_heads (`int`, defaults to `70`):
            The number of heads to use for multi-head attention.
        attention_head_dim (`int`, defaults to `32`):
            The number of channels in each head.
        num_layers (`int`, defaults to `20`):
            The number of layers of Transformer blocks to use.
        num_cross_attention_heads (`int`, *optional*, defaults to `20`):
            The number of heads to use for cross-attention.
        cross_attention_head_dim (`int`, *optional*, defaults to `112`):
            The number of channels in each head for cross-attention.
        cross_attention_dim (`int`, *optional*, defaults to `2240`):
            The number of channels in the cross-attention output.
        caption_channels (`int`, defaults to `2304`):
            The number of channels in the caption embeddings.
        mlp_ratio (`float`, defaults to `2.5`):
            The expansion ratio to use in the GLUMBConv layer.
        dropout (`float`, defaults to `0.0`):
            The dropout probability.
        attention_bias (`bool`, defaults to `False`):
            Whether to use bias in the attention layer.
        sample_size (`int`, defaults to `32`):
            The base size of the input latent.
        patch_size (`int`, defaults to `1`):
            The size of the patches to use in the patch embedding layer.
        norm_elementwise_affine (`bool`, defaults to `False`):
            Whether to use elementwise affinity in the normalization layer.
        norm_eps (`float`, defaults to `1e-6`):
            The epsilon value for the normalization layer.
        qk_norm (`str`, *optional*, defaults to `None`):
            The normalization to use for the query and key.
        timestep_scale (`float`, defaults to `1.0`):
            The scale to use for the timesteps.
    """

    _supports_gradient_checkpointing = True
    _no_split_modules = ["SanaTransformerBlock", "PatchEmbed", "SanaModulatedNorm"]
    _skip_layerwise_casting_patterns = ["patch_embed", "norm"]

    @register_to_config
    def __init__(
        self,
        in_channels: int = 32,
        out_channels: Optional[int] = 32,
        num_attention_heads: int = 70,
        attention_head_dim: int = 32,
        num_layers: int = 20,
        num_cross_attention_heads: Optional[int] = 20,
        cross_attention_head_dim: Optional[int] = 112,
        cross_attention_dim: Optional[int] = 2240,
        caption_channels: int = 2304,
        mlp_ratio: float = 2.5,
        dropout: float = 0.0,
        attention_bias: bool = False,
        sample_size: int = 32,
        patch_size: int = 1,
        norm_elementwise_affine: bool = False,
        norm_eps: float = 1e-6,
        interpolation_scale: Optional[int] = None,
        guidance_embeds: bool = False,
        guidance_embeds_scale: float = 0.1,
        qk_norm: Optional[str] = None,
        timestep_scale: float = 1.0,
    ) -> None:
        super().__init__()

        out_channels = out_channels or in_channels
        inner_dim = num_attention_heads * attention_head_dim

        # 1. Patch Embedding
        self.patch_embed = PatchEmbed(
            height=sample_size,
            width=sample_size,
            patch_size=patch_size,
            in_channels=in_channels,
            embed_dim=inner_dim,
            interpolation_scale=interpolation_scale,
            pos_embed_type="sincos" if interpolation_scale is not None else None,
        )

        # 2. Additional condition embeddings
        if guidance_embeds:
            self.time_embed = SanaCombinedTimestepGuidanceEmbeddings(inner_dim)
        else:
            self.time_embed = AdaLayerNormSingle(inner_dim)

        self.caption_projection = PixArtAlphaTextProjection(in_features=caption_channels, hidden_size=inner_dim)
        self.caption_norm = RMSNorm(inner_dim, eps=1e-5, elementwise_affine=True)

        # 3. Transformer blocks
        self.transformer_blocks = nn.ModuleList(
            [
                SanaTransformerBlock(
                    inner_dim,
                    num_attention_heads,
                    attention_head_dim,
                    dropout=dropout,
                    num_cross_attention_heads=num_cross_attention_heads,
                    cross_attention_head_dim=cross_attention_head_dim,
                    cross_attention_dim=cross_attention_dim,
                    attention_bias=attention_bias,
                    norm_elementwise_affine=norm_elementwise_affine,
                    norm_eps=norm_eps,
                    mlp_ratio=mlp_ratio,
                    qk_norm=qk_norm,
                )
                for _ in range(num_layers)
            ]
        )

        # 4. Output blocks
        self.scale_shift_table = nn.Parameter(torch.randn(2, inner_dim) / inner_dim**0.5)
        self.norm_out = SanaModulatedNorm(inner_dim, elementwise_affine=False, eps=1e-6)
        self.proj_out = nn.Linear(inner_dim, patch_size * patch_size * out_channels)

        self.gradient_checkpointing = False

    @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 register_block_hooks(self, block_indices=None):
        """
        为指定的transformer block注册钩子以获取输出
        
        Args:
            block_indices (list, optional): 要监视的block索引列表，None表示所有block
            
        Returns:
            dict: block_outputs字典，键为block索引，值为对应的输出
        """
        block_outputs = {}
        hooks = []
        
        indices = block_indices if block_indices is not None else range(len(self.transformer_blocks))
        
        for idx in indices:
            # print('idx',idx)
            if idx < 0 or idx >= len(self.transformer_blocks):
                continue
                
            def get_hook(i):
                def hook(module, input, output):
                    block_outputs[i] = output
                return hook
            
            h = self.transformer_blocks[idx].register_forward_hook(get_hook(idx))
            hooks.append(h)
            
        return block_outputs, hooks
    
    def remove_hooks(self, hooks):
        """移除所有注册的钩子"""
        for h in hooks:
            h.remove()


    def forward(
        self,
        hidden_states: torch.Tensor,
        encoder_hidden_states: torch.Tensor,
        timestep: torch.Tensor,
        guidance: Optional[torch.Tensor] = None,
        encoder_attention_mask: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        attention_kwargs: Optional[Dict[str, Any]] = None,
        return_dict: bool = True,
    ) -> Union[Tuple[torch.Tensor, ...], Transformer2DModelOutput]:
        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."
                )

        # ensure attention_mask is a bias, and give it a singleton query_tokens dimension.
        #   we may have done this conversion already, e.g. if we came here via UNet2DConditionModel#forward.
        #   we can tell by counting dims; if ndim == 2: it's a mask rather than a bias.
        # expects mask of shape:
        #   [batch, key_tokens]
        # adds singleton query_tokens dimension:
        #   [batch,                    1, key_tokens]
        # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
        #   [batch,  heads, query_tokens, key_tokens] (e.g. torch sdp attn)
        #   [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
        if attention_mask is not None and attention_mask.ndim == 2:
            # assume that mask is expressed as:
            #   (1 = keep,      0 = discard)
            # convert mask into a bias that can be added to attention scores:
            #       (keep = +0,     discard = -10000.0)
            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)

        # 1. Input
        batch_size, num_channels, height, width = hidden_states.shape
        p = self.config.patch_size
        post_patch_height, post_patch_width = height // p, width // p

        hidden_states = self.patch_embed(hidden_states)

        if guidance is not None:
            timestep, embedded_timestep = self.time_embed(
                timestep, guidance=guidance, hidden_dtype=hidden_states.dtype
            )
        else:
            timestep, embedded_timestep = self.time_embed(
                timestep, batch_size=batch_size, hidden_dtype=hidden_states.dtype
            )

        encoder_hidden_states = self.caption_projection(encoder_hidden_states)
        encoder_hidden_states = encoder_hidden_states.view(batch_size, -1, hidden_states.shape[-1])

        encoder_hidden_states = self.caption_norm(encoder_hidden_states)

        # 2. Transformer blocks
        if torch.is_grad_enabled() and self.gradient_checkpointing:
            for block in self.transformer_blocks:
                hidden_states = self._gradient_checkpointing_func(
                    block,
                    hidden_states,
                    attention_mask,
                    encoder_hidden_states,
                    encoder_attention_mask,
                    timestep,
                    post_patch_height,
                    post_patch_width,
                )

        else:
            for block in self.transformer_blocks:
                hidden_states = block(
                    hidden_states,
                    attention_mask,
                    encoder_hidden_states,
                    encoder_attention_mask,
                    timestep,
                    post_patch_height,
                    post_patch_width,
                )

        # 3. Normalization
        hidden_states = self.norm_out(hidden_states, embedded_timestep, self.scale_shift_table)

        hidden_states = self.proj_out(hidden_states)

        # 5. Unpatchify
        hidden_states = hidden_states.reshape(
            batch_size, post_patch_height, post_patch_width, self.config.patch_size, self.config.patch_size, -1
        )
        hidden_states = hidden_states.permute(0, 5, 1, 3, 2, 4)
        output = hidden_states.reshape(batch_size, -1, post_patch_height * p, post_patch_width * p)

        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)