diffusers/models/transformers/transformer_cogview3plus.py

# Copyright 2025 The CogView team, Tsinghua University & ZhipuAI and 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.


import torch
import torch.nn as nn

from ...configuration_utils import ConfigMixin, register_to_config
from ...utils import logging
from ..attention import AttentionMixin, FeedForward
from ..attention_processor import Attention, CogVideoXAttnProcessor2_0
from ..embeddings import CogView3CombinedTimestepSizeEmbeddings, CogView3PlusPatchEmbed
from ..modeling_outputs import Transformer2DModelOutput
from ..modeling_utils import ModelMixin
from ..normalization import AdaLayerNormContinuous, CogView3PlusAdaLayerNormZeroTextImage


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


class CogView3PlusTransformerBlock(nn.Module):
    r"""
    Transformer block used in [CogView](https://github.com/THUDM/CogView3) model.

    Args:
        dim (`int`):
            The number of channels in the input and output.
        num_attention_heads (`int`):
            The number of heads to use for multi-head attention.
        attention_head_dim (`int`):
            The number of channels in each head.
        time_embed_dim (`int`):
            The number of channels in timestep embedding.
    """

    def __init__(
        self,
        dim: int = 2560,
        num_attention_heads: int = 64,
        attention_head_dim: int = 40,
        time_embed_dim: int = 512,
    ):
        super().__init__()

        self.norm1 = CogView3PlusAdaLayerNormZeroTextImage(embedding_dim=time_embed_dim, dim=dim)

        self.attn1 = Attention(
            query_dim=dim,
            heads=num_attention_heads,
            dim_head=attention_head_dim,
            out_dim=dim,
            bias=True,
            qk_norm="layer_norm",
            elementwise_affine=False,
            eps=1e-6,
            processor=CogVideoXAttnProcessor2_0(),
        )

        self.norm2 = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-5)
        self.norm2_context = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-5)

        self.ff = FeedForward(dim=dim, dim_out=dim, activation_fn="gelu-approximate")

    def forward(
        self,
        hidden_states: torch.Tensor,
        encoder_hidden_states: torch.Tensor,
        emb: torch.Tensor,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        text_seq_length = encoder_hidden_states.size(1)

        # norm & modulate
        (
            norm_hidden_states,
            gate_msa,
            shift_mlp,
            scale_mlp,
            gate_mlp,
            norm_encoder_hidden_states,
            c_gate_msa,
            c_shift_mlp,
            c_scale_mlp,
            c_gate_mlp,
        ) = self.norm1(hidden_states, encoder_hidden_states, emb)

        # attention
        attn_hidden_states, attn_encoder_hidden_states = self.attn1(
            hidden_states=norm_hidden_states, encoder_hidden_states=norm_encoder_hidden_states
        )

        hidden_states = hidden_states + gate_msa.unsqueeze(1) * attn_hidden_states
        encoder_hidden_states = encoder_hidden_states + c_gate_msa.unsqueeze(1) * attn_encoder_hidden_states

        # norm & modulate
        norm_hidden_states = self.norm2(hidden_states)
        norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]

        norm_encoder_hidden_states = self.norm2_context(encoder_hidden_states)
        norm_encoder_hidden_states = norm_encoder_hidden_states * (1 + c_scale_mlp[:, None]) + c_shift_mlp[:, None]

        # feed-forward
        norm_hidden_states = torch.cat([norm_encoder_hidden_states, norm_hidden_states], dim=1)
        ff_output = self.ff(norm_hidden_states)

        hidden_states = hidden_states + gate_mlp.unsqueeze(1) * ff_output[:, text_seq_length:]
        encoder_hidden_states = encoder_hidden_states + c_gate_mlp.unsqueeze(1) * ff_output[:, :text_seq_length]

        if hidden_states.dtype == torch.float16:
            hidden_states = hidden_states.clip(-65504, 65504)
        if encoder_hidden_states.dtype == torch.float16:
            encoder_hidden_states = encoder_hidden_states.clip(-65504, 65504)
        return hidden_states, encoder_hidden_states


class CogView3PlusTransformer2DModel(ModelMixin, AttentionMixin, ConfigMixin):
    r"""
    The Transformer model introduced in [CogView3: Finer and Faster Text-to-Image Generation via Relay
    Diffusion](https://huggingface.co/papers/2403.05121).

    Args:
        patch_size (`int`, defaults to `2`):
            The size of the patches to use in the patch embedding layer.
        in_channels (`int`, defaults to `16`):
            The number of channels in the input.
        num_layers (`int`, defaults to `30`):
            The number of layers of Transformer blocks to use.
        attention_head_dim (`int`, defaults to `40`):
            The number of channels in each head.
        num_attention_heads (`int`, defaults to `64`):
            The number of heads to use for multi-head attention.
        out_channels (`int`, defaults to `16`):
            The number of channels in the output.
        text_embed_dim (`int`, defaults to `4096`):
            Input dimension of text embeddings from the text encoder.
        time_embed_dim (`int`, defaults to `512`):
            Output dimension of timestep embeddings.
        condition_dim (`int`, defaults to `256`):
            The embedding dimension of the input SDXL-style resolution conditions (original_size, target_size,
            crop_coords).
        pos_embed_max_size (`int`, defaults to `128`):
            The maximum resolution of the positional embeddings, from which slices of shape `H x W` are taken and added
            to input patched latents, where `H` and `W` are the latent height and width respectively. A value of 128
            means that the maximum supported height and width for image generation is `128 * vae_scale_factor *
            patch_size => 128 * 8 * 2 => 2048`.
        sample_size (`int`, defaults to `128`):
            The base resolution of input latents. If height/width is not provided during generation, this value is used
            to determine the resolution as `sample_size * vae_scale_factor => 128 * 8 => 1024`
    """

    _supports_gradient_checkpointing = True
    _skip_layerwise_casting_patterns = ["patch_embed", "norm"]
    _no_split_modules = ["CogView3PlusTransformerBlock", "CogView3PlusPatchEmbed"]

    @register_to_config
    def __init__(
        self,
        patch_size: int = 2,
        in_channels: int = 16,
        num_layers: int = 30,
        attention_head_dim: int = 40,
        num_attention_heads: int = 64,
        out_channels: int = 16,
        text_embed_dim: int = 4096,
        time_embed_dim: int = 512,
        condition_dim: int = 256,
        pos_embed_max_size: int = 128,
        sample_size: int = 128,
    ):
        super().__init__()
        self.out_channels = out_channels
        self.inner_dim = num_attention_heads * attention_head_dim

        # CogView3 uses 3 additional SDXL-like conditions - original_size, target_size, crop_coords
        # Each of these are sincos embeddings of shape 2 * condition_dim
        self.pooled_projection_dim = 3 * 2 * condition_dim

        self.patch_embed = CogView3PlusPatchEmbed(
            in_channels=in_channels,
            hidden_size=self.inner_dim,
            patch_size=patch_size,
            text_hidden_size=text_embed_dim,
            pos_embed_max_size=pos_embed_max_size,
        )

        self.time_condition_embed = CogView3CombinedTimestepSizeEmbeddings(
            embedding_dim=time_embed_dim,
            condition_dim=condition_dim,
            pooled_projection_dim=self.pooled_projection_dim,
            timesteps_dim=self.inner_dim,
        )

        self.transformer_blocks = nn.ModuleList(
            [
                CogView3PlusTransformerBlock(
                    dim=self.inner_dim,
                    num_attention_heads=num_attention_heads,
                    attention_head_dim=attention_head_dim,
                    time_embed_dim=time_embed_dim,
                )
                for _ in range(num_layers)
            ]
        )

        self.norm_out = AdaLayerNormContinuous(
            embedding_dim=self.inner_dim,
            conditioning_embedding_dim=time_embed_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

    def forward(
        self,
        hidden_states: torch.Tensor,
        encoder_hidden_states: torch.Tensor,
        timestep: torch.LongTensor,
        original_size: torch.Tensor,
        target_size: torch.Tensor,
        crop_coords: torch.Tensor,
        return_dict: bool = True,
    ) -> tuple[torch.Tensor] | Transformer2DModelOutput:
        """
        The [`CogView3PlusTransformer2DModel`] forward method.

        Args:
            hidden_states (`torch.Tensor`):
                Input `hidden_states` of shape `(batch size, channel, height, width)`.
            encoder_hidden_states (`torch.Tensor`):
                Conditional embeddings (embeddings computed from the input conditions such as prompts) of shape
                `(batch_size, sequence_len, text_embed_dim)`
            timestep (`torch.LongTensor`):
                Used to indicate denoising step.
            original_size (`torch.Tensor`):
                CogView3 uses SDXL-like micro-conditioning for original image size as explained in section 2.2 of
                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
            target_size (`torch.Tensor`):
                CogView3 uses SDXL-like micro-conditioning for target image size as explained in section 2.2 of
                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
            crop_coords (`torch.Tensor`):
                CogView3 uses SDXL-like micro-conditioning for crop coordinates as explained in section 2.2 of
                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
            return_dict (`bool`, *optional*, defaults to `True`):
                Whether or not to return a [`~models.transformer_2d.Transformer2DModelOutput`] instead of a plain
                tuple.

        Returns:
            `torch.Tensor` or [`~models.transformer_2d.Transformer2DModelOutput`]:
                The denoised latents using provided inputs as conditioning.
        """
        height, width = hidden_states.shape[-2:]
        text_seq_length = encoder_hidden_states.shape[1]

        hidden_states = self.patch_embed(
            hidden_states, encoder_hidden_states
        )  # takes care of adding positional embeddings too.
        emb = self.time_condition_embed(timestep, original_size, target_size, crop_coords, hidden_states.dtype)

        encoder_hidden_states = hidden_states[:, :text_seq_length]
        hidden_states = hidden_states[:, text_seq_length:]

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

        hidden_states = self.norm_out(hidden_states, emb)
        hidden_states = self.proj_out(hidden_states)  # (batch_size, height*width, patch_size*patch_size*out_channels)

        # 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, self.out_channels, patch_size, patch_size)
        )
        hidden_states = torch.einsum("nhwcpq->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)