diffusers/models/transformers/cogvideox_transformer_3d.py

# Copyright 2025 The CogVideoX 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.

from typing import Any

import torch
from torch import nn

from ...configuration_utils import ConfigMixin, register_to_config
from ...loaders import PeftAdapterMixin
from ...utils import apply_lora_scale, logging
from ...utils.torch_utils import maybe_allow_in_graph
from ..attention import Attention, AttentionMixin, FeedForward
from ..attention_processor import CogVideoXAttnProcessor2_0, FusedCogVideoXAttnProcessor2_0
from ..cache_utils import CacheMixin
from ..embeddings import CogVideoXPatchEmbed, TimestepEmbedding, Timesteps
from ..modeling_outputs import Transformer2DModelOutput
from ..modeling_utils import ModelMixin
from ..normalization import AdaLayerNorm, CogVideoXLayerNormZero


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


@maybe_allow_in_graph
class CogVideoXBlock(nn.Module):
    r"""
    Transformer block used in [CogVideoX](https://github.com/THUDM/CogVideo) model.

    Parameters:
        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.
        dropout (`float`, defaults to `0.0`):
            The dropout probability to use.
        activation_fn (`str`, defaults to `"gelu-approximate"`):
            Activation function to be used in feed-forward.
        attention_bias (`bool`, defaults to `False`):
            Whether or not to use bias in attention projection layers.
        qk_norm (`bool`, defaults to `True`):
            Whether or not to use normalization after query and key projections in Attention.
        norm_elementwise_affine (`bool`, defaults to `True`):
            Whether to use learnable elementwise affine parameters for normalization.
        norm_eps (`float`, defaults to `1e-5`):
            Epsilon value for normalization layers.
        final_dropout (`bool` defaults to `False`):
            Whether to apply a final dropout after the last feed-forward layer.
        ff_inner_dim (`int`, *optional*, defaults to `None`):
            Custom hidden dimension of Feed-forward layer. If not provided, `4 * dim` is used.
        ff_bias (`bool`, defaults to `True`):
            Whether or not to use bias in Feed-forward layer.
        attention_out_bias (`bool`, defaults to `True`):
            Whether or not to use bias in Attention output projection layer.
    """

    def __init__(
        self,
        dim: int,
        num_attention_heads: int,
        attention_head_dim: int,
        time_embed_dim: int,
        dropout: float = 0.0,
        activation_fn: str = "gelu-approximate",
        attention_bias: bool = False,
        qk_norm: bool = True,
        norm_elementwise_affine: bool = True,
        norm_eps: float = 1e-5,
        final_dropout: bool = True,
        ff_inner_dim: int | None = None,
        ff_bias: bool = True,
        attention_out_bias: bool = True,
    ):
        super().__init__()

        # 1. Self Attention
        self.norm1 = CogVideoXLayerNormZero(time_embed_dim, dim, norm_elementwise_affine, norm_eps, bias=True)

        self.attn1 = Attention(
            query_dim=dim,
            dim_head=attention_head_dim,
            heads=num_attention_heads,
            qk_norm="layer_norm" if qk_norm else None,
            eps=1e-6,
            bias=attention_bias,
            out_bias=attention_out_bias,
            processor=CogVideoXAttnProcessor2_0(),
        )

        # 2. Feed Forward
        self.norm2 = CogVideoXLayerNormZero(time_embed_dim, dim, norm_elementwise_affine, norm_eps, bias=True)

        self.ff = FeedForward(
            dim,
            dropout=dropout,
            activation_fn=activation_fn,
            final_dropout=final_dropout,
            inner_dim=ff_inner_dim,
            bias=ff_bias,
        )

    def forward(
        self,
        hidden_states: torch.Tensor,
        encoder_hidden_states: torch.Tensor,
        temb: torch.Tensor,
        image_rotary_emb: tuple[torch.Tensor, torch.Tensor] | None = None,
        attention_kwargs: dict[str, Any] | None = None,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        text_seq_length = encoder_hidden_states.size(1)
        attention_kwargs = attention_kwargs or {}

        # norm & modulate
        norm_hidden_states, norm_encoder_hidden_states, gate_msa, enc_gate_msa = self.norm1(
            hidden_states, encoder_hidden_states, temb
        )

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

        hidden_states = hidden_states + gate_msa * attn_hidden_states
        encoder_hidden_states = encoder_hidden_states + enc_gate_msa * attn_encoder_hidden_states

        # norm & modulate
        norm_hidden_states, norm_encoder_hidden_states, gate_ff, enc_gate_ff = self.norm2(
            hidden_states, encoder_hidden_states, temb
        )

        # 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_ff * ff_output[:, text_seq_length:]
        encoder_hidden_states = encoder_hidden_states + enc_gate_ff * ff_output[:, :text_seq_length]

        return hidden_states, encoder_hidden_states


class CogVideoXTransformer3DModel(ModelMixin, AttentionMixin, ConfigMixin, PeftAdapterMixin, CacheMixin):
    """
    A Transformer model for video-like data in [CogVideoX](https://github.com/THUDM/CogVideo).

    Parameters:
        num_attention_heads (`int`, defaults to `30`):
            The number of heads to use for multi-head attention.
        attention_head_dim (`int`, defaults to `64`):
            The number of channels in each head.
        in_channels (`int`, defaults to `16`):
            The number of channels in the input.
        out_channels (`int`, *optional*, defaults to `16`):
            The number of channels in the output.
        flip_sin_to_cos (`bool`, defaults to `True`):
            Whether to flip the sin to cos in the time embedding.
        time_embed_dim (`int`, defaults to `512`):
            Output dimension of timestep embeddings.
        ofs_embed_dim (`int`, defaults to `512`):
            Output dimension of "ofs" embeddings used in CogVideoX-5b-I2B in version 1.5
        text_embed_dim (`int`, defaults to `4096`):
            Input dimension of text embeddings from the text encoder.
        num_layers (`int`, defaults to `30`):
            The number of layers of Transformer blocks to use.
        dropout (`float`, defaults to `0.0`):
            The dropout probability to use.
        attention_bias (`bool`, defaults to `True`):
            Whether to use bias in the attention projection layers.
        sample_width (`int`, defaults to `90`):
            The width of the input latents.
        sample_height (`int`, defaults to `60`):
            The height of the input latents.
        sample_frames (`int`, defaults to `49`):
            The number of frames in the input latents. Note that this parameter was incorrectly initialized to 49
            instead of 13 because CogVideoX processed 13 latent frames at once in its default and recommended settings,
            but cannot be changed to the correct value to ensure backwards compatibility. To create a transformer with
            K latent frames, the correct value to pass here would be: ((K - 1) * temporal_compression_ratio + 1).
        patch_size (`int`, defaults to `2`):
            The size of the patches to use in the patch embedding layer.
        temporal_compression_ratio (`int`, defaults to `4`):
            The compression ratio across the temporal dimension. See documentation for `sample_frames`.
        max_text_seq_length (`int`, defaults to `226`):
            The maximum sequence length of the input text embeddings.
        activation_fn (`str`, defaults to `"gelu-approximate"`):
            Activation function to use in feed-forward.
        timestep_activation_fn (`str`, defaults to `"silu"`):
            Activation function to use when generating the timestep embeddings.
        norm_elementwise_affine (`bool`, defaults to `True`):
            Whether to use elementwise affine in normalization layers.
        norm_eps (`float`, defaults to `1e-5`):
            The epsilon value to use in normalization layers.
        spatial_interpolation_scale (`float`, defaults to `1.875`):
            Scaling factor to apply in 3D positional embeddings across spatial dimensions.
        temporal_interpolation_scale (`float`, defaults to `1.0`):
            Scaling factor to apply in 3D positional embeddings across temporal dimensions.
    """

    _skip_layerwise_casting_patterns = ["patch_embed", "norm"]
    _supports_gradient_checkpointing = True
    _no_split_modules = ["CogVideoXBlock", "CogVideoXPatchEmbed"]

    @register_to_config
    def __init__(
        self,
        num_attention_heads: int = 30,
        attention_head_dim: int = 64,
        in_channels: int = 16,
        out_channels: int | None = 16,
        flip_sin_to_cos: bool = True,
        freq_shift: int = 0,
        time_embed_dim: int = 512,
        ofs_embed_dim: int | None = None,
        text_embed_dim: int = 4096,
        num_layers: int = 30,
        dropout: float = 0.0,
        attention_bias: bool = True,
        sample_width: int = 90,
        sample_height: int = 60,
        sample_frames: int = 49,
        patch_size: int = 2,
        patch_size_t: int | None = None,
        temporal_compression_ratio: int = 4,
        max_text_seq_length: int = 226,
        activation_fn: str = "gelu-approximate",
        timestep_activation_fn: str = "silu",
        norm_elementwise_affine: bool = True,
        norm_eps: float = 1e-5,
        spatial_interpolation_scale: float = 1.875,
        temporal_interpolation_scale: float = 1.0,
        use_rotary_positional_embeddings: bool = False,
        use_learned_positional_embeddings: bool = False,
        patch_bias: bool = True,
    ):
        super().__init__()
        inner_dim = num_attention_heads * attention_head_dim

        if not use_rotary_positional_embeddings and use_learned_positional_embeddings:
            raise ValueError(
                "There are no CogVideoX checkpoints available with disable rotary embeddings and learned positional "
                "embeddings. If you're using a custom model and/or believe this should be supported, please open an "
                "issue at https://github.com/huggingface/diffusers/issues."
            )

        # 1. Patch embedding
        self.patch_embed = CogVideoXPatchEmbed(
            patch_size=patch_size,
            patch_size_t=patch_size_t,
            in_channels=in_channels,
            embed_dim=inner_dim,
            text_embed_dim=text_embed_dim,
            bias=patch_bias,
            sample_width=sample_width,
            sample_height=sample_height,
            sample_frames=sample_frames,
            temporal_compression_ratio=temporal_compression_ratio,
            max_text_seq_length=max_text_seq_length,
            spatial_interpolation_scale=spatial_interpolation_scale,
            temporal_interpolation_scale=temporal_interpolation_scale,
            use_positional_embeddings=not use_rotary_positional_embeddings,
            use_learned_positional_embeddings=use_learned_positional_embeddings,
        )
        self.embedding_dropout = nn.Dropout(dropout)

        # 2. Time embeddings and ofs embedding(Only CogVideoX1.5-5B I2V have)

        self.time_proj = Timesteps(inner_dim, flip_sin_to_cos, freq_shift)
        self.time_embedding = TimestepEmbedding(inner_dim, time_embed_dim, timestep_activation_fn)

        self.ofs_proj = None
        self.ofs_embedding = None
        if ofs_embed_dim:
            self.ofs_proj = Timesteps(ofs_embed_dim, flip_sin_to_cos, freq_shift)
            self.ofs_embedding = TimestepEmbedding(
                ofs_embed_dim, ofs_embed_dim, timestep_activation_fn
            )  # same as time embeddings, for ofs

        # 3. Define spatio-temporal transformers blocks
        self.transformer_blocks = nn.ModuleList(
            [
                CogVideoXBlock(
                    dim=inner_dim,
                    num_attention_heads=num_attention_heads,
                    attention_head_dim=attention_head_dim,
                    time_embed_dim=time_embed_dim,
                    dropout=dropout,
                    activation_fn=activation_fn,
                    attention_bias=attention_bias,
                    norm_elementwise_affine=norm_elementwise_affine,
                    norm_eps=norm_eps,
                )
                for _ in range(num_layers)
            ]
        )
        self.norm_final = nn.LayerNorm(inner_dim, norm_eps, norm_elementwise_affine)

        # 4. Output blocks
        self.norm_out = AdaLayerNorm(
            embedding_dim=time_embed_dim,
            output_dim=2 * inner_dim,
            norm_elementwise_affine=norm_elementwise_affine,
            norm_eps=norm_eps,
            chunk_dim=1,
        )

        if patch_size_t is None:
            # For CogVideox 1.0
            output_dim = patch_size * patch_size * out_channels
        else:
            # For CogVideoX 1.5
            output_dim = patch_size * patch_size * patch_size_t * out_channels

        self.proj_out = nn.Linear(inner_dim, output_dim)

        self.gradient_checkpointing = False

    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.fuse_qkv_projections with FusedAttnProcessor2_0->FusedCogVideoXAttnProcessor2_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(FusedCogVideoXAttnProcessor2_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("attention_kwargs")
    def forward(
        self,
        hidden_states: torch.Tensor,
        encoder_hidden_states: torch.Tensor,
        timestep: int | float | torch.LongTensor,
        timestep_cond: torch.Tensor | None = None,
        ofs: int | float | torch.LongTensor | None = None,
        image_rotary_emb: tuple[torch.Tensor, torch.Tensor] | None = None,
        attention_kwargs: dict[str, Any] | None = None,
        return_dict: bool = True,
    ) -> tuple[torch.Tensor] | Transformer2DModelOutput:
        """
        The [`CogVideoXTransformer3DModel`] forward method.

        Args:
            hidden_states (`torch.Tensor` of shape `(batch_size, num_frames, channels, 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.
            timestep (`torch.LongTensor`):
                Used to indicate denoising step.
            timestep_cond (`torch.Tensor`, *optional*):
                Conditional embeddings for timestep. If provided, the embeddings will be summed with the samples passed
                through the `self.time_embedding` layer to obtain the final timestep embeddings.
            ofs (`torch.Tensor`, *optional*):
                Offset embeddings used in CogVideoX-5b-I2V.
            image_rotary_emb (`tuple` of `torch.Tensor`, *optional*):
                Pre-computed rotary positional embeddings.
            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.
        """
        batch_size, num_frames, channels, height, width = hidden_states.shape

        # 1. Time embedding
        timesteps = timestep
        t_emb = self.time_proj(timesteps)

        # timesteps does not contain any weights and will always return f32 tensors
        # but time_embedding might actually be running in fp16. so we need to cast here.
        # there might be better ways to encapsulate this.
        t_emb = t_emb.to(dtype=hidden_states.dtype)
        emb = self.time_embedding(t_emb, timestep_cond)

        if self.ofs_embedding is not None:
            ofs_emb = self.ofs_proj(ofs)
            ofs_emb = ofs_emb.to(dtype=hidden_states.dtype)
            ofs_emb = self.ofs_embedding(ofs_emb)
            emb = emb + ofs_emb

        # 2. Patch embedding
        hidden_states = self.patch_embed(encoder_hidden_states, hidden_states)
        hidden_states = self.embedding_dropout(hidden_states)

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

        # 3. Transformer blocks
        for i, 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,
                    image_rotary_emb,
                    attention_kwargs,
                )
            else:
                hidden_states, encoder_hidden_states = block(
                    hidden_states=hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
                    temb=emb,
                    image_rotary_emb=image_rotary_emb,
                    attention_kwargs=attention_kwargs,
                )

        hidden_states = self.norm_final(hidden_states)

        # 4. Final block
        hidden_states = self.norm_out(hidden_states, temb=emb)
        hidden_states = self.proj_out(hidden_states)

        # 5. Unpatchify
        p = self.config.patch_size
        p_t = self.config.patch_size_t

        if p_t is None:
            output = hidden_states.reshape(batch_size, num_frames, height // p, width // p, -1, p, p)
            output = output.permute(0, 1, 4, 2, 5, 3, 6).flatten(5, 6).flatten(3, 4)
        else:
            output = hidden_states.reshape(
                batch_size, (num_frames + p_t - 1) // p_t, height // p, width // p, -1, p_t, p, p
            )
            output = output.permute(0, 1, 5, 4, 2, 6, 3, 7).flatten(6, 7).flatten(4, 5).flatten(1, 2)

        if not return_dict:
            return (output,)
        return Transformer2DModelOutput(sample=output)