src/model/dit.py

# Copyright 2024 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, Dict, Optional, Tuple, Union

import torch
from torch import nn

from diffusers.configuration_utils import ConfigMixin, register_to_config
from diffusers.loaders import PeftAdapterMixin
from diffusers.pipelines.pipeline_utils import DiffusionPipeline
from diffusers.utils import USE_PEFT_BACKEND, is_torch_version, logging, scale_lora_layers, unscale_lora_layers
from diffusers.utils import is_accelerate_version, is_accelerate_available
from diffusers.utils.torch_utils import maybe_allow_in_graph
from diffusers.models.attention import Attention, FeedForward
from diffusers.models.attention_processor import AttentionProcessor, CogVideoXAttnProcessor2_0, FusedCogVideoXAttnProcessor2_0
from diffusers.models.embeddings import TimestepEmbedding, Timesteps
from diffusers.models.modeling_utils import ModelMixin
from diffusers.models.normalization import AdaLayerNorm, CogVideoXLayerNormZero

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

def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
    """
    This function generates 1D positional embeddings from a grid.

    Args:
        embed_dim (`int`): The embedding dimension `D`
        pos (`torch.Tensor`): 1D tensor of positions with shape `(M,)`

    Returns:
        `torch.Tensor`: Sinusoidal positional embeddings of shape `(M, D)`.
    """
    if embed_dim % 2 != 0:
        raise ValueError("embed_dim must be divisible by 2")

    omega = torch.arange(embed_dim // 2, device=pos.device, dtype=torch.float64)
    omega /= embed_dim / 2.0
    omega = 1.0 / 10000**omega  # (D/2,)

    pos = pos.reshape(-1)  # (M,)
    out = torch.outer(pos, omega)  # (M, D/2), outer product

    emb_sin = torch.sin(out)  # (M, D/2)
    emb_cos = torch.cos(out)  # (M, D/2)

    emb = torch.concat([emb_sin, emb_cos], dim=1)  # (M, D)
    return emb

def get_3d_sincos_pos_embed(
    embed_dim: int,
    spatial_size: Union[int, Tuple[int, int]],
    temporal_size: int,
    spatial_interpolation_scale: float = 1.0,
    temporal_interpolation_scale: float = 1.0,
    device: Optional[torch.device] = None,
    output_type: str = "np",
) -> torch.Tensor:
    r"""
    Creates 3D sinusoidal positional embeddings.

    Args:
        embed_dim (`int`):
            The embedding dimension of inputs. It must be divisible by 16.
        spatial_size (`int` or `Tuple[int, int]`):
            The spatial dimension of positional embeddings. If an integer is provided, the same size is applied to both
            spatial dimensions (height and width).
        temporal_size (`int`):
            The temporal dimension of postional embeddings (number of frames).
        spatial_interpolation_scale (`float`, defaults to 1.0):
            Scale factor for spatial grid interpolation.
        temporal_interpolation_scale (`float`, defaults to 1.0):
            Scale factor for temporal grid interpolation.

    Returns:
        `torch.Tensor`:
            The 3D sinusoidal positional embeddings of shape `[temporal_size, spatial_size[0] * spatial_size[1],
            embed_dim]`.
    """
    
    if embed_dim % 4 != 0:
        raise ValueError("`embed_dim` must be divisible by 4")

    embed_dim_spatial = 3 * embed_dim // 4
    embed_dim_temporal = embed_dim // 4

    # 1. Spatial
    grid_pc = torch.arange(spatial_size, device=device, dtype=torch.float32) / spatial_interpolation_scale
    pos_embed_spatial = get_1d_sincos_pos_embed_from_grid(embed_dim_spatial, grid_pc)

    # 2. Temporal
    grid_t = torch.arange(temporal_size, device=device, dtype=torch.float32) / temporal_interpolation_scale
    pos_embed_temporal = get_1d_sincos_pos_embed_from_grid(embed_dim_temporal, grid_t)

    # 3. Concat
    pos_embed_spatial = pos_embed_spatial[None, :, :]
    pos_embed_spatial = pos_embed_spatial.repeat_interleave(temporal_size, dim=0)  # [T, H*W, D // 4 * 3]
    pos_embed_temporal = pos_embed_temporal[:, None, :]
    pos_embed_temporal = pos_embed_temporal.repeat_interleave(spatial_size, dim=1)  # [T, H*W, D // 4]
    pos_embed = torch.concat([pos_embed_temporal, pos_embed_spatial], dim=-1)  # [T, H*W, D]
    return pos_embed

@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: Optional[int] = 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: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
    ) -> torch.Tensor:
        text_seq_length = encoder_hidden_states.size(1)

        # 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,
        )

        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, ConfigMixin, PeftAdapterMixin):
    """
    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.
    """

    _supports_gradient_checkpointing = True

    @register_to_config
    def __init__(
        self,
        num_attention_heads: int = 8,
        attention_head_dim: int = 64,
        in_channels: int = 3,
        out_channels: Optional[int] = 3,
        flip_sin_to_cos: bool = True,
        freq_shift: int = 0,
        time_embed_dim: int = 512,
        ofs_embed_dim: Optional[int] = None,
        text_embed_dim: int = 4096,
        num_layers: int = 8,
        dropout: float = 0.0,
        attention_bias: bool = True,
        sample_points: int = 2048,
        sample_frames: int = 48,
        patch_size: int = 1,
        patch_size_t: Optional[int] = 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_positional_embeddings: bool = True,
        use_learned_positional_embeddings: bool = False,
        patch_bias: bool = True,
        cond_seq_length: int = 4
    ):
        super().__init__()
        inner_dim = num_attention_heads * attention_head_dim

        if use_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."
            )

        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
        
        if use_positional_embeddings or use_learned_positional_embeddings:
            self.embed_dim = num_attention_heads * attention_head_dim
            self.cond_seq_length = cond_seq_length
            persistent = use_learned_positional_embeddings
            pos_embedding = self._get_positional_embeddings(sample_points, sample_frames)
            self.register_buffer("pos_embedding", pos_embedding, persistent=persistent)

    def _get_positional_embeddings(self, points: int, frames: int, device: Optional[torch.device] = None) -> torch.Tensor:
        pos_embedding = get_3d_sincos_pos_embed(
            self.embed_dim,
            points,
            frames,
            device=device,
            output_type="pt",
        )
        pos_embedding = pos_embedding.flatten(0, 1)
        joint_pos_embedding = pos_embedding.new_zeros(
            1, self.cond_seq_length + points * frames, self.embed_dim, requires_grad=False
        )
        joint_pos_embedding.data[:, self.cond_seq_length:].copy_(pos_embedding)
        return joint_pos_embedding
    
    def _set_gradient_checkpointing(self, module, value=False):
        self.gradient_checkpointing = value

    @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)

    # 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.

        <Tip warning={true}>

        This API is 🧪 experimental.

        </Tip>
        """
        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.

        <Tip warning={true}>

        This API is 🧪 experimental.

        </Tip>

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

    def forward(
        self,
        full_seq: torch.Tensor, # [batch_size]
        timestep: Union[int, float, torch.LongTensor],
        timestep_cond: Optional[torch.Tensor] = None,
        ofs: Optional[Union[int, float, torch.LongTensor]] = None,
        image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
        attention_kwargs: Optional[Dict[str, Any]] = None,
        return_dict: bool = True,
    ):
        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."
                )
        
        # 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=full_seq.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=full_seq.dtype)
            ofs_emb = self.ofs_embedding(ofs_emb)
            emb = emb + ofs_emb

        # 2. Patch embedding
        pos_embedding = self.pos_embedding
        pos_embedding = pos_embedding.to(dtype=full_seq.dtype)
        hidden_states = full_seq + pos_embedding
        
        hidden_states = self.embedding_dropout(hidden_states)
        encoder_hidden_states = hidden_states[:, :self.cond_seq_length]
        hidden_states = hidden_states[:, self.cond_seq_length:]

        # 3. Transformer blocks
        for i, block in enumerate(self.transformer_blocks):
            if torch.is_grad_enabled() and self.gradient_checkpointing:

                def create_custom_forward(module):
                    def custom_forward(*inputs):
                        return module(*inputs)
                    return custom_forward

                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
                hidden_states, encoder_hidden_states = torch.utils.checkpoint.checkpoint(
                    create_custom_forward(block),
                    hidden_states,
                    encoder_hidden_states,
                    emb,
                    image_rotary_emb,
                    **ckpt_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,
                )

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

        if USE_PEFT_BACKEND:
            # remove `lora_scale` from each PEFT layer
            unscale_lora_layers(self, lora_scale)

        return output