consisti2v/models/videoldm_transformer_blocks.py

# Modified from https://github.com/huggingface/diffusers/blob/v0.21.0/src/diffusers/models/transformer_2d.py
from dataclasses import dataclass
from typing import Any, Dict, Optional

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

from einops import rearrange, repeat

from diffusers.configuration_utils import ConfigMixin, register_to_config
from diffusers.models.embeddings import ImagePositionalEmbeddings
from diffusers.utils import BaseOutput, deprecate
from diffusers.models.attention import AdaLayerNorm, AdaLayerNormZero, FeedForward, GatedSelfAttentionDense
from diffusers.models.embeddings import PatchEmbed
from diffusers.models.lora import LoRACompatibleConv, LoRACompatibleLinear
from diffusers.models.modeling_utils import ModelMixin
from diffusers.models.transformer_2d import Transformer2DModelOutput
from diffusers.utils.torch_utils import maybe_allow_in_graph
from diffusers.models.attention_processor import Attention
from diffusers.models.lora import LoRACompatibleLinear

from .videoldm_attention import ConditionalAttention, TemporalConditionalAttention


class Transformer2DConditionModel(ModelMixin, ConfigMixin):
    @register_to_config
    def __init__(
        self,
        num_attention_heads: int = 16,
        attention_head_dim: int = 88,
        in_channels: Optional[int] = None,
        out_channels: Optional[int] = None,
        num_layers: int = 1,
        dropout: float = 0.0,
        norm_num_groups: int = 32,
        cross_attention_dim: Optional[int] = None,
        attention_bias: bool = False,
        sample_size: Optional[int] = None,
        num_vector_embeds: Optional[int] = None,
        patch_size: Optional[int] = None,
        activation_fn: str = "geglu",
        num_embeds_ada_norm: Optional[int] = None,
        use_linear_projection: bool = False,
        only_cross_attention: bool = False,
        double_self_attention: bool = False,
        upcast_attention: bool = False,
        norm_type: str = "layer_norm",
        norm_elementwise_affine: bool = True,
        attention_type: str = "default",
        # additional
        n_frames: int = 8,
        is_temporal: bool = False,
        augment_temporal_attention: bool = False,
        rotary_emb=False,
    ):
        super().__init__()
        self.use_linear_projection = use_linear_projection
        self.num_attention_heads = num_attention_heads
        self.attention_head_dim = attention_head_dim
        inner_dim = num_attention_heads * attention_head_dim

        # 1. Transformer2DModel can process both standard continuous images of shape `(batch_size, num_channels, width, height)` as well as quantized image embeddings of shape `(batch_size, num_image_vectors)`
        # Define whether input is continuous or discrete depending on configuration
        self.is_input_continuous = (in_channels is not None) and (patch_size is None)
        self.is_input_vectorized = num_vector_embeds is not None
        self.is_input_patches = in_channels is not None and patch_size is not None

        if norm_type == "layer_norm" and num_embeds_ada_norm is not None:
            deprecation_message = (
                f"The configuration file of this model: {self.__class__} is outdated. `norm_type` is either not set or"
                " incorrectly set to `'layer_norm'`.Make sure to set `norm_type` to `'ada_norm'` in the config."
                " Please make sure to update the config accordingly as leaving `norm_type` might led to incorrect"
                " results in future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it"
                " would be very nice if you could open a Pull request for the `transformer/config.json` file"
            )
            deprecate("norm_type!=num_embeds_ada_norm", "1.0.0", deprecation_message, standard_warn=False)
            norm_type = "ada_norm"

        if self.is_input_continuous and self.is_input_vectorized:
            raise ValueError(
                f"Cannot define both `in_channels`: {in_channels} and `num_vector_embeds`: {num_vector_embeds}. Make"
                " sure that either `in_channels` or `num_vector_embeds` is None."
            )
        elif self.is_input_vectorized and self.is_input_patches:
            raise ValueError(
                f"Cannot define both `num_vector_embeds`: {num_vector_embeds} and `patch_size`: {patch_size}. Make"
                " sure that either `num_vector_embeds` or `num_patches` is None."
            )
        elif not self.is_input_continuous and not self.is_input_vectorized and not self.is_input_patches:
            raise ValueError(
                f"Has to define `in_channels`: {in_channels}, `num_vector_embeds`: {num_vector_embeds}, or patch_size:"
                f" {patch_size}. Make sure that `in_channels`, `num_vector_embeds` or `num_patches` is not None."
            )

        # 2. Define input layers
        if self.is_input_continuous:
            self.in_channels = in_channels

            self.norm = torch.nn.GroupNorm(num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True)
            if use_linear_projection:
                self.proj_in = LoRACompatibleLinear(in_channels, inner_dim)
            else:
                self.proj_in = LoRACompatibleConv(in_channels, inner_dim, kernel_size=1, stride=1, padding=0)
        elif self.is_input_vectorized:
            assert sample_size is not None, "Transformer2DModel over discrete input must provide sample_size"
            assert num_vector_embeds is not None, "Transformer2DModel over discrete input must provide num_embed"

            self.height = sample_size
            self.width = sample_size
            self.num_vector_embeds = num_vector_embeds
            self.num_latent_pixels = self.height * self.width

            self.latent_image_embedding = ImagePositionalEmbeddings(
                num_embed=num_vector_embeds, embed_dim=inner_dim, height=self.height, width=self.width
            )
        elif self.is_input_patches:
            assert sample_size is not None, "Transformer2DModel over patched input must provide sample_size"

            self.height = sample_size
            self.width = sample_size

            self.patch_size = patch_size
            self.pos_embed = PatchEmbed(
                height=sample_size,
                width=sample_size,
                patch_size=patch_size,
                in_channels=in_channels,
                embed_dim=inner_dim,
            )

        # 3. Define transformers blocks
        self.transformer_blocks = nn.ModuleList(
            [
                BasicConditionalTransformerBlock(
                    inner_dim,
                    num_attention_heads,
                    attention_head_dim,
                    dropout=dropout,
                    cross_attention_dim=cross_attention_dim,
                    activation_fn=activation_fn,
                    num_embeds_ada_norm=num_embeds_ada_norm,
                    attention_bias=attention_bias,
                    only_cross_attention=only_cross_attention,
                    double_self_attention=double_self_attention,
                    upcast_attention=upcast_attention,
                    norm_type=norm_type,
                    norm_elementwise_affine=norm_elementwise_affine,
                    attention_type=attention_type,
                    # additional
                    n_frames=n_frames,
                    is_temporal=is_temporal,
                    augment_temporal_attention=augment_temporal_attention,
                    rotary_emb=rotary_emb,
                )
                for d in range(num_layers)
            ]
        )

        # 4. Define output layers
        self.out_channels = in_channels if out_channels is None else out_channels
        if self.is_input_continuous:
            # TODO: should use out_channels for continuous projections
            if use_linear_projection:
                self.proj_out = LoRACompatibleLinear(inner_dim, in_channels)
            else:
                self.proj_out = LoRACompatibleConv(inner_dim, in_channels, kernel_size=1, stride=1, padding=0)
        elif self.is_input_vectorized:
            self.norm_out = nn.LayerNorm(inner_dim)
            self.out = nn.Linear(inner_dim, self.num_vector_embeds - 1)
        elif self.is_input_patches:
            self.norm_out = nn.LayerNorm(inner_dim, elementwise_affine=False, eps=1e-6)
            self.proj_out_1 = nn.Linear(inner_dim, 2 * inner_dim)
            self.proj_out_2 = nn.Linear(inner_dim, patch_size * patch_size * self.out_channels)

        self.alpha = None
        if is_temporal:
            self.alpha = nn.Parameter(torch.ones(1))

        self.gradient_checkpointing = False

    def forward(
        self,
        hidden_states: torch.Tensor,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        timestep: Optional[torch.LongTensor] = None,
        class_labels: Optional[torch.LongTensor] = None,
        cross_attention_kwargs: Dict[str, Any] = None,
        attention_mask: Optional[torch.Tensor] = None,
        encoder_attention_mask: Optional[torch.Tensor] = None,
        return_dict: bool = True,
        condition_on_first_frame: bool = False,
    ):
        input_states = hidden_states
        input_height, input_width = hidden_states.shape[-2:]
        # 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)

        # Retrieve lora scale.
        lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0

        # 1. Input
        if self.is_input_continuous:
            batch, _, height, width = hidden_states.shape
            residual = hidden_states

            hidden_states = self.norm(hidden_states)
            if not self.use_linear_projection:
                hidden_states = self.proj_in(hidden_states, lora_scale)
                inner_dim = hidden_states.shape[1]
                hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * width, inner_dim)
            else:
                inner_dim = hidden_states.shape[1]
                hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * width, inner_dim)
                hidden_states = self.proj_in(hidden_states, scale=lora_scale)

        elif self.is_input_vectorized:
            hidden_states = self.latent_image_embedding(hidden_states)
        elif self.is_input_patches:
            hidden_states = self.pos_embed(hidden_states)

        # 2. Blocks
        for block in self.transformer_blocks:
            if self.training and self.gradient_checkpointing:
                hidden_states = torch.utils.checkpoint.checkpoint(
                    block,
                    hidden_states,
                    attention_mask,
                    encoder_hidden_states,
                    encoder_attention_mask,
                    timestep,
                    cross_attention_kwargs,
                    class_labels,
                    use_reentrant=False,
                )
            else:
                hidden_states = block(
                    hidden_states,
                    attention_mask=attention_mask,
                    encoder_hidden_states=encoder_hidden_states,
                    encoder_attention_mask=encoder_attention_mask,
                    timestep=timestep,
                    cross_attention_kwargs=cross_attention_kwargs,
                    class_labels=class_labels,
                    # additional
                    condition_on_first_frame=condition_on_first_frame,
                    input_height=input_height,
                    input_width=input_width,
                )

        # 3. Output
        if self.is_input_continuous:
            if not self.use_linear_projection:
                hidden_states = hidden_states.reshape(batch, height, width, inner_dim).permute(0, 3, 1, 2).contiguous()
                hidden_states = self.proj_out(hidden_states, scale=lora_scale)
            else:
                hidden_states = self.proj_out(hidden_states, scale=lora_scale)
                hidden_states = hidden_states.reshape(batch, height, width, inner_dim).permute(0, 3, 1, 2).contiguous()

            output = hidden_states + residual
        elif self.is_input_vectorized:
            hidden_states = self.norm_out(hidden_states)
            logits = self.out(hidden_states)
            # (batch, self.num_vector_embeds - 1, self.num_latent_pixels)
            logits = logits.permute(0, 2, 1)

            # log(p(x_0))
            output = F.log_softmax(logits.double(), dim=1).float()
        elif self.is_input_patches:
            # TODO: cleanup!
            conditioning = self.transformer_blocks[0].norm1.emb(
                timestep, class_labels, hidden_dtype=hidden_states.dtype
            )
            shift, scale = self.proj_out_1(F.silu(conditioning)).chunk(2, dim=1)
            hidden_states = self.norm_out(hidden_states) * (1 + scale[:, None]) + shift[:, None]
            hidden_states = self.proj_out_2(hidden_states)

            # unpatchify
            height = width = int(hidden_states.shape[1] ** 0.5)
            hidden_states = hidden_states.reshape(
                shape=(-1, height, width, self.patch_size, self.patch_size, self.out_channels)
            )
            hidden_states = torch.einsum("nhwpqc->nchpwq", hidden_states)
            output = hidden_states.reshape(
                shape=(-1, self.out_channels, height * self.patch_size, width * self.patch_size)
            )

        if self.alpha is not None:
            with torch.no_grad():
                self.alpha.clamp_(0, 1)

            output = self.alpha * input_states + (1 - self.alpha) * output

        if not return_dict:
            return (output,)

        return Transformer2DModelOutput(sample=output)


@maybe_allow_in_graph
class BasicConditionalTransformerBlock(nn.Module):
    """ transformer block with first frame conditioning """
    def __init__(
        self,
        dim: int,
        num_attention_heads: int,
        attention_head_dim: int,
        dropout=0.0,
        cross_attention_dim: Optional[int] = None,
        activation_fn: str = "geglu",
        num_embeds_ada_norm: Optional[int] = None,
        attention_bias: bool = False,
        only_cross_attention: bool = False,
        double_self_attention: bool = False,
        upcast_attention: bool = False,
        norm_elementwise_affine: bool = True,
        norm_type: str = "layer_norm",
        final_dropout: bool = False,
        attention_type: str = "default",
        # additional
        n_frames: int = 8,
        is_temporal: bool = False,
        augment_temporal_attention: bool = False,
        rotary_emb=False,
    ):
        super().__init__()
        self.n_frames = n_frames
        self.only_cross_attention = only_cross_attention
        self.augment_temporal_attention = augment_temporal_attention
        self.is_temporal = is_temporal

        self.use_ada_layer_norm_zero = (num_embeds_ada_norm is not None) and norm_type == "ada_norm_zero"
        self.use_ada_layer_norm = (num_embeds_ada_norm is not None) and norm_type == "ada_norm"

        if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None:
            raise ValueError(
                f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to"
                f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}."
            )

        # Define 3 blocks. Each block has its own normalization layer.
        # 1. Self-Attn
        if self.use_ada_layer_norm:
            self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm)
        elif self.use_ada_layer_norm_zero:
            self.norm1 = AdaLayerNormZero(dim, num_embeds_ada_norm)
        else:
            self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
        
        if not is_temporal:
            self.attn1 = ConditionalAttention(
                query_dim=dim,
                heads=num_attention_heads,
                dim_head=attention_head_dim,
                dropout=dropout,
                bias=attention_bias,
                cross_attention_dim=cross_attention_dim if only_cross_attention else None,
                upcast_attention=upcast_attention,
            )
        else:
            self.attn1 = TemporalConditionalAttention(
                query_dim=dim,
                heads=num_attention_heads,
                dim_head=attention_head_dim,
                dropout=dropout,
                bias=attention_bias,
                cross_attention_dim=cross_attention_dim if only_cross_attention else None,
                upcast_attention=upcast_attention,
                # additional
                n_frames=n_frames,
                rotary_emb=rotary_emb,
            )

        # 2. Cross-Attn
        if cross_attention_dim is not None or double_self_attention:
            # We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
            # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
            # the second cross attention block.
            self.norm2 = (
                AdaLayerNorm(dim, num_embeds_ada_norm)
                if self.use_ada_layer_norm
                else nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
            )
            if not is_temporal:
                self.attn2 = ConditionalAttention(
                    query_dim=dim,
                    cross_attention_dim=cross_attention_dim if not double_self_attention else None,
                    heads=num_attention_heads,
                    dim_head=attention_head_dim,
                    dropout=dropout,
                    bias=attention_bias,
                    upcast_attention=upcast_attention,
                )  # is self-attn if encoder_hidden_states is none
            else:
                self.attn2 = TemporalConditionalAttention(
                    query_dim=dim,
                    cross_attention_dim=cross_attention_dim if not double_self_attention else None,
                    heads=num_attention_heads,
                    dim_head=attention_head_dim,
                    dropout=dropout,
                    bias=attention_bias,
                    upcast_attention=upcast_attention,
                    # additional
                    n_frames=n_frames,
                    rotary_emb=rotary_emb,
                )
        else:
            self.norm2 = None
            self.attn2 = None

        # 3. Feed-forward
        self.norm3 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
        self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn, final_dropout=final_dropout)

        # 4. Fuser
        if attention_type == "gated" or attention_type == "gated-text-image":
            self.fuser = GatedSelfAttentionDense(dim, cross_attention_dim, num_attention_heads, attention_head_dim)

        # let chunk size default to None
        self._chunk_size = None
        self._chunk_dim = 0

    def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int):
        # Sets chunk feed-forward
        self._chunk_size = chunk_size
        self._chunk_dim = dim

    def forward(
        self,
        hidden_states: torch.FloatTensor,
        attention_mask: Optional[torch.FloatTensor] = None,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
        timestep: Optional[torch.LongTensor] = None,
        cross_attention_kwargs: Dict[str, Any] = None,
        class_labels: Optional[torch.LongTensor] = None,
        condition_on_first_frame: bool = False,
        input_height: Optional[int] = None,
        input_width: Optional[int] = None,
    ):
        # Notice that normalization is always applied before the real computation in the following blocks.
        # 0. Self-Attention
        if self.use_ada_layer_norm:
            norm_hidden_states = self.norm1(hidden_states, timestep)
        elif self.use_ada_layer_norm_zero:
            norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(
                hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype
            )
        else:
            norm_hidden_states = self.norm1(hidden_states)

        # 1. Retrieve lora scale.
        lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0

        # 2. Prepare GLIGEN inputs
        cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {}
        gligen_kwargs = cross_attention_kwargs.pop("gligen", None)

        if condition_on_first_frame:
            first_frame_hidden_states = rearrange(norm_hidden_states, '(b f) d h -> b f d h', f=self.n_frames)[:, 0, :, :]
            first_frame_hidden_states = repeat(first_frame_hidden_states, 'b d h -> b f d h', f=self.n_frames)
            first_frame_hidden_states = rearrange(first_frame_hidden_states, 'b f d h -> (b f) d h')
            first_frame_concat_hidden_states = torch.cat((norm_hidden_states, first_frame_hidden_states), dim=1)
            attn_output = self.attn1(
                norm_hidden_states,
                encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else first_frame_concat_hidden_states,
                attention_mask=attention_mask,
                **cross_attention_kwargs,
            )
        elif self.is_temporal and self.augment_temporal_attention:
            first_frame_hidden_states = rearrange(norm_hidden_states, '(b f) d h -> b f d h', f=self.n_frames)[:, 0, :, :]
            first_frame_hidden_states = rearrange(first_frame_hidden_states, 'b (h w) c -> b h w c', h=input_height, w=input_width)
            first_frame_hidden_states = first_frame_hidden_states.permute(0, 3, 1, 2)
            padded_first_frame = torch.nn.functional.pad(first_frame_hidden_states, (1, 1, 1, 1), "replicate")
            first_frame_windows = padded_first_frame.unfold(2, 3, 1).unfold(3, 3, 1)
            mask = torch.tensor([[1, 1, 1], [1, 0, 1], [1, 1, 1]], dtype=torch.bool)
            adjacent_slices = first_frame_windows[:, :, :, :, mask]
            attn_output = self.attn1(
                norm_hidden_states,
                encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
                attention_mask=attention_mask,
                adjacent_slices=adjacent_slices,
                **cross_attention_kwargs,
            )
        else:
            attn_output = self.attn1(
                norm_hidden_states,
                encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
                attention_mask=attention_mask,
                **cross_attention_kwargs,
            )
        if self.use_ada_layer_norm_zero:
            attn_output = gate_msa.unsqueeze(1) * attn_output
        hidden_states = attn_output + hidden_states

        # 2.5 GLIGEN Control
        if gligen_kwargs is not None:
            hidden_states = self.fuser(hidden_states, gligen_kwargs["objs"])
        # 2.5 ends

        # 3. Cross-Attention
        if self.attn2 is not None:
            norm_hidden_states = (
                self.norm2(hidden_states, timestep) if self.use_ada_layer_norm else self.norm2(hidden_states)
            )

            attn_output = self.attn2(
                norm_hidden_states,
                encoder_hidden_states=encoder_hidden_states,
                attention_mask=encoder_attention_mask,
                **cross_attention_kwargs,
            )
            hidden_states = attn_output + hidden_states

        # 4. Feed-forward
        norm_hidden_states = self.norm3(hidden_states)

        if self.use_ada_layer_norm_zero:
            norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]

        if self._chunk_size is not None:
            # "feed_forward_chunk_size" can be used to save memory
            if norm_hidden_states.shape[self._chunk_dim] % self._chunk_size != 0:
                raise ValueError(
                    f"`hidden_states` dimension to be chunked: {norm_hidden_states.shape[self._chunk_dim]} has to be divisible by chunk size: {self._chunk_size}. Make sure to set an appropriate `chunk_size` when calling `unet.enable_forward_chunking`."
                )

            num_chunks = norm_hidden_states.shape[self._chunk_dim] // self._chunk_size
            ff_output = torch.cat(
                [
                    self.ff(hid_slice, scale=lora_scale)
                    for hid_slice in norm_hidden_states.chunk(num_chunks, dim=self._chunk_dim)
                ],
                dim=self._chunk_dim,
            )
        else:
            ff_output = self.ff(norm_hidden_states, scale=lora_scale)

        if self.use_ada_layer_norm_zero:
            ff_output = gate_mlp.unsqueeze(1) * ff_output

        hidden_states = ff_output + hidden_states

        return hidden_states