pyramid_dit/modeling_pyramid_mmdit.py

import torch
import torch.nn as nn
import os
import torch.nn.functional as F

from einops import rearrange
from diffusers.utils.torch_utils import randn_tensor
from diffusers.models.modeling_utils import ModelMixin
from diffusers.configuration_utils import ConfigMixin, register_to_config
from diffusers.utils import is_torch_version
from typing import Any, Callable, Dict, List, Optional, Union
from tqdm import tqdm

from .modeling_embedding import PatchEmbed3D, CombinedTimestepConditionEmbeddings
from .modeling_normalization import AdaLayerNormContinuous
from .modeling_mmdit_block import JointTransformerBlock

from trainer_misc import (
    is_sequence_parallel_initialized,
    get_sequence_parallel_group,
    get_sequence_parallel_world_size,
    get_sequence_parallel_rank,
    all_to_all,
)

from IPython import embed


def rope(pos: torch.Tensor, dim: int, theta: int) -> torch.Tensor:
    assert dim % 2 == 0, "The dimension must be even."

    scale = torch.arange(0, dim, 2, dtype=torch.float64, device=pos.device) / dim
    omega = 1.0 / (theta**scale)

    batch_size, seq_length = pos.shape
    out = torch.einsum("...n,d->...nd", pos, omega)
    cos_out = torch.cos(out)
    sin_out = torch.sin(out)

    stacked_out = torch.stack([cos_out, -sin_out, sin_out, cos_out], dim=-1)
    out = stacked_out.view(batch_size, -1, dim // 2, 2, 2)
    return out.float()


class EmbedNDRoPE(nn.Module):
    def __init__(self, dim: int, theta: int, axes_dim: List[int]):
        super().__init__()
        self.dim = dim
        self.theta = theta
        self.axes_dim = axes_dim

    def forward(self, ids: torch.Tensor) -> torch.Tensor:
        n_axes = ids.shape[-1]
        emb = torch.cat(
            [rope(ids[..., i], self.axes_dim[i], self.theta) for i in range(n_axes)],
            dim=-3,
        )
        return emb.unsqueeze(2)


class PyramidDiffusionMMDiT(ModelMixin, ConfigMixin):
    _supports_gradient_checkpointing = True

    @register_to_config
    def __init__(
        self,
        sample_size: int = 128,
        patch_size: int = 2,
        in_channels: int = 16,
        num_layers: int = 24,
        attention_head_dim: int = 64,
        num_attention_heads: int = 24,
        caption_projection_dim: int = 1152,
        pooled_projection_dim: int = 2048,
        pos_embed_max_size: int = 192,
        max_num_frames: int = 200,
        qk_norm: str = 'rms_norm',
        pos_embed_type: str = 'rope',
        temp_pos_embed_type: str = 'sincos',
        joint_attention_dim: int = 4096,
        use_gradient_checkpointing: bool = False,
        use_flash_attn: bool = True,
        use_temporal_causal: bool = False,
        use_t5_mask: bool = False,
        add_temp_pos_embed: bool = False,
        interp_condition_pos: bool = False,
    ):
        super().__init__()

        self.out_channels = in_channels
        self.inner_dim = num_attention_heads * attention_head_dim
        assert temp_pos_embed_type in ['rope', 'sincos']

        # The input latent embeder, using the name pos_embed to remain the same with SD#
        self.pos_embed = PatchEmbed3D(
            height=sample_size,
            width=sample_size,
            patch_size=patch_size,
            in_channels=in_channels,
            embed_dim=self.inner_dim,
            pos_embed_max_size=pos_embed_max_size,  # hard-code for now.
            max_num_frames=max_num_frames,
            pos_embed_type=pos_embed_type,
            temp_pos_embed_type=temp_pos_embed_type,
            add_temp_pos_embed=add_temp_pos_embed,
            interp_condition_pos=interp_condition_pos,
        )

        # The RoPE EMbedding
        if pos_embed_type == 'rope':
            self.rope_embed = EmbedNDRoPE(self.inner_dim, 10000, axes_dim=[16, 24, 24])
        else:
            self.rope_embed = None

        if temp_pos_embed_type == 'rope':
            self.temp_rope_embed = EmbedNDRoPE(self.inner_dim, 10000, axes_dim=[attention_head_dim])
        else:
            self.temp_rope_embed = None

        self.time_text_embed = CombinedTimestepConditionEmbeddings(
            embedding_dim=self.inner_dim, pooled_projection_dim=self.config.pooled_projection_dim,
        )
        self.context_embedder = nn.Linear(self.config.joint_attention_dim, self.config.caption_projection_dim)

        self.transformer_blocks = nn.ModuleList(
            [
                JointTransformerBlock(
                    dim=self.inner_dim,
                    num_attention_heads=num_attention_heads,
                    attention_head_dim=self.inner_dim,
                    qk_norm=qk_norm,
                    context_pre_only=i == num_layers - 1,
                    use_flash_attn=use_flash_attn,
                )
                for i in range(num_layers)
            ]
        )

        self.norm_out = AdaLayerNormContinuous(self.inner_dim, self.inner_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 = use_gradient_checkpointing
        self.patch_size = patch_size
        self.use_flash_attn = use_flash_attn
        self.use_temporal_causal = use_temporal_causal
        self.pos_embed_type = pos_embed_type
        self.temp_pos_embed_type = temp_pos_embed_type
        self.add_temp_pos_embed = add_temp_pos_embed

        if self.use_temporal_causal:
            print("Using temporal causal attention")
            assert self.use_flash_attn is False, "The flash attention does not support temporal causal"
        
        if interp_condition_pos:
            print("We interp the position embedding of condition latents")

        # init weights
        self.initialize_weights()

    def initialize_weights(self):
        # Initialize transformer layers:
        def _basic_init(module):
            if isinstance(module, (nn.Linear, nn.Conv2d, nn.Conv3d)):
                torch.nn.init.xavier_uniform_(module.weight)
                if module.bias is not None:
                    nn.init.constant_(module.bias, 0)
        self.apply(_basic_init)

        # Initialize patch_embed like nn.Linear (instead of nn.Conv2d):
        w = self.pos_embed.proj.weight.data
        nn.init.xavier_uniform_(w.view([w.shape[0], -1]))
        nn.init.constant_(self.pos_embed.proj.bias, 0)

        # Initialize all the conditioning to normal init
        nn.init.normal_(self.time_text_embed.timestep_embedder.linear_1.weight, std=0.02)
        nn.init.normal_(self.time_text_embed.timestep_embedder.linear_2.weight, std=0.02)
        nn.init.normal_(self.time_text_embed.text_embedder.linear_1.weight, std=0.02)
        nn.init.normal_(self.time_text_embed.text_embedder.linear_2.weight, std=0.02)
        nn.init.normal_(self.context_embedder.weight, std=0.02)

        # Zero-out adaLN modulation layers in DiT blocks:
        for block in self.transformer_blocks:
            nn.init.constant_(block.norm1.linear.weight, 0)
            nn.init.constant_(block.norm1.linear.bias, 0)
            nn.init.constant_(block.norm1_context.linear.weight, 0)
            nn.init.constant_(block.norm1_context.linear.bias, 0)

        # Zero-out output layers:
        nn.init.constant_(self.norm_out.linear.weight, 0)
        nn.init.constant_(self.norm_out.linear.bias, 0)
        nn.init.constant_(self.proj_out.weight, 0)
        nn.init.constant_(self.proj_out.bias, 0)

    @torch.no_grad()
    def _prepare_latent_image_ids(self, batch_size, temp, height, width, device):
        latent_image_ids = torch.zeros(temp, height, width, 3)
        latent_image_ids[..., 0] = latent_image_ids[..., 0] + torch.arange(temp)[:, None, None]
        latent_image_ids[..., 1] = latent_image_ids[..., 1] + torch.arange(height)[None, :, None]
        latent_image_ids[..., 2] = latent_image_ids[..., 2] + torch.arange(width)[None, None, :]

        latent_image_ids = latent_image_ids[None, :].repeat(batch_size, 1, 1, 1, 1)
        latent_image_ids = rearrange(latent_image_ids, 'b t h w c -> b (t h w) c')
        return latent_image_ids.to(device=device)

    @torch.no_grad()
    def _prepare_pyramid_latent_image_ids(self, batch_size, temp_list, height_list, width_list, device):
        base_width = width_list[-1]; base_height = height_list[-1]
        assert base_width == max(width_list)
        assert base_height == max(height_list)

        image_ids_list = []
        for temp, height, width in zip(temp_list, height_list, width_list):
            latent_image_ids = torch.zeros(temp, height, width, 3)

            if height != base_height:
                height_pos = F.interpolate(torch.arange(base_height)[None, None, :].float(), height, mode='linear').squeeze(0, 1)
            else:
                height_pos = torch.arange(base_height).float()
            if width != base_width:
                width_pos = F.interpolate(torch.arange(base_width)[None, None, :].float(), width, mode='linear').squeeze(0, 1)
            else:
                width_pos = torch.arange(base_width).float()

            latent_image_ids[..., 0] = latent_image_ids[..., 0] + torch.arange(temp)[:, None, None]  
            latent_image_ids[..., 1] = latent_image_ids[..., 1] + height_pos[None, :, None]
            latent_image_ids[..., 2] = latent_image_ids[..., 2] + width_pos[None, None, :]
            latent_image_ids = latent_image_ids[None, :].repeat(batch_size, 1, 1, 1, 1)
            latent_image_ids = rearrange(latent_image_ids, 'b t h w c -> b (t h w) c').to(device)
            image_ids_list.append(latent_image_ids)
    
        return image_ids_list

    @torch.no_grad()
    def _prepare_temporal_rope_ids(self, batch_size, temp, height, width, device, start_time_stamp=0):
        latent_image_ids = torch.zeros(temp, height, width, 1)
        latent_image_ids[..., 0] = latent_image_ids[..., 0] + torch.arange(start_time_stamp, start_time_stamp + temp)[:, None, None]
        latent_image_ids = latent_image_ids[None, :].repeat(batch_size, 1, 1, 1, 1)
        latent_image_ids = rearrange(latent_image_ids, 'b t h w c -> b (t h w) c')
        return latent_image_ids.to(device=device)

    @torch.no_grad()
    def _prepare_pyramid_temporal_rope_ids(self, sample, batch_size, device):
        image_ids_list = []

        for i_b, sample_ in enumerate(sample):
            if not isinstance(sample_, list):
                sample_ = [sample_]

            cur_image_ids = []
            start_time_stamp = 0

            for clip_ in sample_:
                _, _, temp, height, width = clip_.shape
                height = height // self.patch_size
                width = width // self.patch_size
                cur_image_ids.append(self._prepare_temporal_rope_ids(batch_size, temp, height, width, device, start_time_stamp=start_time_stamp))
                start_time_stamp += temp

            cur_image_ids = torch.cat(cur_image_ids, dim=1)
            image_ids_list.append(cur_image_ids)

        return image_ids_list

    def merge_input(self, sample, encoder_hidden_length, encoder_attention_mask):
        """
            Merge the input video with different resolutions into one sequence
            Sample: From low resolution to high resolution
        """
        if isinstance(sample[0], list):
            device = sample[0][-1].device
            pad_batch_size = sample[0][-1].shape[0]
        else:
            device = sample[0].device
            pad_batch_size = sample[0].shape[0]

        num_stages = len(sample)
        height_list = [];width_list = [];temp_list = []
        trainable_token_list = []

        for i_b, sample_ in enumerate(sample):
            if isinstance(sample_, list):
                sample_ = sample_[-1]
            _, _, temp, height, width = sample_.shape
            height = height // self.patch_size
            width = width // self.patch_size
            temp_list.append(temp)
            height_list.append(height)
            width_list.append(width)
            trainable_token_list.append(height * width * temp)

        # prepare the RoPE embedding if needed
        if self.pos_embed_type == 'rope':
            # TODO: support the 3D Rope for video
            raise NotImplementedError("Not compatible with video generation now")
            text_ids = torch.zeros(pad_batch_size, encoder_hidden_length, 3).to(device=device)
            image_ids_list = self._prepare_pyramid_latent_image_ids(pad_batch_size, temp_list, height_list, width_list, device)
            input_ids_list = [torch.cat([text_ids, image_ids], dim=1) for image_ids in image_ids_list]
            image_rotary_emb = [self.rope_embed(input_ids) for input_ids in input_ids_list]  # [bs, seq_len, 1, head_dim // 2, 2, 2]
        else:
            if self.temp_pos_embed_type == 'rope' and self.add_temp_pos_embed:
                image_ids_list = self._prepare_pyramid_temporal_rope_ids(sample, pad_batch_size, device)
                text_ids = torch.zeros(pad_batch_size, encoder_attention_mask.shape[1], 1).to(device=device)    
                input_ids_list = [torch.cat([text_ids, image_ids], dim=1) for image_ids in image_ids_list]
                image_rotary_emb = [self.temp_rope_embed(input_ids) for input_ids in input_ids_list]  # [bs, seq_len, 1, head_dim // 2, 2, 2]

                if is_sequence_parallel_initialized():
                    sp_group = get_sequence_parallel_group()
                    sp_group_size = get_sequence_parallel_world_size()
                    image_rotary_emb = [all_to_all(x_.repeat(1, 1, sp_group_size, 1, 1, 1), sp_group, sp_group_size, scatter_dim=2, gather_dim=0) for x_ in image_rotary_emb]
                    input_ids_list = [all_to_all(input_ids.repeat(1, 1, sp_group_size), sp_group, sp_group_size, scatter_dim=2, gather_dim=0) for input_ids in input_ids_list]

            else:
                image_rotary_emb = None

        hidden_states = self.pos_embed(sample)  # hidden states is a list of [b c t h w] b = real_b // num_stages
        hidden_length = []
    
        for i_b in range(num_stages):
            hidden_length.append(hidden_states[i_b].shape[1])

        # prepare the attention mask
        if self.use_flash_attn:
            attention_mask = None
            indices_list = []
            for i_p, length in enumerate(hidden_length):
                pad_attention_mask = torch.ones((pad_batch_size, length), dtype=encoder_attention_mask.dtype).to(device)
                pad_attention_mask = torch.cat([encoder_attention_mask[i_p::num_stages], pad_attention_mask], dim=1)
                
                if is_sequence_parallel_initialized():
                    sp_group = get_sequence_parallel_group()
                    sp_group_size = get_sequence_parallel_world_size()
                    pad_attention_mask = all_to_all(pad_attention_mask.unsqueeze(2).repeat(1, 1, sp_group_size), sp_group, sp_group_size, scatter_dim=2, gather_dim=0)
                    pad_attention_mask = pad_attention_mask.squeeze(2)

                seqlens_in_batch = pad_attention_mask.sum(dim=-1, dtype=torch.int32)
                indices = torch.nonzero(pad_attention_mask.flatten(), as_tuple=False).flatten()

                indices_list.append(
                    {
                        'indices': indices,
                        'seqlens_in_batch': seqlens_in_batch,
                    }
                )
            encoder_attention_mask = indices_list
        else:
            assert encoder_attention_mask.shape[1] == encoder_hidden_length
            real_batch_size = encoder_attention_mask.shape[0]
            # prepare text ids
            text_ids = torch.arange(1, real_batch_size + 1, dtype=encoder_attention_mask.dtype).unsqueeze(1).repeat(1, encoder_hidden_length)
            text_ids = text_ids.to(device)
            text_ids[encoder_attention_mask == 0] = 0

            # prepare image ids
            image_ids = torch.arange(1, real_batch_size + 1, dtype=encoder_attention_mask.dtype).unsqueeze(1).repeat(1, max(hidden_length))
            image_ids = image_ids.to(device)
            image_ids_list = []
            for i_p, length in enumerate(hidden_length):
                image_ids_list.append(image_ids[i_p::num_stages][:, :length])

            if is_sequence_parallel_initialized():
                sp_group = get_sequence_parallel_group()
                sp_group_size = get_sequence_parallel_world_size()
                text_ids = all_to_all(text_ids.unsqueeze(2).repeat(1, 1, sp_group_size), sp_group, sp_group_size, scatter_dim=2, gather_dim=0).squeeze(2)
                image_ids_list = [all_to_all(image_ids_.unsqueeze(2).repeat(1, 1, sp_group_size), sp_group, sp_group_size, scatter_dim=2, gather_dim=0).squeeze(2) for image_ids_ in image_ids_list]

            attention_mask = []
            for i_p in range(len(hidden_length)):
                image_ids = image_ids_list[i_p]
                token_ids = torch.cat([text_ids[i_p::num_stages], image_ids], dim=1)
                stage_attention_mask = rearrange(token_ids, 'b i -> b 1 i 1') == rearrange(token_ids, 'b j -> b 1 1 j')  # [bs, 1, q_len, k_len]
                if self.use_temporal_causal:
                    input_order_ids = input_ids_list[i_p].squeeze(2)
                    temporal_causal_mask = rearrange(input_order_ids, 'b i -> b 1 i 1') >= rearrange(input_order_ids, 'b j -> b 1 1 j')
                    stage_attention_mask = stage_attention_mask & temporal_causal_mask
                attention_mask.append(stage_attention_mask)

        return hidden_states, hidden_length, temp_list, height_list, width_list, trainable_token_list, encoder_attention_mask, attention_mask, image_rotary_emb

    def split_output(self, batch_hidden_states, hidden_length, temps, heights, widths, trainable_token_list):
        # To split the hidden states
        batch_size = batch_hidden_states.shape[0]
        output_hidden_list = []
        batch_hidden_states = torch.split(batch_hidden_states, hidden_length, dim=1)

        if is_sequence_parallel_initialized():
            sp_group_size = get_sequence_parallel_world_size()
            batch_size = batch_size // sp_group_size

        for i_p, length in enumerate(hidden_length):
            width, height, temp = widths[i_p], heights[i_p], temps[i_p]
            trainable_token_num = trainable_token_list[i_p]
            hidden_states = batch_hidden_states[i_p]

            if is_sequence_parallel_initialized():
                sp_group = get_sequence_parallel_group()
                sp_group_size = get_sequence_parallel_world_size()
                hidden_states = all_to_all(hidden_states, sp_group, sp_group_size, scatter_dim=0, gather_dim=1)

            # only the trainable token are taking part in loss computation
            hidden_states = hidden_states[:, -trainable_token_num:]

            # unpatchify
            hidden_states = hidden_states.reshape(
                shape=(batch_size, temp, height, width, self.patch_size, self.patch_size, self.out_channels)
            )
            hidden_states = rearrange(hidden_states, "b t h w p1 p2 c -> b t (h p1) (w p2) c")
            hidden_states = rearrange(hidden_states, "b t h w c -> b c t h w")
            output_hidden_list.append(hidden_states)

        return output_hidden_list

    def forward(
        self,
        sample: torch.FloatTensor, # [num_stages]
        encoder_hidden_states: torch.FloatTensor = None,
        encoder_attention_mask: torch.FloatTensor = None,
        pooled_projections: torch.FloatTensor = None,
        timestep_ratio: torch.FloatTensor = None,
    ):
        # Get the timestep embedding
        temb = self.time_text_embed(timestep_ratio, pooled_projections)
        encoder_hidden_states = self.context_embedder(encoder_hidden_states)
        encoder_hidden_length = encoder_hidden_states.shape[1]

        # Get the input sequence
        hidden_states, hidden_length, temps, heights, widths, trainable_token_list, encoder_attention_mask, \
                attention_mask, image_rotary_emb = self.merge_input(sample, encoder_hidden_length, encoder_attention_mask)
        
        # split the long latents if necessary
        if is_sequence_parallel_initialized():
            sp_group = get_sequence_parallel_group()
            sp_group_size = get_sequence_parallel_world_size()
            
            # sync the input hidden states
            batch_hidden_states = []
            for i_p, hidden_states_ in enumerate(hidden_states):
                assert hidden_states_.shape[1] % sp_group_size == 0, "The sequence length should be divided by sequence parallel size"
                hidden_states_ = all_to_all(hidden_states_, sp_group, sp_group_size, scatter_dim=1, gather_dim=0)
                hidden_length[i_p] = hidden_length[i_p] // sp_group_size
                batch_hidden_states.append(hidden_states_)

            # sync the encoder hidden states
            hidden_states = torch.cat(batch_hidden_states, dim=1)
            encoder_hidden_states = all_to_all(encoder_hidden_states, sp_group, sp_group_size, scatter_dim=1, gather_dim=0)
            temb = all_to_all(temb.unsqueeze(1).repeat(1, sp_group_size, 1), sp_group, sp_group_size, scatter_dim=1, gather_dim=0)
            temb = temb.squeeze(1)
        else:
            hidden_states = torch.cat(hidden_states, dim=1)

        # print(hidden_length)
        for i_b, block in enumerate(self.transformer_blocks):
            if self.training and self.gradient_checkpointing and (i_b >= 2):
                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 {}
                encoder_hidden_states, hidden_states = torch.utils.checkpoint.checkpoint(
                    create_custom_forward(block),
                    hidden_states,
                    encoder_hidden_states,
                    encoder_attention_mask,
                    temb,
                    attention_mask,
                    hidden_length,
                    image_rotary_emb,
                    **ckpt_kwargs,
                )

            else:
                encoder_hidden_states, hidden_states = block(
                    hidden_states=hidden_states, 
                    encoder_hidden_states=encoder_hidden_states,
                    encoder_attention_mask=encoder_attention_mask,
                    temb=temb,
                    attention_mask=attention_mask,
                    hidden_length=hidden_length,
                    image_rotary_emb=image_rotary_emb,
                )

        hidden_states = self.norm_out(hidden_states, temb, hidden_length=hidden_length)
        hidden_states = self.proj_out(hidden_states)

        output = self.split_output(hidden_states, hidden_length, temps, heights, widths, trainable_token_list)

        return output