wan/distributed/xdit_context_parallel.py

# Copyright 2024-2025 The Alibaba Wan Team Authors. All rights reserved.
import numpy as np
import torch
import torch.nn as nn
import torch.cuda.amp as amp
from xfuser.core.distributed import (
    get_sequence_parallel_rank,
    get_sequence_parallel_world_size,
    get_sp_group,
)
from einops import rearrange
from xfuser.core.long_ctx_attention import xFuserLongContextAttention
import xformers.ops

from ..modules.model import sinusoidal_embedding_1d
from ..utils.multitalk_utils import get_attn_map_with_target, split_token_counts_and_frame_ids, normalize_and_scale
from ..modules.attention import SingleStreamAttention, SingleStreamMutiAttention


def pad_freqs(original_tensor, target_len):
    seq_len, s1, s2 = original_tensor.shape
    pad_size = target_len - seq_len
    padding_tensor = torch.ones(
        pad_size,
        s1,
        s2,
        dtype=original_tensor.dtype,
        device=original_tensor.device)
    padded_tensor = torch.cat([original_tensor, padding_tensor], dim=0)
    return padded_tensor


@amp.autocast(enabled=False)
def rope_apply(x, grid_sizes, freqs):
    """
    x:          [B, L, N, C].
    grid_sizes: [B, 3].
    freqs:      [M, C // 2].
    """
    s, n, c = x.size(1), x.size(2), x.size(3) // 2
    # split freqs
    freqs = freqs.split([c - 2 * (c // 3), c // 3, c // 3], dim=1) # [[N, head_dim/2], [N, head_dim/2], [N, head_dim/2]] # T H W 极坐标

    # loop over samples
    output = []
    for i, (f, h, w) in enumerate(grid_sizes.tolist()):
        seq_len = f * h * w

        # precompute multipliers
        x_i = torch.view_as_complex(x[i, :s].to(torch.float64).reshape(
            s, n, -1, 2)) # [L, N, C/2] # 极坐标
        freqs_i = torch.cat([
            freqs[0][:f].view(f, 1, 1, -1).expand(f, h, w, -1),
            freqs[1][:h].view(1, h, 1, -1).expand(f, h, w, -1),
            freqs[2][:w].view(1, 1, w, -1).expand(f, h, w, -1)
        ],
                            dim=-1).reshape(seq_len, 1, -1) # seq_lens, 1,  3 * dim / 2 (T H W)

        # apply rotary embedding
        sp_size = get_sequence_parallel_world_size()
        sp_rank = get_sequence_parallel_rank()
        freqs_i = pad_freqs(freqs_i, s * sp_size)
        s_per_rank = s
        freqs_i_rank = freqs_i[(sp_rank * s_per_rank):((sp_rank + 1) *
                                                       s_per_rank), :, :]
        x_i = torch.view_as_real(x_i * freqs_i_rank).flatten(2)
        x_i = torch.cat([x_i, x[i, s:]])

        # append to collection
        output.append(x_i)
    return torch.stack(output).float()


def usp_dit_forward_vace(self, x, vace_context, seq_len, kwargs):
    # embeddings
    c = [self.vace_patch_embedding(u.unsqueeze(0)) for u in vace_context]
    c = [u.flatten(2).transpose(1, 2) for u in c]
    c = torch.cat([
        torch.cat([u, u.new_zeros(1, seq_len - u.size(1), u.size(2))], dim=1)
        for u in c
    ])

    # arguments
    new_kwargs = dict(x=x)
    new_kwargs.update(kwargs)

    # Context Parallel
    c = torch.chunk(
        c, get_sequence_parallel_world_size(),
        dim=1)[get_sequence_parallel_rank()]

    hints = []
    for block in self.vace_blocks:
        c, c_skip = block(c, **new_kwargs)
        hints.append(c_skip)
    return hints


def usp_dit_forward(
    self,
    x,
    t,
    context,
    seq_len,
    vace_context=None,
    vace_context_scale=1.0,
    clip_fea=None,
    y=None,
):
    """
    x:              A list of videos each with shape [C, T, H, W].
    t:              [B].
    context:        A list of text embeddings each with shape [L, C].
    """
    if self.model_type == 'i2v':
        assert clip_fea is not None and y is not None
    # params
    device = self.patch_embedding.weight.device
    if self.freqs.device != device:
        self.freqs = self.freqs.to(device)

    if self.model_type != 'vace' and y is not None:
        x = [torch.cat([u, v], dim=0) for u, v in zip(x, y)]

    # embeddings
    x = [self.patch_embedding(u.unsqueeze(0)) for u in x]
    grid_sizes = torch.stack(
        [torch.tensor(u.shape[2:], dtype=torch.long) for u in x])
    x = [u.flatten(2).transpose(1, 2) for u in x]
    seq_lens = torch.tensor([u.size(1) for u in x], dtype=torch.long)
    assert seq_lens.max() <= seq_len
    x = torch.cat([
        torch.cat([u, u.new_zeros(1, seq_len - u.size(1), u.size(2))], dim=1)
        for u in x
    ])

    # time embeddings
    with amp.autocast(dtype=torch.float32):
        e = self.time_embedding(
            sinusoidal_embedding_1d(self.freq_dim, t).float())
        e0 = self.time_projection(e).unflatten(1, (6, self.dim))
        assert e.dtype == torch.float32 and e0.dtype == torch.float32

    # context
    context_lens = None
    context = self.text_embedding(
        torch.stack([
            torch.cat([u, u.new_zeros(self.text_len - u.size(0), u.size(1))])
            for u in context
        ]))

    if self.model_type != 'vace' and clip_fea is not None:
        context_clip = self.img_emb(clip_fea)  # bs x 257 x dim
        context = torch.concat([context_clip, context], dim=1)

    # arguments
    kwargs = dict(
        e=e0,
        seq_lens=seq_lens,
        grid_sizes=grid_sizes,
        freqs=self.freqs,
        context=context,
        context_lens=context_lens)
    
    # Context Parallel
    x = torch.chunk(
        x, get_sequence_parallel_world_size(),
        dim=1)[get_sequence_parallel_rank()]

    for block in self.blocks:
        x = block(x, **kwargs)

    # head
    x = self.head(x, e)

    # Context Parallel
    x = get_sp_group().all_gather(x, dim=1)

    # unpatchify
    x = self.unpatchify(x, grid_sizes)
    return [u.float() for u in x]


def usp_attn_forward(self,
                     x,
                     seq_lens,
                     grid_sizes,
                     freqs,
                     dtype=torch.bfloat16):
    b, s, n, d = *x.shape[:2], self.num_heads, self.head_dim
    half_dtypes = (torch.float16, torch.bfloat16)

    def half(x):
        return x if x.dtype in half_dtypes else x.to(dtype)

    # query, key, value function
    def qkv_fn(x):
        q = self.norm_q(self.q(x)).view(b, s, n, d)
        k = self.norm_k(self.k(x)).view(b, s, n, d)
        v = self.v(x).view(b, s, n, d)
        return q, k, v

    q, k, v = qkv_fn(x)
    q = rope_apply(q, grid_sizes, freqs)
    k = rope_apply(k, grid_sizes, freqs)

    # TODO: We should use unpaded q,k,v for attention.
    # k_lens = seq_lens // get_sequence_parallel_world_size()
    # if k_lens is not None:
    #     q = torch.cat([u[:l] for u, l in zip(q, k_lens)]).unsqueeze(0)
    #     k = torch.cat([u[:l] for u, l in zip(k, k_lens)]).unsqueeze(0)
    #     v = torch.cat([u[:l] for u, l in zip(v, k_lens)]).unsqueeze(0)

    x = xFuserLongContextAttention()(
        None,
        query=half(q),
        key=half(k),
        value=half(v),
        window_size=self.window_size)

    # TODO: padding after attention.
    # x = torch.cat([x, x.new_zeros(b, s - x.size(1), n, d)], dim=1)

    # output
    x = x.flatten(2)
    x = self.o(x)
    return x




def usp_dit_forward_multitalk(
    self,
    x,
    t,
    context,
    seq_len,
    clip_fea=None,
    y=None,
    audio=None,
    ref_target_masks=None,
):
    """
    x:              A list of videos each with shape [C, T, H, W].
    t:              [B].
    context:        A list of text embeddings each with shape [L, C].
    """
    
    assert clip_fea is not None and y is not None
    # params
    device = self.patch_embedding.weight.device
    if self.freqs.device != device:
        self.freqs = self.freqs.to(device)

    _, T, H, W = x[0].shape
    N_t = T // self.patch_size[0]
    N_h = H // self.patch_size[1]
    N_w = W // self.patch_size[2]

    if y is not None:
        x = [torch.cat([u, v], dim=0) for u, v in zip(x, y)]
    x[0] = x[0].to(context[0].dtype)

    # embeddings
    x = [self.patch_embedding(u.unsqueeze(0)) for u in x]
    grid_sizes = torch.stack(
        [torch.tensor(u.shape[2:], dtype=torch.long) for u in x])
    x = [u.flatten(2).transpose(1, 2) for u in x]
    seq_lens = torch.tensor([u.size(1) for u in x], dtype=torch.long)
    assert seq_lens.max() <= seq_len
    x = torch.cat([
        torch.cat([u, u.new_zeros(1, seq_len - u.size(1), u.size(2))], dim=1)
        for u in x
    ])

    # time embeddings
    with amp.autocast(dtype=torch.float32):
        e = self.time_embedding(
            sinusoidal_embedding_1d(self.freq_dim, t).float())
        e0 = self.time_projection(e).unflatten(1, (6, self.dim))
        assert e.dtype == torch.float32 and e0.dtype == torch.float32

    # context
    context_lens = None
    context = self.text_embedding(
        torch.stack([
            torch.cat([u, u.new_zeros(self.text_len - u.size(0), u.size(1))])
            for u in context
        ]))

    if clip_fea is not None:
        context_clip = self.img_emb(clip_fea)  
        context = torch.concat([context_clip, context], dim=1)

    # get audio token
    audio_cond = audio.to(device=x.device, dtype=x.dtype)
    first_frame_audio_emb_s = audio_cond[:, :1, ...] 
    latter_frame_audio_emb = audio_cond[:, 1:, ...] 
    latter_frame_audio_emb = rearrange(latter_frame_audio_emb, "b (n_t n) w s c -> b n_t n w s c", n=self.vae_scale) 
    middle_index = self.audio_window // 2
    latter_first_frame_audio_emb = latter_frame_audio_emb[:, :, :1, :middle_index+1, ...] 
    latter_first_frame_audio_emb = rearrange(latter_first_frame_audio_emb, "b n_t n w s c -> b n_t (n w) s c") 
    latter_last_frame_audio_emb = latter_frame_audio_emb[:, :, -1:, middle_index:, ...] 
    latter_last_frame_audio_emb = rearrange(latter_last_frame_audio_emb, "b n_t n w s c -> b n_t (n w) s c") 
    latter_middle_frame_audio_emb = latter_frame_audio_emb[:, :, 1:-1, middle_index:middle_index+1, ...] 
    latter_middle_frame_audio_emb = rearrange(latter_middle_frame_audio_emb, "b n_t n w s c -> b n_t (n w) s c") 
    latter_frame_audio_emb_s = torch.concat([latter_first_frame_audio_emb, latter_middle_frame_audio_emb, latter_last_frame_audio_emb], dim=2) 
    audio_embedding = self.audio_proj(first_frame_audio_emb_s, latter_frame_audio_emb_s) 
    human_num = len(audio_embedding)
    audio_embedding = torch.concat(audio_embedding.split(1), dim=2).to(x.dtype)


    # convert ref_target_masks to token_ref_target_masks
    if ref_target_masks is not None:
        ref_target_masks = ref_target_masks.unsqueeze(0).to(torch.float32) 
        token_ref_target_masks = nn.functional.interpolate(ref_target_masks, size=(N_h, N_w), mode='nearest') 
        token_ref_target_masks = token_ref_target_masks.squeeze(0) 
        token_ref_target_masks = (token_ref_target_masks > 0)
        token_ref_target_masks = token_ref_target_masks.view(token_ref_target_masks.shape[0], -1) 
        token_ref_target_masks = token_ref_target_masks.to(x.dtype)
    
    if self.enable_teacache:
        modulated_inp = e0 if self.use_ret_steps else e
        if self.cnt%3==0: # cond
            if self.cnt < self.ret_steps or self.cnt >= self.cutoff_steps:
                should_calc_cond = True
                self.accumulated_rel_l1_distance_cond = 0
            else:
                rescale_func = np.poly1d(self.coefficients)
                self.accumulated_rel_l1_distance_cond += rescale_func(((modulated_inp-self.previous_e0_cond).abs().mean() / self.previous_e0_cond.abs().mean()).cpu().item())
                # print("accumulated_rel_l1_distance_even", self.accumulated_rel_l1_distance_even)
                if self.accumulated_rel_l1_distance_cond < self.teacache_thresh:
                    should_calc_cond = False
                else:
                    should_calc_cond = True
                    self.accumulated_rel_l1_distance_cond = 0
            self.previous_e0_cond = modulated_inp.clone()
        elif self.cnt%3==1: # drop_text
            if self.cnt < self.ret_steps or self.cnt >= self.cutoff_steps:
                should_calc_drop_text = True
                self.accumulated_rel_l1_distance_drop_text = 0
            else:
                rescale_func = np.poly1d(self.coefficients)
                self.accumulated_rel_l1_distance_drop_text += rescale_func(((modulated_inp-self.previous_e0_drop_text).abs().mean() / self.previous_e0_drop_text.abs().mean()).cpu().item())
                if self.accumulated_rel_l1_distance_drop_text < self.teacache_thresh:
                    should_calc_drop_text = False
                else:
                    should_calc_drop_text = True
                    self.accumulated_rel_l1_distance_drop_text = 0
            self.previous_e0_drop_text = modulated_inp.clone()
        else: # uncond
            if self.cnt < self.ret_steps or self.cnt >= self.cutoff_steps:
                should_calc_uncond = True
                self.accumulated_rel_l1_distance_uncond = 0
            else:
                rescale_func = np.poly1d(self.coefficients)
                self.accumulated_rel_l1_distance_uncond += rescale_func(((modulated_inp-self.previous_e0_uncond).abs().mean() / self.previous_e0_uncond.abs().mean()).cpu().item())
                if self.accumulated_rel_l1_distance_uncond < self.teacache_thresh:
                    should_calc_uncond = False
                else:
                    should_calc_uncond = True
                    self.accumulated_rel_l1_distance_uncond = 0
            self.previous_e0_uncond = modulated_inp.clone()

    # Context Parallel
    x = torch.chunk(
        x, get_sequence_parallel_world_size(),
        dim=1)[get_sequence_parallel_rank()]

    # arguments
    kwargs = dict(
        e=e0,
        seq_lens=seq_lens,
        grid_sizes=grid_sizes,
        freqs=self.freqs,
        context=context,
        context_lens=context_lens,
        audio_embedding=audio_embedding,
        ref_target_masks=token_ref_target_masks,
        human_num=human_num,
        )

    if self.enable_teacache:
        if self.cnt%3==0:
            if not should_calc_cond:
                x +=  self.previous_residual_cond
            else:
                ori_x = x.clone()
                for block in self.blocks:
                    x = block(x, **kwargs)
                self.previous_residual_cond = x - ori_x
        elif self.cnt%3==1:
            if not should_calc_drop_text:
                x +=  self.previous_residual_drop_text
            else:
                ori_x = x.clone()
                for block in self.blocks:
                    x = block(x, **kwargs)
                self.previous_residual_drop_text = x - ori_x
        else:
            if not should_calc_uncond:
                x +=  self.previous_residual_uncond
            else:
                ori_x = x.clone()
                for block in self.blocks:
                    x = block(x, **kwargs)
                self.previous_residual_uncond = x - ori_x
    else:
        for block in self.blocks:
            x = block(x, **kwargs)

    # head
    x = self.head(x, e)

    # Context Parallel
    x = get_sp_group().all_gather(x, dim=1)

    # unpatchify
    x = self.unpatchify(x, grid_sizes)
    if self.enable_teacache:
        self.cnt += 1
        if self.cnt >= self.num_steps:
            self.cnt = 0
        
    return torch.stack(x).float()


def usp_attn_forward_multitalk(self,
                     x,
                     seq_lens,
                     grid_sizes,
                     freqs,
                     dtype=torch.bfloat16,
                     ref_target_masks=None):
    b, s, n, d = *x.shape[:2], self.num_heads, self.head_dim
    half_dtypes = (torch.float16, torch.bfloat16)

    def half(x):
        return x if x.dtype in half_dtypes else x.to(dtype)

    # query, key, value function
    def qkv_fn(x):
        q = self.norm_q(self.q(x)).view(b, s, n, d)
        k = self.norm_k(self.k(x)).view(b, s, n, d)
        v = self.v(x).view(b, s, n, d)
        return q, k, v

    q, k, v = qkv_fn(x)
    q = rope_apply(q, grid_sizes, freqs)
    k = rope_apply(k, grid_sizes, freqs)


    x = xFuserLongContextAttention()(
        None,
        query=half(q),
        key=half(k),
        value=half(v),
        window_size=self.window_size)


    # output
    x = x.flatten(2)
    x = self.o(x)

    with torch.no_grad():
        x_ref_attn_map = get_attn_map_with_target(q.type_as(x), k.type_as(x), grid_sizes[0], 
                                            ref_target_masks=ref_target_masks, enable_sp=True) 

    return x, x_ref_attn_map




def usp_crossattn_multi_forward_multitalk(self, 
                                        x: torch.Tensor, 
                                        encoder_hidden_states: torch.Tensor,  # 1, 21, 64, C
                                        shape=None, 
                                        x_ref_attn_map=None,
                                        human_num=None) -> torch.Tensor:
        
        N_t, N_h, N_w = shape 
        sp_size = get_sequence_parallel_world_size()
        sp_rank = get_sequence_parallel_rank()
        audio_tokens_per_frame = 32
        visual_seqlen, frame_ids = split_token_counts_and_frame_ids(N_t, N_h * N_w, sp_size, sp_rank)
        encoder_hidden_states = encoder_hidden_states[:, min(frame_ids):max(frame_ids)+1, ...]
        encoder_hidden_states = rearrange(encoder_hidden_states, "B T N C -> B (T N) C")
        N_a = len(frame_ids)
        kv_seq = [audio_tokens_per_frame * human_num] * N_a

        if human_num == 1:
            return super(SingleStreamMutiAttention, self).forward(x, encoder_hidden_states, shape, enable_sp=True, kv_seq=kv_seq)


        # get q for hidden_state
        B, N, C = x.shape
        q = self.q_linear(x) 
        q_shape = (B, N, self.num_heads, self.head_dim) 
        q = q.view(q_shape).permute((0, 2, 1, 3))

        if self.qk_norm:
            q = self.q_norm(q)

        max_values = x_ref_attn_map.max(1).values[:, None, None] 
        min_values = x_ref_attn_map.min(1).values[:, None, None] 
        max_min_values = torch.cat([max_values, min_values], dim=2)
        max_min_values = get_sp_group().all_gather(max_min_values, dim=1)

        human1_max_value, human1_min_value = max_min_values[0, :, 0].max(), max_min_values[0, :, 1].min()
        human2_max_value, human2_min_value = max_min_values[1, :, 0].max(), max_min_values[1, :, 1].min()

        human1 = normalize_and_scale(x_ref_attn_map[0], (human1_min_value, human1_max_value), (self.rope_h1[0], self.rope_h1[1]))
        human2 = normalize_and_scale(x_ref_attn_map[1], (human2_min_value, human2_max_value), (self.rope_h2[0], self.rope_h2[1]))
        back   = torch.full((x_ref_attn_map.size(1),), self.rope_bak, dtype=human1.dtype).to(human1.device)
        max_indices = x_ref_attn_map.argmax(dim=0)
        normalized_map = torch.stack([human1, human2, back], dim=1)
        normalized_pos = normalized_map[range(x_ref_attn_map.size(1)), max_indices] # N 
        q = self.rope_1d(q, normalized_pos)
 
        encoder_kv = self.kv_linear(encoder_hidden_states) 
        encoder_kv_shape = (B, encoder_hidden_states.size(1), 2, self.num_heads, self.head_dim)
        encoder_kv = encoder_kv.view(encoder_kv_shape).permute((2, 0, 3, 1, 4)) 
        encoder_k, encoder_v = encoder_kv.unbind(0) # B H N C

        if self.qk_norm:
            encoder_k = self.add_k_norm(encoder_k)

        # position embedding for condition audio embeddings
        per_frame = torch.zeros(audio_tokens_per_frame * human_num, dtype=encoder_k.dtype).to(encoder_k.device)
        per_frame[:audio_tokens_per_frame] = (self.rope_h1[0] + self.rope_h1[1]) / 2
        per_frame[audio_tokens_per_frame:] = (self.rope_h2[0] + self.rope_h2[1]) / 2
        encoder_pos = torch.concat([per_frame]*N_a, dim=0)
        encoder_k = self.rope_1d(encoder_k, encoder_pos)

        # get attn
        q = rearrange(q, "B H M K -> B M H K")
        encoder_k = rearrange(encoder_k, "B H M K -> B M H K")
        encoder_v = rearrange(encoder_v, "B H M K -> B M H K")
        attn_bias = xformers.ops.fmha.attn_bias.BlockDiagonalMask.from_seqlens(visual_seqlen, kv_seq)
        x = xformers.ops.memory_efficient_attention(q, encoder_k, encoder_v, attn_bias=attn_bias, op=None,)
        x = rearrange(x, "B M H K -> B H M K")

        # linear transform
        x_output_shape = (B, N, C)
        x = x.transpose(1, 2) 
        x = x.reshape(x_output_shape) 
        x = self.proj(x) 
        x = self.proj_drop(x)

        return x