Track bundled vendor runtime sources

vendor/ComfyUI/comfy/ldm/wan/ar_model.py

@@ -0,0 +1,276 @@
+"""
+CausalWanModel: Wan 2.1 backbone with KV-cached causal self-attention for
+autoregressive (frame-by-frame) video generation via Causal Forcing.
+
+Weight-compatible with the standard WanModel -- same layer names, same shapes.
+The difference is purely in the forward pass: this model processes one temporal
+block at a time and maintains a KV cache across blocks.
+
+Reference: https://github.com/thu-ml/Causal-Forcing
+"""
+
+import torch
+import torch.nn as nn
+
+from comfy.ldm.modules.attention import optimized_attention
+from comfy.ldm.flux.math import apply_rope1
+from comfy.ldm.wan.model import (
+    sinusoidal_embedding_1d,
+    repeat_e,
+    WanModel,
+    WanAttentionBlock,
+)
+import comfy.ldm.common_dit
+import comfy.model_management
+
+
+class CausalWanSelfAttention(nn.Module):
+    """Self-attention with KV cache support for autoregressive inference."""
+
+    def __init__(self, dim, num_heads, window_size=(-1, -1), qk_norm=True,
+                 eps=1e-6, operation_settings={}):
+        assert dim % num_heads == 0
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+        self.qk_norm = qk_norm
+        self.eps = eps
+
+        ops = operation_settings.get("operations")
+        device = operation_settings.get("device")
+        dtype = operation_settings.get("dtype")
+
+        self.q = ops.Linear(dim, dim, device=device, dtype=dtype)
+        self.k = ops.Linear(dim, dim, device=device, dtype=dtype)
+        self.v = ops.Linear(dim, dim, device=device, dtype=dtype)
+        self.o = ops.Linear(dim, dim, device=device, dtype=dtype)
+        self.norm_q = ops.RMSNorm(dim, eps=eps, elementwise_affine=True, device=device, dtype=dtype) if qk_norm else nn.Identity()
+        self.norm_k = ops.RMSNorm(dim, eps=eps, elementwise_affine=True, device=device, dtype=dtype) if qk_norm else nn.Identity()
+
+    def forward(self, x, freqs, kv_cache=None, transformer_options={}):
+        b, s, n, d = *x.shape[:2], self.num_heads, self.head_dim
+
+        q = apply_rope1(self.norm_q(self.q(x)).view(b, s, n, d), freqs)
+        k = apply_rope1(self.norm_k(self.k(x)).view(b, s, n, d), freqs)
+        v = self.v(x).view(b, s, n, d)
+
+        if kv_cache is None:
+            x = optimized_attention(
+                q.view(b, s, n * d),
+                k.view(b, s, n * d),
+                v.view(b, s, n * d),
+                heads=self.num_heads,
+                transformer_options=transformer_options,
+            )
+        else:
+            end = kv_cache["end"]
+            new_end = end + s
+
+            # Roped K and plain V go into cache
+            kv_cache["k"][:, end:new_end] = k
+            kv_cache["v"][:, end:new_end] = v
+            kv_cache["end"] = new_end
+
+            x = optimized_attention(
+                q.view(b, s, n * d),
+                kv_cache["k"][:, :new_end].view(b, new_end, n * d),
+                kv_cache["v"][:, :new_end].view(b, new_end, n * d),
+                heads=self.num_heads,
+                transformer_options=transformer_options,
+            )
+
+        x = self.o(x)
+        return x
+
+
+class CausalWanAttentionBlock(WanAttentionBlock):
+    """Transformer block with KV-cached self-attention and cross-attention caching."""
+
+    def __init__(self, cross_attn_type, dim, ffn_dim, num_heads,
+                 window_size=(-1, -1), qk_norm=True, cross_attn_norm=False,
+                 eps=1e-6, operation_settings={}):
+        super().__init__(cross_attn_type, dim, ffn_dim, num_heads,
+                         window_size, qk_norm, cross_attn_norm, eps,
+                         operation_settings=operation_settings)
+        self.self_attn = CausalWanSelfAttention(
+            dim, num_heads, window_size, qk_norm, eps,
+            operation_settings=operation_settings)
+
+    def forward(self, x, e, freqs, context, context_img_len=257,
+                kv_cache=None, crossattn_cache=None, transformer_options={}):
+        if e.ndim < 4:
+            e = (comfy.model_management.cast_to(self.modulation, dtype=x.dtype, device=x.device) + e).chunk(6, dim=1)
+        else:
+            e = (comfy.model_management.cast_to(self.modulation, dtype=x.dtype, device=x.device).unsqueeze(0) + e).unbind(2)
+
+        # Self-attention with optional KV cache
+        x = x.contiguous()
+        y = self.self_attn(
+            torch.addcmul(repeat_e(e[0], x), self.norm1(x), 1 + repeat_e(e[1], x)),
+            freqs, kv_cache=kv_cache, transformer_options=transformer_options)
+        x = torch.addcmul(x, y, repeat_e(e[2], x))
+        del y
+
+        # Cross-attention with optional caching
+        if crossattn_cache is not None and crossattn_cache.get("is_init"):
+            q = self.cross_attn.norm_q(self.cross_attn.q(self.norm3(x)))
+            x_ca = optimized_attention(
+                q, crossattn_cache["k"], crossattn_cache["v"],
+                heads=self.num_heads, transformer_options=transformer_options)
+            x = x + self.cross_attn.o(x_ca)
+        else:
+            x = x + self.cross_attn(self.norm3(x), context, context_img_len=context_img_len, transformer_options=transformer_options)
+            if crossattn_cache is not None:
+                crossattn_cache["k"] = self.cross_attn.norm_k(self.cross_attn.k(context))
+                crossattn_cache["v"] = self.cross_attn.v(context)
+                crossattn_cache["is_init"] = True
+
+        # FFN
+        y = self.ffn(torch.addcmul(repeat_e(e[3], x), self.norm2(x), 1 + repeat_e(e[4], x)))
+        x = torch.addcmul(x, y, repeat_e(e[5], x))
+        return x
+
+
+class CausalWanModel(WanModel):
+    """
+    Wan 2.1 diffusion backbone with causal KV-cache support.
+
+    Same weight structure as WanModel -- loads identical state dicts.
+    Adds forward_block() for frame-by-frame autoregressive inference.
+    """
+
+    def __init__(self,
+                 model_type='t2v',
+                 patch_size=(1, 2, 2),
+                 text_len=512,
+                 in_dim=16,
+                 dim=2048,
+                 ffn_dim=8192,
+                 freq_dim=256,
+                 text_dim=4096,
+                 out_dim=16,
+                 num_heads=16,
+                 num_layers=32,
+                 window_size=(-1, -1),
+                 qk_norm=True,
+                 cross_attn_norm=True,
+                 eps=1e-6,
+                 image_model=None,
+                 device=None,
+                 dtype=None,
+                 operations=None):
+        super().__init__(
+            model_type=model_type, patch_size=patch_size, text_len=text_len,
+            in_dim=in_dim, dim=dim, ffn_dim=ffn_dim, freq_dim=freq_dim,
+            text_dim=text_dim, out_dim=out_dim, num_heads=num_heads,
+            num_layers=num_layers, window_size=window_size, qk_norm=qk_norm,
+            cross_attn_norm=cross_attn_norm, eps=eps, image_model=image_model,
+            wan_attn_block_class=CausalWanAttentionBlock,
+            device=device, dtype=dtype, operations=operations)
+
+    def forward_block(self, x, timestep, context, start_frame,
+                      kv_caches, crossattn_caches, clip_fea=None):
+        """
+        Forward one temporal block for autoregressive inference.
+
+        Args:
+            x: [B, C, block_frames, H, W] input latent for the current block
+            timestep: [B, block_frames] per-frame timesteps
+            context: [B, L, text_dim] raw text embeddings (pre-text_embedding)
+            start_frame: temporal frame index for RoPE offset
+            kv_caches: list of per-layer KV cache dicts
+            crossattn_caches: list of per-layer cross-attention cache dicts
+            clip_fea: optional CLIP features for I2V
+
+        Returns:
+            flow_pred: [B, C_out, block_frames, H, W] flow prediction
+        """
+        x = comfy.ldm.common_dit.pad_to_patch_size(x, self.patch_size)
+        bs, c, t, h, w = x.shape
+
+        x = self.patch_embedding(x.float()).to(x.dtype)
+        grid_sizes = x.shape[2:]
+        x = x.flatten(2).transpose(1, 2)
+
+        # Per-frame time embedding
+        e = self.time_embedding(
+            sinusoidal_embedding_1d(self.freq_dim, timestep.flatten()).to(dtype=x.dtype))
+        e = e.reshape(timestep.shape[0], -1, e.shape[-1])
+        e0 = self.time_projection(e).unflatten(2, (6, self.dim))
+
+        # Text embedding (reuses crossattn_cache after first block)
+        context = self.text_embedding(context)
+
+        context_img_len = None
+        if clip_fea is not None and self.img_emb is not None:
+            context_clip = self.img_emb(clip_fea)
+            context = torch.concat([context_clip, context], dim=1)
+            context_img_len = clip_fea.shape[-2]
+
+        # RoPE for current block's temporal position
+        freqs = self.rope_encode(t, h, w, t_start=start_frame, device=x.device, dtype=x.dtype)
+
+        # Transformer blocks
+        for i, block in enumerate(self.blocks):
+            x = block(x, e=e0, freqs=freqs, context=context,
+                      context_img_len=context_img_len,
+                      kv_cache=kv_caches[i],
+                      crossattn_cache=crossattn_caches[i])
+
+        # Head
+        x = self.head(x, e)
+
+        # Unpatchify
+        x = self.unpatchify(x, grid_sizes)
+        return x[:, :, :t, :h, :w]
+
+    def init_kv_caches(self, batch_size, max_seq_len, device, dtype):
+        """Create fresh KV caches for all layers."""
+        caches = []
+        for _ in range(self.num_layers):
+            caches.append({
+                "k": torch.zeros(batch_size, max_seq_len, self.num_heads, self.head_dim, device=device, dtype=dtype),
+                "v": torch.zeros(batch_size, max_seq_len, self.num_heads, self.head_dim, device=device, dtype=dtype),
+                "end": 0,
+            })
+        return caches
+
+    def init_crossattn_caches(self, batch_size, device, dtype):
+        """Create fresh cross-attention caches for all layers."""
+        caches = []
+        for _ in range(self.num_layers):
+            caches.append({"is_init": False})
+        return caches
+
+    def reset_kv_caches(self, kv_caches):
+        """Reset KV caches to empty (reuse allocated memory)."""
+        for cache in kv_caches:
+            cache["end"] = 0
+
+    def reset_crossattn_caches(self, crossattn_caches):
+        """Reset cross-attention caches."""
+        for cache in crossattn_caches:
+            cache["is_init"] = False
+
+    @property
+    def head_dim(self):
+        return self.dim // self.num_heads
+
+    def forward(self, x, timestep, context, clip_fea=None, time_dim_concat=None, transformer_options={}, **kwargs):
+        ar_state = transformer_options.get("ar_state")
+        if ar_state is not None:
+            bs = x.shape[0]
+            block_frames = x.shape[2]
+            t_per_frame = timestep.unsqueeze(1).expand(bs, block_frames)
+            return self.forward_block(
+                x=x, timestep=t_per_frame, context=context,
+                start_frame=ar_state["start_frame"],
+                kv_caches=ar_state["kv_caches"],
+                crossattn_caches=ar_state["crossattn_caches"],
+                clip_fea=clip_fea,
+            )
+
+        return super().forward(x, timestep, context, clip_fea=clip_fea,
+                               time_dim_concat=time_dim_concat,
+                               transformer_options=transformer_options, **kwargs)

vendor/ComfyUI/comfy/ldm/wan/model_animate.py

@@ -0,0 +1,551 @@
+from torch import nn
+import torch
+from typing import Tuple, Optional
+from einops import rearrange
+import torch.nn.functional as F
+import math
+from .model import WanModel, sinusoidal_embedding_1d
+from comfy.ldm.modules.attention import optimized_attention
+import comfy.model_management
+
+class CausalConv1d(nn.Module):
+
+    def __init__(self, chan_in, chan_out, kernel_size=3, stride=1, dilation=1, pad_mode="replicate", operations=None, **kwargs):
+        super().__init__()
+
+        self.pad_mode = pad_mode
+        padding = (kernel_size - 1, 0)  # T
+        self.time_causal_padding = padding
+
+        self.conv = operations.Conv1d(chan_in, chan_out, kernel_size, stride=stride, dilation=dilation, **kwargs)
+
+    def forward(self, x):
+        x = F.pad(x, self.time_causal_padding, mode=self.pad_mode)
+        return self.conv(x)
+
+
+class FaceEncoder(nn.Module):
+    def __init__(self, in_dim: int, hidden_dim: int, num_heads=int, dtype=None, device=None, operations=None):
+        factory_kwargs = {"dtype": dtype, "device": device}
+        super().__init__()
+
+        self.num_heads = num_heads
+        self.conv1_local = CausalConv1d(in_dim, 1024 * num_heads, 3, stride=1, operations=operations, **factory_kwargs)
+        self.norm1 = operations.LayerNorm(hidden_dim // 8, elementwise_affine=False, eps=1e-6, **factory_kwargs)
+        self.act = nn.SiLU()
+        self.conv2 = CausalConv1d(1024, 1024, 3, stride=2, operations=operations, **factory_kwargs)
+        self.conv3 = CausalConv1d(1024, 1024, 3, stride=2, operations=operations, **factory_kwargs)
+
+        self.out_proj = operations.Linear(1024, hidden_dim, **factory_kwargs)
+        self.norm1 = operations.LayerNorm(1024, elementwise_affine=False, eps=1e-6, **factory_kwargs)
+
+        self.norm2 = operations.LayerNorm(1024, elementwise_affine=False, eps=1e-6, **factory_kwargs)
+
+        self.norm3 = operations.LayerNorm(1024, elementwise_affine=False, eps=1e-6, **factory_kwargs)
+
+        self.padding_tokens = nn.Parameter(torch.empty(1, 1, 1, hidden_dim, **factory_kwargs))
+
+    def forward(self, x):
+
+        x = rearrange(x, "b t c -> b c t")
+        b, c, t = x.shape
+
+        x = self.conv1_local(x)
+        x = rearrange(x, "b (n c) t -> (b n) t c", n=self.num_heads)
+
+        x = self.norm1(x)
+        x = self.act(x)
+        x = rearrange(x, "b t c -> b c t")
+        x = self.conv2(x)
+        x = rearrange(x, "b c t -> b t c")
+        x = self.norm2(x)
+        x = self.act(x)
+        x = rearrange(x, "b t c -> b c t")
+        x = self.conv3(x)
+        x = rearrange(x, "b c t -> b t c")
+        x = self.norm3(x)
+        x = self.act(x)
+        x = self.out_proj(x)
+        x = rearrange(x, "(b n) t c -> b t n c", b=b)
+        padding = comfy.model_management.cast_to(self.padding_tokens, dtype=x.dtype, device=x.device).repeat(b, x.shape[1], 1, 1)
+        x = torch.cat([x, padding], dim=-2)
+        x_local = x.clone()
+
+        return x_local
+
+
+def get_norm_layer(norm_layer, operations=None):
+    """
+    Get the normalization layer.
+
+    Args:
+        norm_layer (str): The type of normalization layer.
+
+    Returns:
+        norm_layer (nn.Module): The normalization layer.
+    """
+    if norm_layer == "layer":
+        return operations.LayerNorm
+    elif norm_layer == "rms":
+        return operations.RMSNorm
+    else:
+        raise NotImplementedError(f"Norm layer {norm_layer} is not implemented")
+
+
+class FaceAdapter(nn.Module):
+    def __init__(
+        self,
+        hidden_dim: int,
+        heads_num: int,
+        qk_norm: bool = True,
+        qk_norm_type: str = "rms",
+        num_adapter_layers: int = 1,
+        dtype=None, device=None, operations=None
+    ):
+
+        factory_kwargs = {"dtype": dtype, "device": device}
+        super().__init__()
+        self.hidden_size = hidden_dim
+        self.heads_num = heads_num
+        self.fuser_blocks = nn.ModuleList(
+            [
+                FaceBlock(
+                    self.hidden_size,
+                    self.heads_num,
+                    qk_norm=qk_norm,
+                    qk_norm_type=qk_norm_type,
+                    operations=operations,
+                    **factory_kwargs,
+                )
+                for _ in range(num_adapter_layers)
+            ]
+        )
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        motion_embed: torch.Tensor,
+        idx: int,
+        freqs_cis_q: Tuple[torch.Tensor, torch.Tensor] = None,
+        freqs_cis_k: Tuple[torch.Tensor, torch.Tensor] = None,
+    ) -> torch.Tensor:
+
+        return self.fuser_blocks[idx](x, motion_embed, freqs_cis_q, freqs_cis_k)
+
+
+
+class FaceBlock(nn.Module):
+    def __init__(
+        self,
+        hidden_size: int,
+        heads_num: int,
+        qk_norm: bool = True,
+        qk_norm_type: str = "rms",
+        qk_scale: float = None,
+        dtype: Optional[torch.dtype] = None,
+        device: Optional[torch.device] = None,
+        operations=None
+    ):
+        factory_kwargs = {"device": device, "dtype": dtype}
+        super().__init__()
+
+        self.deterministic = False
+        self.hidden_size = hidden_size
+        self.heads_num = heads_num
+        head_dim = hidden_size // heads_num
+        self.scale = qk_scale or head_dim**-0.5
+
+        self.linear1_kv = operations.Linear(hidden_size, hidden_size * 2, **factory_kwargs)
+        self.linear1_q = operations.Linear(hidden_size, hidden_size, **factory_kwargs)
+
+        self.linear2 = operations.Linear(hidden_size, hidden_size, **factory_kwargs)
+
+        qk_norm_layer = get_norm_layer(qk_norm_type, operations=operations)
+        self.q_norm = (
+            qk_norm_layer(head_dim, elementwise_affine=True, eps=1e-6, **factory_kwargs) if qk_norm else nn.Identity()
+        )
+        self.k_norm = (
+            qk_norm_layer(head_dim, elementwise_affine=True, eps=1e-6, **factory_kwargs) if qk_norm else nn.Identity()
+        )
+
+        self.pre_norm_feat = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, **factory_kwargs)
+
+        self.pre_norm_motion = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, **factory_kwargs)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        motion_vec: torch.Tensor,
+        motion_mask: Optional[torch.Tensor] = None,
+        # use_context_parallel=False,
+    ) -> torch.Tensor:
+
+        B, T, N, C = motion_vec.shape
+        T_comp = T
+
+        x_motion = self.pre_norm_motion(motion_vec)
+        x_feat = self.pre_norm_feat(x)
+
+        kv = self.linear1_kv(x_motion)
+        q = self.linear1_q(x_feat)
+
+        k, v = rearrange(kv, "B L N (K H D) -> K B L N H D", K=2, H=self.heads_num)
+        q = rearrange(q, "B S (H D) -> B S H D", H=self.heads_num)
+
+        # Apply QK-Norm if needed.
+        q = self.q_norm(q).to(v)
+        k = self.k_norm(k).to(v)
+
+        k = rearrange(k, "B L N H D -> (B L) N H D")
+        v = rearrange(v, "B L N H D -> (B L) N H D")
+
+        q = rearrange(q, "B (L S) H D -> (B L) S (H D)", L=T_comp)
+
+        attn = optimized_attention(q, k, v, heads=self.heads_num)
+
+        attn = rearrange(attn, "(B L) S C -> B (L S) C", L=T_comp)
+
+        output = self.linear2(attn)
+
+        if motion_mask is not None:
+            output = output * rearrange(motion_mask, "B T H W -> B (T H W)").unsqueeze(-1)
+
+        return output
+
+# https://github.com/XPixelGroup/BasicSR/blob/8d56e3a045f9fb3e1d8872f92ee4a4f07f886b0a/basicsr/ops/upfirdn2d/upfirdn2d.py#L162
+def upfirdn2d_native(input, kernel, up_x, up_y, down_x, down_y, pad_x0, pad_x1, pad_y0, pad_y1):
+    _, minor, in_h, in_w = input.shape
+    kernel_h, kernel_w = kernel.shape
+
+    out = input.view(-1, minor, in_h, 1, in_w, 1)
+    out = F.pad(out, [0, up_x - 1, 0, 0, 0, up_y - 1, 0, 0])
+    out = out.view(-1, minor, in_h * up_y, in_w * up_x)
+
+    out = F.pad(out, [max(pad_x0, 0), max(pad_x1, 0), max(pad_y0, 0), max(pad_y1, 0)])
+    out = out[:, :, max(-pad_y0, 0): out.shape[2] - max(-pad_y1, 0), max(-pad_x0, 0): out.shape[3] - max(-pad_x1, 0)]
+
+    out = out.reshape([-1, 1, in_h * up_y + pad_y0 + pad_y1, in_w * up_x + pad_x0 + pad_x1])
+    w = torch.flip(kernel, [0, 1]).view(1, 1, kernel_h, kernel_w)
+    out = F.conv2d(out, w)
+    out = out.reshape(-1, minor, in_h * up_y + pad_y0 + pad_y1 - kernel_h + 1, in_w * up_x + pad_x0 + pad_x1 - kernel_w + 1)
+    return out[:, :, ::down_y, ::down_x]
+
+def upfirdn2d(input, kernel, up=1, down=1, pad=(0, 0)):
+    return upfirdn2d_native(input, kernel, up, up, down, down, pad[0], pad[1], pad[0], pad[1])
+
+# https://github.com/XPixelGroup/BasicSR/blob/8d56e3a045f9fb3e1d8872f92ee4a4f07f886b0a/basicsr/ops/fused_act/fused_act.py#L81
+class FusedLeakyReLU(torch.nn.Module):
+    def __init__(self, channel, negative_slope=0.2, scale=2 ** 0.5, dtype=None, device=None):
+        super().__init__()
+        self.bias = torch.nn.Parameter(torch.empty(1, channel, 1, 1, dtype=dtype, device=device))
+        self.negative_slope = negative_slope
+        self.scale = scale
+
+    def forward(self, input):
+        return fused_leaky_relu(input, comfy.model_management.cast_to(self.bias, device=input.device, dtype=input.dtype), self.negative_slope, self.scale)
+
+def fused_leaky_relu(input, bias, negative_slope=0.2, scale=2 ** 0.5):
+    return F.leaky_relu(input + bias, negative_slope) * scale
+
+class Blur(torch.nn.Module):
+    def __init__(self, kernel, pad, dtype=None, device=None):
+        super().__init__()
+        kernel = torch.tensor(kernel, dtype=dtype, device=device)
+        kernel = kernel[None, :] * kernel[:, None]
+        kernel = kernel / kernel.sum()
+        self.register_buffer('kernel', kernel)
+        self.pad = pad
+
+    def forward(self, input):
+        return upfirdn2d(input, comfy.model_management.cast_to(self.kernel, dtype=input.dtype, device=input.device), pad=self.pad)
+
+#https://github.com/XPixelGroup/BasicSR/blob/8d56e3a045f9fb3e1d8872f92ee4a4f07f886b0a/basicsr/archs/stylegan2_arch.py#L590
+class ScaledLeakyReLU(torch.nn.Module):
+    def __init__(self, negative_slope=0.2):
+        super().__init__()
+        self.negative_slope = negative_slope
+
+    def forward(self, input):
+        return F.leaky_relu(input, negative_slope=self.negative_slope)
+
+# https://github.com/XPixelGroup/BasicSR/blob/8d56e3a045f9fb3e1d8872f92ee4a4f07f886b0a/basicsr/archs/stylegan2_arch.py#L605
+class EqualConv2d(torch.nn.Module):
+    def __init__(self, in_channel, out_channel, kernel_size, stride=1, padding=0, bias=True, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.weight = torch.nn.Parameter(torch.empty(out_channel, in_channel, kernel_size, kernel_size, device=device, dtype=dtype))
+        self.scale = 1 / math.sqrt(in_channel * kernel_size ** 2)
+        self.stride = stride
+        self.padding = padding
+        self.bias = torch.nn.Parameter(torch.empty(out_channel, device=device, dtype=dtype)) if bias else None
+
+    def forward(self, input):
+        if self.bias is None:
+            bias = None
+        else:
+            bias = comfy.model_management.cast_to(self.bias, device=input.device, dtype=input.dtype)
+
+        return F.conv2d(input, comfy.model_management.cast_to(self.weight, device=input.device, dtype=input.dtype) * self.scale, bias=bias, stride=self.stride, padding=self.padding)
+
+# https://github.com/XPixelGroup/BasicSR/blob/8d56e3a045f9fb3e1d8872f92ee4a4f07f886b0a/basicsr/archs/stylegan2_arch.py#L134
+class EqualLinear(torch.nn.Module):
+    def __init__(self, in_dim, out_dim, bias=True, bias_init=0, lr_mul=1, activation=None, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.weight = torch.nn.Parameter(torch.empty(out_dim, in_dim, device=device, dtype=dtype))
+        self.bias = torch.nn.Parameter(torch.empty(out_dim, device=device, dtype=dtype)) if bias else None
+        self.activation = activation
+        self.scale = (1 / math.sqrt(in_dim)) * lr_mul
+        self.lr_mul = lr_mul
+
+    def forward(self, input):
+        if self.bias is None:
+            bias = None
+        else:
+            bias = comfy.model_management.cast_to(self.bias, device=input.device, dtype=input.dtype) * self.lr_mul
+
+        if self.activation:
+            out = F.linear(input, comfy.model_management.cast_to(self.weight, device=input.device, dtype=input.dtype) * self.scale)
+            return fused_leaky_relu(out, bias)
+        return F.linear(input, comfy.model_management.cast_to(self.weight, device=input.device, dtype=input.dtype) * self.scale, bias=bias)
+
+# https://github.com/XPixelGroup/BasicSR/blob/8d56e3a045f9fb3e1d8872f92ee4a4f07f886b0a/basicsr/archs/stylegan2_arch.py#L654
+class ConvLayer(torch.nn.Sequential):
+    def __init__(self, in_channel, out_channel, kernel_size, downsample=False, blur_kernel=[1, 3, 3, 1], bias=True, activate=True, dtype=None, device=None, operations=None):
+        layers = []
+
+        if downsample:
+            factor = 2
+            p = (len(blur_kernel) - factor) + (kernel_size - 1)
+            layers.append(Blur(blur_kernel, pad=((p + 1) // 2, p // 2)))
+            stride, padding = 2, 0
+        else:
+            stride, padding = 1, kernel_size // 2
+
+        layers.append(EqualConv2d(in_channel, out_channel, kernel_size, padding=padding, stride=stride, bias=bias and not activate, dtype=dtype, device=device, operations=operations))
+
+        if activate:
+            layers.append(FusedLeakyReLU(out_channel) if bias else ScaledLeakyReLU(0.2))
+
+        super().__init__(*layers)
+
+# https://github.com/XPixelGroup/BasicSR/blob/8d56e3a045f9fb3e1d8872f92ee4a4f07f886b0a/basicsr/archs/stylegan2_arch.py#L704
+class ResBlock(torch.nn.Module):
+    def __init__(self, in_channel, out_channel, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.conv1 = ConvLayer(in_channel, in_channel, 3, dtype=dtype, device=device, operations=operations)
+        self.conv2 = ConvLayer(in_channel, out_channel, 3, downsample=True, dtype=dtype, device=device, operations=operations)
+        self.skip = ConvLayer(in_channel, out_channel, 1, downsample=True, activate=False, bias=False, dtype=dtype, device=device, operations=operations)
+
+    def forward(self, input):
+        out = self.conv2(self.conv1(input))
+        skip = self.skip(input)
+        return (out + skip) / math.sqrt(2)
+
+
+class EncoderApp(torch.nn.Module):
+    def __init__(self, w_dim=512, dtype=None, device=None, operations=None):
+        super().__init__()
+        kwargs = {"device": device, "dtype": dtype, "operations": operations}
+
+        self.convs = torch.nn.ModuleList([
+            ConvLayer(3, 32, 1, **kwargs), ResBlock(32, 64, **kwargs),
+            ResBlock(64, 128, **kwargs), ResBlock(128, 256, **kwargs),
+            ResBlock(256, 512, **kwargs), ResBlock(512, 512, **kwargs),
+            ResBlock(512, 512, **kwargs), ResBlock(512, 512, **kwargs),
+            EqualConv2d(512, w_dim, 4, padding=0, bias=False, **kwargs)
+        ])
+
+    def forward(self, x):
+        h = x
+        for conv in self.convs:
+            h = conv(h)
+        return h.squeeze(-1).squeeze(-1)
+
+class Encoder(torch.nn.Module):
+    def __init__(self, dim=512, motion_dim=20, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.net_app = EncoderApp(dim, dtype=dtype, device=device, operations=operations)
+        self.fc = torch.nn.Sequential(*[EqualLinear(dim, dim, dtype=dtype, device=device, operations=operations) for _ in range(4)] + [EqualLinear(dim, motion_dim, dtype=dtype, device=device, operations=operations)])
+
+    def encode_motion(self, x):
+        return self.fc(self.net_app(x))
+
+class Direction(torch.nn.Module):
+    def __init__(self, motion_dim, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.weight = torch.nn.Parameter(torch.empty(512, motion_dim, device=device, dtype=dtype))
+        self.motion_dim = motion_dim
+
+    def forward(self, input):
+        stabilized_weight = comfy.model_management.cast_to(self.weight, device=input.device, dtype=input.dtype) + 1e-8 * torch.eye(512, self.motion_dim, device=input.device, dtype=input.dtype)
+        Q, _ = torch.linalg.qr(stabilized_weight.float())
+        if input is None:
+            return Q
+        return torch.sum(input.unsqueeze(-1) * Q.T.to(input.dtype), dim=1)
+
+class Synthesis(torch.nn.Module):
+    def __init__(self, motion_dim, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.direction = Direction(motion_dim, dtype=dtype, device=device, operations=operations)
+
+class Generator(torch.nn.Module):
+    def __init__(self, style_dim=512, motion_dim=20, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.enc = Encoder(style_dim, motion_dim, dtype=dtype, device=device, operations=operations)
+        self.dec = Synthesis(motion_dim, dtype=dtype, device=device, operations=operations)
+
+    def get_motion(self, img):
+        motion_feat = self.enc.encode_motion(img)
+        return self.dec.direction(motion_feat)
+
+class AnimateWanModel(WanModel):
+    r"""
+    Wan diffusion backbone supporting both text-to-video and image-to-video.
+    """
+
+    def __init__(self,
+                 model_type='animate',
+                 patch_size=(1, 2, 2),
+                 text_len=512,
+                 in_dim=16,
+                 dim=2048,
+                 ffn_dim=8192,
+                 freq_dim=256,
+                 text_dim=4096,
+                 out_dim=16,
+                 num_heads=16,
+                 num_layers=32,
+                 window_size=(-1, -1),
+                 qk_norm=True,
+                 cross_attn_norm=True,
+                 eps=1e-6,
+                 flf_pos_embed_token_number=None,
+                 motion_encoder_dim=512,
+                 image_model=None,
+                 device=None,
+                 dtype=None,
+                 operations=None,
+                 ):
+
+        super().__init__(model_type='i2v', patch_size=patch_size, text_len=text_len, in_dim=in_dim, dim=dim, ffn_dim=ffn_dim, freq_dim=freq_dim, text_dim=text_dim, out_dim=out_dim, num_heads=num_heads, num_layers=num_layers, window_size=window_size, qk_norm=qk_norm, cross_attn_norm=cross_attn_norm, eps=eps, flf_pos_embed_token_number=flf_pos_embed_token_number, image_model=image_model, device=device, dtype=dtype, operations=operations)
+
+        self.pose_patch_embedding = operations.Conv3d(
+            16, dim, kernel_size=patch_size, stride=patch_size, device=device, dtype=dtype
+        )
+
+        self.motion_encoder = Generator(style_dim=512, motion_dim=20, device=device, dtype=dtype, operations=operations)
+
+        self.face_adapter = FaceAdapter(
+            heads_num=self.num_heads,
+            hidden_dim=self.dim,
+            num_adapter_layers=self.num_layers // 5,
+            device=device, dtype=dtype, operations=operations
+        )
+
+        self.face_encoder = FaceEncoder(
+            in_dim=motion_encoder_dim,
+            hidden_dim=self.dim,
+            num_heads=4,
+            device=device, dtype=dtype, operations=operations
+        )
+
+    def after_patch_embedding(self, x, pose_latents, face_pixel_values):
+        if pose_latents is not None:
+            pose_latents = self.pose_patch_embedding(pose_latents)
+            x[:, :, 1:pose_latents.shape[2] + 1] += pose_latents[:, :, :x.shape[2] - 1]
+
+        if face_pixel_values is None:
+            return x, None
+
+        b, c, T, h, w = face_pixel_values.shape
+        face_pixel_values = rearrange(face_pixel_values, "b c t h w -> (b t) c h w")
+        encode_bs = 8
+        face_pixel_values_tmp = []
+        for i in range(math.ceil(face_pixel_values.shape[0] / encode_bs)):
+            face_pixel_values_tmp.append(self.motion_encoder.get_motion(face_pixel_values[i * encode_bs: (i + 1) * encode_bs]))
+
+        motion_vec = torch.cat(face_pixel_values_tmp)
+
+        motion_vec = rearrange(motion_vec, "(b t) c -> b t c", t=T)
+        motion_vec = self.face_encoder(motion_vec)
+
+        B, L, H, C = motion_vec.shape
+        pad_face = torch.zeros(B, 1, H, C).type_as(motion_vec)
+        motion_vec = torch.cat([pad_face, motion_vec], dim=1)
+
+        if motion_vec.shape[1] < x.shape[2]:
+            B, L, H, C = motion_vec.shape
+            pad = torch.zeros(B, x.shape[2] - motion_vec.shape[1], H, C).type_as(motion_vec)
+            motion_vec = torch.cat([motion_vec, pad], dim=1)
+        else:
+            motion_vec = motion_vec[:, :x.shape[2]]
+        return x, motion_vec
+
+    def forward_orig(
+        self,
+        x,
+        t,
+        context,
+        clip_fea=None,
+        pose_latents=None,
+        face_pixel_values=None,
+        freqs=None,
+        transformer_options={},
+        **kwargs,
+    ):
+        # embeddings
+        x = self.patch_embedding(x.float()).to(x.dtype)
+        x, motion_vec = self.after_patch_embedding(x, pose_latents, face_pixel_values)
+        grid_sizes = x.shape[2:]
+        x = x.flatten(2).transpose(1, 2)
+
+        # time embeddings
+        e = self.time_embedding(
+            sinusoidal_embedding_1d(self.freq_dim, t.flatten()).to(dtype=x[0].dtype))
+        e = e.reshape(t.shape[0], -1, e.shape[-1])
+        e0 = self.time_projection(e).unflatten(2, (6, self.dim))
+
+        full_ref = None
+        if self.ref_conv is not None:
+            full_ref = kwargs.get("reference_latent", None)
+            if full_ref is not None:
+                full_ref = self.ref_conv(full_ref).flatten(2).transpose(1, 2)
+                x = torch.concat((full_ref, x), dim=1)
+
+        # context
+        context = self.text_embedding(context)
+
+        context_img_len = None
+        if clip_fea is not None:
+            if self.img_emb is not None:
+                context_clip = self.img_emb(clip_fea)  # bs x 257 x dim
+                context = torch.concat([context_clip, context], dim=1)
+            context_img_len = clip_fea.shape[-2]
+
+        patches_replace = transformer_options.get("patches_replace", {})
+        blocks_replace = patches_replace.get("dit", {})
+        transformer_options["total_blocks"] = len(self.blocks)
+        transformer_options["block_type"] = "double"
+        for i, block in enumerate(self.blocks):
+            transformer_options["block_index"] = i
+            if ("double_block", i) in blocks_replace:
+                def block_wrap(args):
+                    out = {}
+                    out["img"] = block(args["img"], context=args["txt"], e=args["vec"], freqs=args["pe"], context_img_len=context_img_len, transformer_options=args["transformer_options"])
+                    return out
+                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": e0, "pe": freqs, "transformer_options": transformer_options}, {"original_block": block_wrap})
+                x = out["img"]
+            else:
+                x = block(x, e=e0, freqs=freqs, context=context, context_img_len=context_img_len, transformer_options=transformer_options)
+
+            if i % 5 == 0 and motion_vec is not None:
+                x = x + self.face_adapter.fuser_blocks[i // 5](x, motion_vec)
+
+        # head
+        x = self.head(x, e)
+
+        if full_ref is not None:
+            x = x[:, full_ref.shape[1]:]
+
+        # unpatchify
+        x = self.unpatchify(x, grid_sizes)
+        return x

vendor/ComfyUI/comfy/ldm/wan/model_multitalk.py

@@ -0,0 +1,500 @@
+import torch
+from einops import rearrange, repeat
+import comfy
+from comfy.ldm.modules.attention import optimized_attention
+
+
+def calculate_x_ref_attn_map(visual_q, ref_k, ref_target_masks, split_num=8):
+    scale = 1.0 / visual_q.shape[-1] ** 0.5
+    visual_q = visual_q.transpose(1, 2) * scale
+
+    B, H, x_seqlens, K = visual_q.shape
+
+    x_ref_attn_maps = []
+    for class_idx, ref_target_mask in enumerate(ref_target_masks):
+        ref_target_mask = ref_target_mask.view(1, 1, 1, -1)
+
+        x_ref_attnmap = torch.zeros(B, H, x_seqlens, device=visual_q.device, dtype=visual_q.dtype)
+        chunk_size = min(max(x_seqlens // split_num, 1), x_seqlens)
+
+        for i in range(0, x_seqlens, chunk_size):
+            end_i = min(i + chunk_size, x_seqlens)
+
+            attn_chunk = visual_q[:, :, i:end_i] @ ref_k.permute(0, 2, 3, 1)  # B, H, chunk, ref_seqlens
+
+            # Apply softmax
+            attn_max = attn_chunk.max(dim=-1, keepdim=True).values
+            attn_chunk = (attn_chunk - attn_max).exp()
+            attn_sum = attn_chunk.sum(dim=-1, keepdim=True)
+            attn_chunk = attn_chunk / (attn_sum + 1e-8)
+
+            # Apply mask and sum
+            masked_attn = attn_chunk * ref_target_mask
+            x_ref_attnmap[:, :, i:end_i] = masked_attn.sum(-1) / (ref_target_mask.sum() + 1e-8)
+
+            del attn_chunk, masked_attn
+
+        # Average across heads
+        x_ref_attnmap = x_ref_attnmap.mean(dim=1)  # B, x_seqlens
+        x_ref_attn_maps.append(x_ref_attnmap)
+
+    del visual_q, ref_k
+
+    return torch.cat(x_ref_attn_maps, dim=0)
+
+def get_attn_map_with_target(visual_q, ref_k, shape, ref_target_masks=None, split_num=2):
+    """Args:
+        query (torch.tensor): B M H K
+        key (torch.tensor): B M H K
+        shape (tuple): (N_t, N_h, N_w)
+        ref_target_masks: [B, N_h * N_w]
+    """
+
+    N_t, N_h, N_w = shape
+
+    x_seqlens = N_h * N_w
+    ref_k     = ref_k[:, :x_seqlens]
+    _, seq_lens, heads, _ = visual_q.shape
+    class_num, _ = ref_target_masks.shape
+    x_ref_attn_maps = torch.zeros(class_num, seq_lens).to(visual_q)
+
+    split_chunk = heads // split_num
+
+    for i in range(split_num):
+        x_ref_attn_maps_perhead = calculate_x_ref_attn_map(
+            visual_q[:, :, i*split_chunk:(i+1)*split_chunk, :],
+            ref_k[:, :, i*split_chunk:(i+1)*split_chunk, :],
+            ref_target_masks
+            )
+        x_ref_attn_maps += x_ref_attn_maps_perhead
+
+    return x_ref_attn_maps / split_num
+
+
+def normalize_and_scale(column, source_range, target_range, epsilon=1e-8):
+    source_min, source_max = source_range
+    new_min, new_max = target_range
+    normalized = (column - source_min) / (source_max - source_min + epsilon)
+    scaled = normalized * (new_max - new_min) + new_min
+    return scaled
+
+
+def rotate_half(x):
+    x = rearrange(x, "... (d r) -> ... d r", r=2)
+    x1, x2 = x.unbind(dim=-1)
+    x = torch.stack((-x2, x1), dim=-1)
+    return rearrange(x, "... d r -> ... (d r)")
+
+
+def get_audio_embeds(encoded_audio, audio_start, audio_end):
+    audio_embs = []
+    human_num = len(encoded_audio)
+    audio_frames = encoded_audio[0].shape[0]
+
+    indices = (torch.arange(4 + 1) - 2) * 1
+
+    for human_idx in range(human_num):
+        if audio_end > audio_frames: # in case of not enough audio for current window, pad with first audio frame as that's most likely silence
+            pad_len = audio_end - audio_frames
+            pad_shape = list(encoded_audio[human_idx].shape)
+            pad_shape[0] = pad_len
+            pad_tensor = encoded_audio[human_idx][:1].repeat(pad_len, *([1] * (encoded_audio[human_idx].dim() - 1)))
+            encoded_audio_in = torch.cat([encoded_audio[human_idx], pad_tensor], dim=0)
+        else:
+            encoded_audio_in = encoded_audio[human_idx]
+        center_indices = torch.arange(audio_start, audio_end, 1).unsqueeze(1) + indices.unsqueeze(0)
+        center_indices = torch.clamp(center_indices, min=0, max=encoded_audio_in.shape[0] - 1)
+        audio_emb = encoded_audio_in[center_indices].unsqueeze(0)
+        audio_embs.append(audio_emb)
+
+    return torch.cat(audio_embs, dim=0)
+
+
+def project_audio_features(audio_proj, encoded_audio, audio_start, audio_end):
+    audio_embs = get_audio_embeds(encoded_audio, audio_start, audio_end)
+
+    first_frame_audio_emb_s = audio_embs[:, :1, ...]
+    latter_frame_audio_emb = audio_embs[:, 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=4)
+
+    middle_index = audio_proj.seq_len // 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.cat([latter_first_frame_audio_emb, latter_middle_frame_audio_emb, latter_last_frame_audio_emb], dim=2)
+
+    audio_emb = audio_proj(first_frame_audio_emb_s, latter_frame_audio_emb_s)
+    audio_emb = torch.cat(audio_emb.split(1), dim=2)
+
+    return audio_emb
+
+
+class RotaryPositionalEmbedding1D(torch.nn.Module):
+    def __init__(self,
+                 head_dim,
+                 ):
+        super().__init__()
+        self.head_dim = head_dim
+        self.base = 10000
+
+    def precompute_freqs_cis_1d(self, pos_indices):
+        freqs = 1.0 / (self.base ** (torch.arange(0, self.head_dim, 2)[: (self.head_dim // 2)].float() / self.head_dim))
+        freqs = freqs.to(pos_indices.device)
+        freqs = torch.einsum("..., f -> ... f", pos_indices.float(), freqs)
+        freqs = repeat(freqs, "... n -> ... (n r)", r=2)
+        return freqs
+
+    def forward(self, x, pos_indices):
+        freqs_cis = self.precompute_freqs_cis_1d(pos_indices)
+
+        x_ = x.float()
+
+        freqs_cis = freqs_cis.float().to(x.device)
+        cos, sin = freqs_cis.cos(), freqs_cis.sin()
+        cos, sin = rearrange(cos, 'n d -> 1 1 n d'), rearrange(sin, 'n d -> 1 1 n d')
+        x_ = (x_ * cos) + (rotate_half(x_) * sin)
+
+        return x_.type_as(x)
+
+class SingleStreamAttention(torch.nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        encoder_hidden_states_dim: int,
+        num_heads: int,
+        qkv_bias: bool,
+        device=None, dtype=None, operations=None
+    ) -> None:
+        super().__init__()
+        self.dim = dim
+        self.encoder_hidden_states_dim = encoder_hidden_states_dim
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+
+        self.q_linear = operations.Linear(dim, dim, bias=qkv_bias, device=device, dtype=dtype)
+        self.proj = operations.Linear(dim, dim, device=device, dtype=dtype)
+        self.kv_linear = operations.Linear(encoder_hidden_states_dim, dim * 2, bias=qkv_bias, device=device, dtype=dtype)
+
+    def forward(self, x: torch.Tensor, encoder_hidden_states: torch.Tensor, shape=None) -> torch.Tensor:
+        N_t, N_h, N_w = shape
+
+        expected_tokens = N_t * N_h * N_w
+        actual_tokens = x.shape[1]
+        x_extra = None
+
+        if actual_tokens != expected_tokens:
+            x_extra = x[:, -N_h * N_w:, :]
+            x = x[:, :-N_h * N_w, :]
+            N_t = N_t - 1
+
+        B = x.shape[0]
+        S = N_h * N_w
+        x = x.view(B * N_t, S, self.dim)
+
+        # get q for hidden_state
+        q = self.q_linear(x).view(B * N_t, S, self.num_heads, self.head_dim)
+
+        # get kv from encoder_hidden_states # shape: (B, N, num_heads, head_dim)
+        kv = self.kv_linear(encoder_hidden_states)
+        encoder_k, encoder_v = kv.view(B * N_t, encoder_hidden_states.shape[1], 2, self.num_heads, self.head_dim).unbind(2)
+
+        #print("q.shape", q.shape) #torch.Size([21, 1024, 40, 128])
+        x = optimized_attention(
+            q.transpose(1, 2),
+            encoder_k.transpose(1, 2),
+            encoder_v.transpose(1, 2),
+            heads=self.num_heads, skip_reshape=True, skip_output_reshape=True).transpose(1, 2)
+
+        # linear transform
+        x = self.proj(x.reshape(B * N_t, S, self.dim))
+        x = x.view(B, N_t * S, self.dim)
+
+        if x_extra is not None:
+            x = torch.cat([x, torch.zeros_like(x_extra)], dim=1)
+
+        return x
+
+class SingleStreamMultiAttention(SingleStreamAttention):
+    def __init__(
+        self,
+        dim: int,
+        encoder_hidden_states_dim: int,
+        num_heads: int,
+        qkv_bias: bool,
+        class_range: int = 24,
+        class_interval: int = 4,
+        device=None, dtype=None, operations=None
+    ) -> None:
+        super().__init__(
+            dim=dim,
+            encoder_hidden_states_dim=encoder_hidden_states_dim,
+            num_heads=num_heads,
+            qkv_bias=qkv_bias,
+            device=device,
+            dtype=dtype,
+            operations=operations
+        )
+
+        # Rotary-embedding layout parameters
+        self.class_interval = class_interval
+        self.class_range = class_range
+        self.max_humans = self.class_range // self.class_interval
+
+        # Constant bucket used for background tokens
+        self.rope_bak = int(self.class_range // 2)
+
+        self.rope_1d = RotaryPositionalEmbedding1D(self.head_dim)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        shape=None,
+        x_ref_attn_map=None
+    ) -> torch.Tensor:
+        encoder_hidden_states = encoder_hidden_states.squeeze(0).to(x.device)
+        human_num = x_ref_attn_map.shape[0] if x_ref_attn_map is not None else 1
+        # Single-speaker fall-through
+        if human_num <= 1:
+            return super().forward(x, encoder_hidden_states, shape)
+
+        N_t, N_h, N_w = shape
+
+        x_extra = None
+        if x.shape[0] * N_t != encoder_hidden_states.shape[0]:
+            x_extra = x[:, -N_h * N_w:, :]
+            x = x[:, :-N_h * N_w, :]
+            N_t = N_t - 1
+        x = rearrange(x, "B (N_t S) C -> (B N_t) S C", N_t=N_t)
+
+        # Query projection
+        B, N, C = x.shape
+        q = self.q_linear(x)
+        q = q.view(B, N, self.num_heads, self.head_dim).permute(0, 2, 1, 3)
+
+        # Use `class_range` logic for 2 speakers
+        rope_h1 = (0, self.class_interval)
+        rope_h2 = (self.class_range - self.class_interval, self.class_range)
+        rope_bak = int(self.class_range // 2)
+
+        # Normalize and scale attention maps for each speaker
+        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)
+
+        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), rope_h1)
+        human2 = normalize_and_scale(x_ref_attn_map[1], (human2_min_value, human2_max_value), rope_h2)
+        back = torch.full((x_ref_attn_map.size(1),), rope_bak, dtype=human1.dtype, device=human1.device)
+
+        # Token-wise speaker dominance
+        max_indices = x_ref_attn_map.argmax(dim=0)
+        normalized_map = torch.stack([human1, human2, back], dim=1)
+        normalized_pos = normalized_map[torch.arange(x_ref_attn_map.size(1)), max_indices]
+
+        # Apply rotary to Q
+        q = rearrange(q, "(B N_t) H S C -> B H (N_t S) C", N_t=N_t)
+        q = self.rope_1d(q, normalized_pos)
+        q = rearrange(q, "B H (N_t S) C -> (B N_t) H S C", N_t=N_t)
+
+        # Keys / Values
+        _, N_a, _ = encoder_hidden_states.shape
+        encoder_kv = self.kv_linear(encoder_hidden_states)
+        encoder_kv = encoder_kv.view(B, N_a, 2, self.num_heads, self.head_dim).permute(2, 0, 3, 1, 4)
+        encoder_k, encoder_v = encoder_kv.unbind(0)
+
+        # Rotary for keys – assign centre of each speaker bucket to its context tokens
+        per_frame = torch.zeros(N_a, dtype=encoder_k.dtype, device=encoder_k.device)
+        per_frame[: per_frame.size(0) // 2] = (rope_h1[0] + rope_h1[1]) / 2
+        per_frame[per_frame.size(0) // 2 :] = (rope_h2[0] + rope_h2[1]) / 2
+        encoder_pos = torch.cat([per_frame] * N_t, dim=0)
+
+        encoder_k = rearrange(encoder_k, "(B N_t) H S C -> B H (N_t S) C", N_t=N_t)
+        encoder_k = self.rope_1d(encoder_k, encoder_pos)
+        encoder_k = rearrange(encoder_k, "B H (N_t S) C -> (B N_t) H S C", N_t=N_t)
+
+        # Final attention
+        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")
+
+        x = optimized_attention(
+            q.transpose(1, 2),
+            encoder_k.transpose(1, 2),
+            encoder_v.transpose(1, 2),
+            heads=self.num_heads, skip_reshape=True, skip_output_reshape=True).transpose(1, 2)
+
+        # Linear projection
+        x = x.reshape(B, N, C)
+        x = self.proj(x)
+
+        # Restore original layout
+        x = rearrange(x, "(B N_t) S C -> B (N_t S) C", N_t=N_t)
+        if x_extra is not None:
+            x = torch.cat([x, torch.zeros_like(x_extra)], dim=1)
+
+        return x
+
+
+class MultiTalkAudioProjModel(torch.nn.Module):
+    def __init__(
+        self,
+        seq_len: int = 5,
+        seq_len_vf: int = 12,
+        blocks: int = 12,
+        channels: int = 768,
+        intermediate_dim: int = 512,
+        out_dim: int = 768,
+        context_tokens: int = 32,
+        device=None, dtype=None, operations=None
+    ):
+        super().__init__()
+
+        self.seq_len = seq_len
+        self.blocks = blocks
+        self.channels = channels
+        self.input_dim = seq_len * blocks * channels
+        self.input_dim_vf = seq_len_vf * blocks * channels
+        self.intermediate_dim = intermediate_dim
+        self.context_tokens = context_tokens
+        self.out_dim = out_dim
+
+        # define multiple linear layers
+        self.proj1 = operations.Linear(self.input_dim, intermediate_dim, device=device, dtype=dtype)
+        self.proj1_vf = operations.Linear(self.input_dim_vf, intermediate_dim, device=device, dtype=dtype)
+        self.proj2 = operations.Linear(intermediate_dim, intermediate_dim, device=device, dtype=dtype)
+        self.proj3 = operations.Linear(intermediate_dim, context_tokens * out_dim, device=device, dtype=dtype)
+        self.norm = operations.LayerNorm(out_dim, device=device, dtype=dtype)
+
+    def forward(self, audio_embeds, audio_embeds_vf):
+        video_length = audio_embeds.shape[1] + audio_embeds_vf.shape[1]
+        B, _, _, S, C = audio_embeds.shape
+
+        # process audio of first frame
+        audio_embeds = rearrange(audio_embeds, "bz f w b c -> (bz f) w b c")
+        batch_size, window_size, blocks, channels = audio_embeds.shape
+        audio_embeds = audio_embeds.view(batch_size, window_size * blocks * channels)
+
+        # process audio of latter frame
+        audio_embeds_vf = rearrange(audio_embeds_vf, "bz f w b c -> (bz f) w b c")
+        batch_size_vf, window_size_vf, blocks_vf, channels_vf = audio_embeds_vf.shape
+        audio_embeds_vf = audio_embeds_vf.view(batch_size_vf, window_size_vf * blocks_vf * channels_vf)
+
+        # first projection
+        audio_embeds = torch.relu(self.proj1(audio_embeds))
+        audio_embeds_vf = torch.relu(self.proj1_vf(audio_embeds_vf))
+        audio_embeds = rearrange(audio_embeds, "(bz f) c -> bz f c", bz=B)
+        audio_embeds_vf = rearrange(audio_embeds_vf, "(bz f) c -> bz f c", bz=B)
+        audio_embeds_c = torch.concat([audio_embeds, audio_embeds_vf], dim=1)
+        batch_size_c, N_t, C_a = audio_embeds_c.shape
+        audio_embeds_c = audio_embeds_c.view(batch_size_c*N_t, C_a)
+
+        # second projection
+        audio_embeds_c = torch.relu(self.proj2(audio_embeds_c))
+
+        context_tokens = self.proj3(audio_embeds_c).reshape(batch_size_c*N_t, self.context_tokens, self.out_dim)
+
+        # normalization and reshape
+        context_tokens = self.norm(context_tokens)
+        context_tokens = rearrange(context_tokens, "(bz f) m c -> bz f m c", f=video_length)
+
+        return context_tokens
+
+
+class WanMultiTalkAttentionBlock(torch.nn.Module):
+    def __init__(self, in_dim=5120, out_dim=768, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.audio_cross_attn = SingleStreamMultiAttention(in_dim, out_dim, num_heads=40, qkv_bias=True, device=device, dtype=dtype, operations=operations)
+        self.norm_x = operations.LayerNorm(in_dim, device=device, dtype=dtype, elementwise_affine=True)
+
+
+class MultiTalkGetAttnMapPatch:
+    def __init__(self, ref_target_masks=None):
+        self.ref_target_masks = ref_target_masks
+
+    def __call__(self, kwargs):
+        transformer_options = kwargs.get("transformer_options", {})
+        x = kwargs["x"]
+
+        if self.ref_target_masks is not None:
+            x_ref_attn_map = get_attn_map_with_target(kwargs["q"], kwargs["k"], transformer_options["grid_sizes"], ref_target_masks=self.ref_target_masks.to(x.device))
+            transformer_options["x_ref_attn_map"] = x_ref_attn_map
+        return x
+
+
+class MultiTalkCrossAttnPatch:
+    def __init__(self, model_patch, audio_scale=1.0, ref_target_masks=None):
+        self.model_patch = model_patch
+        self.audio_scale = audio_scale
+        self.ref_target_masks = ref_target_masks
+
+    def __call__(self, kwargs):
+        transformer_options = kwargs.get("transformer_options", {})
+        block_idx = transformer_options.get("block_index", None)
+        x = kwargs["x"]
+        if block_idx is None:
+            return torch.zeros_like(x)
+
+        audio_embeds = transformer_options.get("audio_embeds")
+        x_ref_attn_map = transformer_options.pop("x_ref_attn_map", None)
+
+        norm_x = self.model_patch.model.blocks[block_idx].norm_x(x)
+        x_audio = self.model_patch.model.blocks[block_idx].audio_cross_attn(
+            norm_x, audio_embeds.to(x.dtype),
+            shape=transformer_options["grid_sizes"],
+            x_ref_attn_map=x_ref_attn_map
+        )
+        x = x + x_audio * self.audio_scale
+        return x
+
+    def models(self):
+        return [self.model_patch]
+
+class MultiTalkApplyModelWrapper:
+    def __init__(self, init_latents):
+        self.init_latents = init_latents
+
+    def __call__(self, executor, x, *args, **kwargs):
+        x[:, :, :self.init_latents.shape[2]] = self.init_latents.to(x)
+        samples = executor(x, *args, **kwargs)
+        return samples
+
+
+class InfiniteTalkOuterSampleWrapper:
+    def __init__(self, motion_frames_latent, model_patch, is_extend=False):
+        self.motion_frames_latent = motion_frames_latent
+        self.model_patch = model_patch
+        self.is_extend = is_extend
+
+    def __call__(self, executor, *args, **kwargs):
+        model_patcher = executor.class_obj.model_patcher
+        model_options = executor.class_obj.model_options
+        process_latent_in = model_patcher.model.process_latent_in
+
+        # for InfiniteTalk, model input first latent(s) need to always be replaced on every step
+        if self.motion_frames_latent is not None:
+            wrappers = model_options["transformer_options"]["wrappers"]
+            w = wrappers.setdefault(comfy.patcher_extension.WrappersMP.APPLY_MODEL, {})
+            w["MultiTalk_apply_model"] = [MultiTalkApplyModelWrapper(process_latent_in(self.motion_frames_latent))]
+
+        # run the sampling process
+        result = executor(*args, **kwargs)
+
+        # insert motion frames before decoding
+        if self.is_extend:
+            overlap = self.motion_frames_latent.shape[2]
+            result = torch.cat([self.motion_frames_latent.to(result), result[:, :, overlap:]], dim=2)
+
+        return result
+
+    def to(self, device_or_dtype):
+        if isinstance(device_or_dtype, torch.device):
+            if self.motion_frames_latent is not None:
+                self.motion_frames_latent = self.motion_frames_latent.to(device_or_dtype)
+        return self

vendor/ComfyUI/comfy/ldm/wan/model_wandancer.py

@@ -0,0 +1,251 @@
+import torch
+import torch.nn as nn
+import comfy
+from comfy.ldm.modules.attention import optimized_attention
+from comfy.ldm.flux.math import apply_rope1
+from comfy.ldm.flux.layers import EmbedND
+
+from .model import AudioInjector_WAN, WanModel, MLPProj, Head, sinusoidal_embedding_1d
+
+
+class MusicSelfAttention(nn.Module):
+    def __init__(self, dim, num_heads, device=None, dtype=None, operations=None):
+        assert dim % num_heads == 0
+        super().__init__()
+        self.embed_dim = dim
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+
+        self.q_proj = operations.Linear(dim, dim, device=device, dtype=dtype)
+        self.k_proj = operations.Linear(dim, dim, device=device, dtype=dtype)
+        self.v_proj = operations.Linear(dim, dim, device=device, dtype=dtype)
+        self.out_proj = operations.Linear(dim, dim, device=device, dtype=dtype)
+
+    def forward(self, x, freqs):
+        b, s, n, d = *x.shape[:2], self.num_heads, self.head_dim
+
+        q = self.q_proj(x).view(b, s, n, d)
+        q = apply_rope1(q, freqs)
+
+        k = self.k_proj(x).view(b, s, n, d)
+        k = apply_rope1(k, freqs)
+
+        x = optimized_attention(
+            q.view(b, s, n * d),
+            k.view(b, s, n * d),
+            self.v_proj(x).view(b, s, n * d),
+            heads=self.num_heads,
+        )
+
+        return self.out_proj(x)
+
+
+class MusicEncoderLayer(nn.Module):
+    def __init__(self, dim: int, num_heads: int, ffn_dim: int, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.self_attn = MusicSelfAttention(dim, num_heads, device=device, dtype=dtype, operations=operations)
+
+        self.linear1 = operations.Linear(dim, ffn_dim, device=device, dtype=dtype)
+        self.linear2 = operations.Linear(ffn_dim, dim, device=device, dtype=dtype)
+
+        self.norm1 = operations.LayerNorm(dim, device=device, dtype=dtype)
+        self.norm2 = operations.LayerNorm(dim, device=device, dtype=dtype)
+
+    def forward(self, x: torch.Tensor, freqs: torch.Tensor) -> torch.Tensor:
+        x = x + self.self_attn(self.norm1(x), freqs=freqs)
+        x = x + self.linear2(torch.nn.functional.gelu(self.linear1(self.norm2(x)))) # ffn
+        return x
+
+
+class WanDancerModel(WanModel):
+    def __init__(self,
+                 model_type='wandancer',
+                 patch_size=(1, 2, 2),
+                 text_len=512,
+                 in_dim=16,
+                 dim=5120,
+                 ffn_dim=8192,
+                 freq_dim=256,
+                 text_dim=4096,
+                 out_dim=16,
+                 num_heads=16,
+                 num_layers=40,
+                 window_size=(-1, -1),
+                 qk_norm=True,
+                 cross_attn_norm=True,
+                 eps=1e-6,
+                 in_dim_ref_conv=None,
+                 image_model=None,
+                 device=None, dtype=None, operations=None,
+                 audio_inject_layers=[0, 4, 8, 12, 16, 20, 24, 27],
+                 music_dim = 256,
+                 music_heads = 4,
+                 music_feature_dim = 35,
+                 music_latent_dim = 256
+                 ):
+
+        super().__init__(model_type='i2v', patch_size=patch_size, text_len=text_len, in_dim=in_dim, dim=dim, ffn_dim=ffn_dim, freq_dim=freq_dim, text_dim=text_dim, out_dim=out_dim,
+                         num_heads=num_heads, num_layers=num_layers, window_size=window_size, qk_norm=qk_norm, cross_attn_norm=cross_attn_norm, eps=eps, image_model=image_model, in_dim_ref_conv=in_dim_ref_conv,
+                         device=device, dtype=dtype, operations=operations)
+
+        self.dtype = dtype
+        operation_settings = {"operations": operations, "device": device, "dtype": dtype}
+
+        self.patch_embedding_global = operations.Conv3d(in_dim, dim, kernel_size=patch_size, stride=patch_size, device=operation_settings.get("device"), dtype=torch.float32)
+        self.img_emb_refimage = MLPProj(1280, dim, operation_settings=operation_settings)
+        self.head_global = Head(dim, out_dim, patch_size, eps, operation_settings=operation_settings)
+
+        self.music_injector = AudioInjector_WAN(
+            dim=self.dim,
+            num_heads=self.num_heads,
+            inject_layer=audio_inject_layers,
+            root_net=self,
+            enable_adain=False,
+            dtype=dtype, device=device, operations=operations
+        )
+
+        self.music_projection = operations.Linear(music_feature_dim, music_latent_dim, device=device, dtype=dtype)
+        self.music_encoder = nn.ModuleList([MusicEncoderLayer(dim=music_dim, num_heads=music_heads, ffn_dim=1024, device=device, dtype=dtype, operations=operations) for _ in range(2)])
+        music_head_dim = music_dim // music_heads
+        self.music_rope_embedder = EmbedND(dim=music_head_dim, theta=10000.0, axes_dim=[music_head_dim])
+
+    def forward_orig(self, x, t, context, clip_fea=None, clip_fea_ref=None, freqs=None, audio_embed=None, fps=30, audio_inject_scale=1.0, transformer_options={}, **kwargs):
+        # embeddings
+        if int(fps + 0.5) != 30:
+            x = self.patch_embedding_global(x.float()).to(x.dtype)
+        else:
+            x = self.patch_embedding(x.float()).to(x.dtype)
+
+        grid_sizes = x.shape[2:]
+        latent_frames = grid_sizes[0]
+        transformer_options["grid_sizes"] = grid_sizes
+        x = x.flatten(2).transpose(1, 2)
+        seq_len = x.size(1)
+
+        # time embeddings
+        e = self.time_embedding(sinusoidal_embedding_1d(self.freq_dim, t.flatten()).to(dtype=x[0].dtype))
+        e = e.reshape(t.shape[0], -1, e.shape[-1])
+        e0 = self.time_projection(e).unflatten(2, (6, self.dim))
+
+        full_ref = None
+        if self.ref_conv is not None: # model has the weight, but this wasn't used in the original pipeline
+            full_ref = kwargs.get("reference_latent", None)
+            if full_ref is not None:
+                full_ref = self.ref_conv(full_ref).flatten(2).transpose(1, 2)
+                x = torch.concat((full_ref, x), dim=1)
+
+        # context
+        context = self.text_embedding(context)
+
+        audio_emb = None
+        if audio_embed is not None: # encode music feature，[1, frame_num, 35] -> [1, F*8, dim]
+            music_feature = self.music_projection(audio_embed)
+
+            music_seq_len = music_feature.shape[1]
+            music_ids = torch.arange(music_seq_len, device=music_feature.device, dtype=music_feature.dtype).reshape(1, -1, 1) # create 1D position IDs
+            music_freqs = self.music_rope_embedder(music_ids).movedim(1, 2)
+
+            # apply encoder layers
+            for layer in self.music_encoder:
+                music_feature = layer(music_feature, music_freqs)
+
+            # interpolate
+            audio_emb = torch.nn.functional.interpolate(music_feature.unsqueeze(1), size=(latent_frames * 8, self.dim), mode='bilinear').squeeze(1)
+
+        context_img_len = 0
+        if self.img_emb is not None and clip_fea is not None:
+            context_clip = self.img_emb(clip_fea)  # bs x 257 x dim
+            context = torch.cat([context_clip, context], dim=1)
+            context_img_len += clip_fea.shape[-2]
+        if self.img_emb_refimage is not None and clip_fea_ref is not None:
+            context_clip_ref = self.img_emb_refimage(clip_fea_ref)
+            context = torch.cat([context_clip_ref, context], dim=1)
+            context_img_len += clip_fea_ref.shape[-2]
+
+        patches_replace = transformer_options.get("patches_replace", {})
+        blocks_replace = patches_replace.get("dit", {})
+        transformer_options["total_blocks"] = len(self.blocks)
+        transformer_options["block_type"] = "double"
+        for i, block in enumerate(self.blocks):
+            transformer_options["block_index"] = i
+            if ("double_block", i) in blocks_replace:
+                def block_wrap(args):
+                    out = {}
+                    out["img"] = block(args["img"], context=args["txt"], e=args["vec"], freqs=args["pe"], context_img_len=context_img_len, transformer_options=args["transformer_options"])
+                    return out
+                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": e0, "pe": freqs, "transformer_options": transformer_options}, {"original_block": block_wrap})
+                x = out["img"]
+            else:
+                x = block(x, e=e0, freqs=freqs, context=context, context_img_len=context_img_len, transformer_options=transformer_options)
+            if audio_emb is not None:
+                x = self.music_injector(x, i, audio_emb, audio_emb_global=None, seq_len=seq_len, scale=audio_inject_scale)
+
+        # head
+        if int(fps + 0.5) != 30:
+            x = self.head_global(x, e)
+        else:
+            x = self.head(x, e)
+
+        if full_ref is not None:
+            x = x[:, full_ref.shape[1]:]
+
+        # unpatchify
+        x = self.unpatchify(x, grid_sizes)
+        return x
+
+    def _forward(self, x, timestep, context, clip_fea=None, time_dim_concat=None, transformer_options={}, clip_fea_ref=None, fps=30, audio_inject_scale=1.0, **kwargs):
+        bs, c, t, h, w = x.shape
+        x = comfy.ldm.common_dit.pad_to_patch_size(x, self.patch_size)
+
+        t_len = t
+        if time_dim_concat is not None:
+            time_dim_concat = comfy.ldm.common_dit.pad_to_patch_size(time_dim_concat, self.patch_size)
+            x = torch.cat([x, time_dim_concat], dim=2)
+            t_len = x.shape[2]
+
+        freqs = self.rope_encode(t_len, h, w, device=x.device, dtype=x.dtype, fps=fps, transformer_options=transformer_options)
+        return self.forward_orig(x, timestep, context, clip_fea=clip_fea, clip_fea_ref=clip_fea_ref, freqs=freqs, fps=fps, audio_inject_scale=audio_inject_scale, transformer_options=transformer_options, **kwargs)[:, :, :t, :h, :w]
+
+    def rope_encode(self, t, h, w, t_start=0, steps_t=None, steps_h=None, steps_w=None, fps=30, device=None, dtype=None, transformer_options={}):
+        patch_size = self.patch_size
+        t_len = ((t + (patch_size[0] // 2)) // patch_size[0])
+        h_len = ((h + (patch_size[1] // 2)) // patch_size[1])
+        w_len = ((w + (patch_size[2] // 2)) // patch_size[2])
+
+        if steps_t is None:
+            steps_t = t_len
+        if steps_h is None:
+            steps_h = h_len
+        if steps_w is None:
+            steps_w = w_len
+
+        h_start = 0
+        w_start = 0
+        rope_options = transformer_options.get("rope_options", None)
+        if rope_options is not None:
+            t_len = (t_len - 1.0) * rope_options.get("scale_t", 1.0) + 1.0
+            h_len = (h_len - 1.0) * rope_options.get("scale_y", 1.0) + 1.0
+            w_len = (w_len - 1.0) * rope_options.get("scale_x", 1.0) + 1.0
+
+            t_start += rope_options.get("shift_t", 0.0)
+            h_start += rope_options.get("shift_y", 0.0)
+            w_start += rope_options.get("shift_x", 0.0)
+
+        img_ids = torch.zeros((steps_t, steps_h, steps_w, 3), device=device, dtype=dtype)
+
+        if int(fps + 0.5) != 30:
+            time_scale = 30.0 / fps # how many time units each frame represents relative to 30fps
+            positions_new = torch.arange(steps_t, device=device, dtype=dtype) * time_scale + t_start
+            total_frames_at_30fps = int(time_scale * steps_t + 0.5)
+            positions_new[-1] = t_start + (total_frames_at_30fps - 1)
+
+            img_ids[:, :, :, 0] = img_ids[:, :, :, 0] + positions_new.reshape(-1, 1, 1)
+        else:
+            img_ids[:, :, :, 0] = img_ids[:, :, :, 0] + torch.linspace(t_start, t_start + (t_len - 1), steps=steps_t, device=device, dtype=dtype).reshape(-1, 1, 1)
+
+        img_ids[:, :, :, 1] = img_ids[:, :, :, 1] + torch.linspace(h_start, h_start + (h_len - 1), steps=steps_h, device=device, dtype=dtype).reshape(1, -1, 1)
+        img_ids[:, :, :, 2] = img_ids[:, :, :, 2] + torch.linspace(w_start, w_start + (w_len - 1), steps=steps_w, device=device, dtype=dtype).reshape(1, 1, -1)
+        img_ids = img_ids.reshape(1, -1, img_ids.shape[-1])
+
+        freqs = self.rope_embedder(img_ids).movedim(1, 2)
+        return freqs

vendor/ComfyUI/comfy/ldm/wan/vae.py

@@ -0,0 +1,511 @@
+# original version: https://github.com/Wan-Video/Wan2.1/blob/main/wan/modules/vae.py
+# Copyright 2024-2025 The Alibaba Wan Team Authors. All rights reserved.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange
+from comfy.ldm.modules.diffusionmodules.model import vae_attention, torch_cat_if_needed
+
+import comfy.ops
+ops = comfy.ops.disable_weight_init
+
+CACHE_T = 2
+
+
+class CausalConv3d(ops.Conv3d):
+    """
+    Causal 3d convolusion.
+    """
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self._padding = 2 * self.padding[0]
+        self.padding = (0, self.padding[1], self.padding[2])
+
+    def forward(self, x, cache_x=None, cache_list=None, cache_idx=None):
+        if cache_list is not None:
+            cache_x = cache_list[cache_idx]
+            cache_list[cache_idx] = None
+
+        if cache_x is None and x.shape[2] == 1:
+            #Fast path - the op will pad for use by truncating the weight
+            #and save math on a pile of zeros.
+            return super().forward(x, autopad="causal_zero")
+
+        if self._padding > 0:
+            padding_needed = self._padding
+            if cache_x is not None:
+                cache_x = cache_x.to(x.device)
+                padding_needed = max(0, padding_needed - cache_x.shape[2])
+            padding_shape = list(x.shape)
+            padding_shape[2] = padding_needed
+            padding = torch.zeros(padding_shape, device=x.device, dtype=x.dtype)
+            x = torch_cat_if_needed([padding, cache_x, x], dim=2)
+            del cache_x
+
+        return super().forward(x)
+
+
+class RMS_norm(nn.Module):
+
+    def __init__(self, dim, channel_first=True, images=True, bias=False):
+        super().__init__()
+        broadcastable_dims = (1, 1, 1) if not images else (1, 1)
+        shape = (dim, *broadcastable_dims) if channel_first else (dim,)
+
+        self.channel_first = channel_first
+        self.scale = dim**0.5
+        self.gamma = nn.Parameter(torch.ones(shape))
+        self.bias = nn.Parameter(torch.zeros(shape)) if bias else None
+
+    def forward(self, x):
+        return F.normalize(
+            x, dim=(1 if self.channel_first else -1)) * self.scale * self.gamma.to(x) + (self.bias.to(x) if self.bias is not None else 0)
+
+
+class Resample(nn.Module):
+
+    def __init__(self, dim, mode):
+        assert mode in ('none', 'upsample2d', 'upsample3d', 'downsample2d',
+                        'downsample3d')
+        super().__init__()
+        self.dim = dim
+        self.mode = mode
+
+        # layers
+        if mode == 'upsample2d':
+            self.resample = nn.Sequential(
+                nn.Upsample(scale_factor=(2., 2.), mode='nearest-exact'),
+                ops.Conv2d(dim, dim // 2, 3, padding=1))
+        elif mode == 'upsample3d':
+            self.resample = nn.Sequential(
+                nn.Upsample(scale_factor=(2., 2.), mode='nearest-exact'),
+                ops.Conv2d(dim, dim // 2, 3, padding=1))
+            self.time_conv = CausalConv3d(
+                dim, dim * 2, (3, 1, 1), padding=(1, 0, 0))
+
+        elif mode == 'downsample2d':
+            self.resample = nn.Sequential(
+                nn.ZeroPad2d((0, 1, 0, 1)),
+                ops.Conv2d(dim, dim, 3, stride=(2, 2)))
+        elif mode == 'downsample3d':
+            self.resample = nn.Sequential(
+                nn.ZeroPad2d((0, 1, 0, 1)),
+                ops.Conv2d(dim, dim, 3, stride=(2, 2)))
+            self.time_conv = CausalConv3d(
+                dim, dim, (3, 1, 1), stride=(2, 1, 1), padding=(0, 0, 0))
+
+        else:
+            self.resample = nn.Identity()
+
+    def forward(self, x, feat_cache=None, feat_idx=[0], final=False):
+        b, c, t, h, w = x.size()
+        if self.mode == 'upsample3d':
+            if feat_cache is not None:
+                idx = feat_idx[0]
+                if feat_cache[idx] is None:
+                    feat_cache[idx] = 'Rep'
+                    feat_idx[0] += 1
+                else:
+
+                    cache_x = x[:, :, -CACHE_T:, :, :]
+                    if feat_cache[idx] == 'Rep':
+                        x = self.time_conv(x)
+                    else:
+                        x = self.time_conv(x, feat_cache[idx])
+                    feat_cache[idx] = cache_x
+                    feat_idx[0] += 1
+
+                    x = x.reshape(b, 2, c, t, h, w)
+                    x = torch.stack((x[:, 0, :, :, :, :], x[:, 1, :, :, :, :]),
+                                    3)
+                    x = x.reshape(b, c, t * 2, h, w)
+        t = x.shape[2]
+        x = rearrange(x, 'b c t h w -> (b t) c h w')
+        x = self.resample(x)
+        x = rearrange(x, '(b t) c h w -> b c t h w', t=t)
+
+        if self.mode == 'downsample3d':
+            if feat_cache is not None:
+                idx = feat_idx[0]
+                if feat_cache[idx] is None:
+                    feat_cache[idx] = x
+                else:
+
+                    cache_x = x[:, :, -1:, :, :]
+                    x = self.time_conv(
+                        torch.cat([feat_cache[idx][:, :, -1:, :, :], x], 2))
+                    feat_cache[idx] = cache_x
+
+                    deferred_x = feat_cache[idx + 1]
+                    if deferred_x is not None:
+                        x = torch.cat([deferred_x, x], 2)
+                        feat_cache[idx + 1] = None
+
+                    if x.shape[2] == 1 and not final:
+                        feat_cache[idx + 1] = x
+                        x = None
+
+                feat_idx[0] += 2
+        return x
+
+
+class ResidualBlock(nn.Module):
+
+    def __init__(self, in_dim, out_dim, dropout=0.0):
+        super().__init__()
+        self.in_dim = in_dim
+        self.out_dim = out_dim
+
+        # layers
+        self.residual = nn.Sequential(
+            RMS_norm(in_dim, images=False), nn.SiLU(),
+            CausalConv3d(in_dim, out_dim, 3, padding=1),
+            RMS_norm(out_dim, images=False), nn.SiLU(), nn.Dropout(dropout),
+            CausalConv3d(out_dim, out_dim, 3, padding=1))
+        self.shortcut = CausalConv3d(in_dim, out_dim, 1) \
+            if in_dim != out_dim else nn.Identity()
+
+    def forward(self, x, feat_cache=None, feat_idx=[0], final=False):
+        old_x = x
+        for layer in self.residual:
+            if isinstance(layer, CausalConv3d) and feat_cache is not None:
+                idx = feat_idx[0]
+                cache_x = x[:, :, -CACHE_T:, :, :]
+                x = layer(x, cache_list=feat_cache, cache_idx=idx)
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+            else:
+                x = layer(x)
+        return x + self.shortcut(old_x)
+
+
+class AttentionBlock(nn.Module):
+    """
+    Causal self-attention with a single head.
+    """
+
+    def __init__(self, dim):
+        super().__init__()
+        self.dim = dim
+
+        # layers
+        self.norm = RMS_norm(dim)
+        self.to_qkv = ops.Conv2d(dim, dim * 3, 1)
+        self.proj = ops.Conv2d(dim, dim, 1)
+        self.optimized_attention = vae_attention()
+
+    def forward(self, x, feat_cache=None, feat_idx=[0], final=False):
+        identity = x
+        b, c, t, h, w = x.size()
+        x = rearrange(x, 'b c t h w -> (b t) c h w')
+        x = self.norm(x)
+        # compute query, key, value
+
+        q, k, v = self.to_qkv(x).chunk(3, dim=1)
+        x = self.optimized_attention(q, k, v)
+
+        # output
+        x = self.proj(x)
+        x = rearrange(x, '(b t) c h w-> b c t h w', t=t)
+        return x + identity
+
+
+class Encoder3d(nn.Module):
+
+    def __init__(self,
+                 dim=128,
+                 z_dim=4,
+                 input_channels=3,
+                 dim_mult=[1, 2, 4, 4],
+                 num_res_blocks=2,
+                 attn_scales=[],
+                 temperal_downsample=[True, True, False],
+                 dropout=0.0):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_downsample = temperal_downsample
+
+        # dimensions
+        dims = [dim * u for u in [1] + dim_mult]
+        scale = 1.0
+
+        # init block
+        self.conv1 = CausalConv3d(input_channels, dims[0], 3, padding=1)
+
+        # downsample blocks
+        downsamples = []
+        for i, (in_dim, out_dim) in enumerate(zip(dims[:-1], dims[1:])):
+            # residual (+attention) blocks
+            for _ in range(num_res_blocks):
+                downsamples.append(ResidualBlock(in_dim, out_dim, dropout))
+                if scale in attn_scales:
+                    downsamples.append(AttentionBlock(out_dim))
+                in_dim = out_dim
+
+            # downsample block
+            if i != len(dim_mult) - 1:
+                mode = 'downsample3d' if temperal_downsample[
+                    i] else 'downsample2d'
+                downsamples.append(Resample(out_dim, mode=mode))
+                scale /= 2.0
+        self.downsamples = nn.Sequential(*downsamples)
+
+        # middle blocks
+        self.middle = nn.Sequential(
+            ResidualBlock(out_dim, out_dim, dropout), AttentionBlock(out_dim),
+            ResidualBlock(out_dim, out_dim, dropout))
+
+        # output blocks
+        self.head = nn.Sequential(
+            RMS_norm(out_dim, images=False), nn.SiLU(),
+            CausalConv3d(out_dim, z_dim, 3, padding=1))
+
+    def forward(self, x, feat_cache=None, feat_idx=[0], final=False):
+        if feat_cache is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :]
+            x = self.conv1(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            x = self.conv1(x)
+
+        ## downsamples
+        for layer in self.downsamples:
+            if feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx, final=final)
+                if x is None:
+                    return None
+            else:
+                x = layer(x)
+
+        ## middle
+        for layer in self.middle:
+            if feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx, final=final)
+            else:
+                x = layer(x)
+
+        ## head
+        for layer in self.head:
+            if isinstance(layer, CausalConv3d) and feat_cache is not None:
+                idx = feat_idx[0]
+                cache_x = x[:, :, -CACHE_T:, :, :]
+                x = layer(x, feat_cache[idx])
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+            else:
+                x = layer(x)
+        return x
+
+
+class Decoder3d(nn.Module):
+
+    def __init__(self,
+                 dim=128,
+                 z_dim=4,
+                 output_channels=3,
+                 dim_mult=[1, 2, 4, 4],
+                 num_res_blocks=2,
+                 attn_scales=[],
+                 temperal_upsample=[False, True, True],
+                 dropout=0.0):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_upsample = temperal_upsample
+
+        # dimensions
+        dims = [dim * u for u in [dim_mult[-1]] + dim_mult[::-1]]
+        scale = 1.0 / 2**(len(dim_mult) - 2)
+
+        # init block
+        self.conv1 = CausalConv3d(z_dim, dims[0], 3, padding=1)
+
+        # middle blocks
+        self.middle = nn.Sequential(
+            ResidualBlock(dims[0], dims[0], dropout), AttentionBlock(dims[0]),
+            ResidualBlock(dims[0], dims[0], dropout))
+
+        # upsample blocks
+        upsamples = []
+        for i, (in_dim, out_dim) in enumerate(zip(dims[:-1], dims[1:])):
+            # residual (+attention) blocks
+            if i == 1 or i == 2 or i == 3:
+                in_dim = in_dim // 2
+            for _ in range(num_res_blocks + 1):
+                upsamples.append(ResidualBlock(in_dim, out_dim, dropout))
+                if scale in attn_scales:
+                    upsamples.append(AttentionBlock(out_dim))
+                in_dim = out_dim
+
+            # upsample block
+            if i != len(dim_mult) - 1:
+                mode = 'upsample3d' if temperal_upsample[i] else 'upsample2d'
+                upsamples.append(Resample(out_dim, mode=mode))
+                scale *= 2.0
+        self.upsamples = nn.Sequential(*upsamples)
+
+        # output blocks
+        self.head = nn.Sequential(
+            RMS_norm(out_dim, images=False), nn.SiLU(),
+            CausalConv3d(out_dim, output_channels, 3, padding=1))
+
+    def run_up(self, layer_idx, x_ref, feat_cache, feat_idx, out_chunks):
+        x = x_ref[0]
+        x_ref[0] = None
+        if layer_idx >= len(self.upsamples):
+            for layer in self.head:
+                if isinstance(layer, CausalConv3d) and feat_cache is not None:
+                    cache_x = x[:, :, -CACHE_T:, :, :]
+                    x = layer(x, feat_cache[feat_idx[0]])
+                    feat_cache[feat_idx[0]] = cache_x
+                    feat_idx[0] += 1
+                else:
+                    x = layer(x)
+            out_chunks.append(x)
+            return
+
+        layer = self.upsamples[layer_idx]
+        if feat_cache is not None:
+            x = layer(x, feat_cache, feat_idx)
+        else:
+            x = layer(x)
+
+        if isinstance(layer, Resample) and layer.mode == 'upsample3d' and x.shape[2] > 2:
+            for frame_idx in range(0, x.shape[2], 2):
+                self.run_up(
+                    layer_idx + 1,
+                    [x[:, :, frame_idx:frame_idx + 2, :, :]],
+                    feat_cache,
+                    feat_idx.copy(),
+                    out_chunks,
+                )
+            del x
+            return
+
+        next_x_ref = [x]
+        del x
+        self.run_up(layer_idx + 1, next_x_ref, feat_cache, feat_idx, out_chunks)
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        ## conv1
+        if feat_cache is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :]
+            x = self.conv1(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            x = self.conv1(x)
+
+        ## middle
+        for layer in self.middle:
+            if feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        out_chunks = []
+
+        self.run_up(0, [x], feat_cache, feat_idx, out_chunks)
+        return out_chunks
+
+
+def count_cache_layers(model):
+    count = 0
+    for m in model.modules():
+        if isinstance(m, CausalConv3d) or (isinstance(m, Resample) and m.mode == 'downsample3d'):
+            count += 1
+    return count
+
+
+class WanVAE(nn.Module):
+
+    def __init__(self,
+                 dim=128,
+                 z_dim=4,
+                 dim_mult=[1, 2, 4, 4],
+                 num_res_blocks=2,
+                 attn_scales=[],
+                 temperal_downsample=[True, True, False],
+                 image_channels=3,
+                 conv_out_channels=3,
+                 dropout=0.0):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_downsample = temperal_downsample
+        self.temperal_upsample = temperal_downsample[::-1]
+
+        # modules
+        self.encoder = Encoder3d(dim, z_dim * 2, image_channels, dim_mult, num_res_blocks,
+                                 attn_scales, self.temperal_downsample, dropout)
+        self.conv1 = CausalConv3d(z_dim * 2, z_dim * 2, 1)
+        self.conv2 = CausalConv3d(z_dim, z_dim, 1)
+        self.decoder = Decoder3d(dim, z_dim, conv_out_channels, dim_mult, num_res_blocks,
+                                 attn_scales, self.temperal_upsample, dropout)
+
+    def encode(self, x):
+        conv_idx = [0]
+        ## cache
+        t = x.shape[2]
+        t = 1 + ((t - 1) // 4) * 4
+        iter_ = 1 + (t - 1) // 2
+        feat_map = None
+        if iter_ > 1:
+            feat_map = [None] * count_cache_layers(self.encoder)
+        ## 对encode输入的x，按时间拆分为1、2、2、2....(总帧数先按4N+1向下取整)
+        for i in range(iter_):
+            conv_idx = [0]
+            if i == 0:
+                out = self.encoder(
+                    x[:, :, :1, :, :],
+                    feat_cache=feat_map,
+                    feat_idx=conv_idx)
+            else:
+                out_ = self.encoder(
+                    x[:, :, 1 + 2 * (i - 1):1 + 2 * i, :, :],
+                    feat_cache=feat_map,
+                    feat_idx=conv_idx,
+                    final=(i == (iter_ - 1)))
+                if out_ is None:
+                    continue
+                out = torch.cat([out, out_], 2)
+
+        mu, log_var = self.conv1(out).chunk(2, dim=1)
+        return mu
+
+    def decode(self, z):
+        # z: [b,c,t,h,w]
+        iter_ = 1 + z.shape[2] // 2
+        feat_map = None
+        if iter_ > 1:
+            feat_map = [None] * count_cache_layers(self.decoder)
+        x = self.conv2(z)
+        for i in range(iter_):
+            conv_idx = [0]
+            if i == 0:
+                out = self.decoder(
+                    x[:, :, i:i + 1, :, :],
+                    feat_cache=feat_map,
+                    feat_idx=conv_idx)
+            else:
+                out_ = self.decoder(
+                    x[:, :, 1 + 2 * (i - 1):1 + 2 * i, :, :],
+                    feat_cache=feat_map,
+                    feat_idx=conv_idx)
+                out += out_
+        return torch.cat(out, 2)

vendor/ComfyUI/comfy/ldm/wan/vae2_2.py

@@ -0,0 +1,717 @@
+# original version: https://github.com/Wan-Video/Wan2.2/blob/main/wan/modules/vae2_2.py
+# Copyright 2024-2025 The Alibaba Wan Team Authors. All rights reserved.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange
+from .vae import AttentionBlock, CausalConv3d, RMS_norm
+
+import comfy.ops
+ops = comfy.ops.disable_weight_init
+
+CACHE_T = 2
+
+
+class Resample(nn.Module):
+
+    def __init__(self, dim, mode):
+        assert mode in (
+            "none",
+            "upsample2d",
+            "upsample3d",
+            "downsample2d",
+            "downsample3d",
+        )
+        super().__init__()
+        self.dim = dim
+        self.mode = mode
+
+        # layers
+        if mode == "upsample2d":
+            self.resample = nn.Sequential(
+                nn.Upsample(scale_factor=(2.0, 2.0), mode="nearest-exact"),
+                ops.Conv2d(dim, dim, 3, padding=1),
+            )
+        elif mode == "upsample3d":
+            self.resample = nn.Sequential(
+                nn.Upsample(scale_factor=(2.0, 2.0), mode="nearest-exact"),
+                ops.Conv2d(dim, dim, 3, padding=1),
+                # ops.Conv2d(dim, dim//2, 3, padding=1)
+            )
+            self.time_conv = CausalConv3d(
+                dim, dim * 2, (3, 1, 1), padding=(1, 0, 0))
+        elif mode == "downsample2d":
+            self.resample = nn.Sequential(
+                nn.ZeroPad2d((0, 1, 0, 1)),
+                ops.Conv2d(dim, dim, 3, stride=(2, 2)))
+        elif mode == "downsample3d":
+            self.resample = nn.Sequential(
+                nn.ZeroPad2d((0, 1, 0, 1)),
+                ops.Conv2d(dim, dim, 3, stride=(2, 2)))
+            self.time_conv = CausalConv3d(
+                dim, dim, (3, 1, 1), stride=(2, 1, 1), padding=(0, 0, 0))
+        else:
+            self.resample = nn.Identity()
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        b, c, t, h, w = x.size()
+        if self.mode == "upsample3d":
+            if feat_cache is not None:
+                idx = feat_idx[0]
+                if feat_cache[idx] is None:
+                    feat_cache[idx] = "Rep"
+                    feat_idx[0] += 1
+                else:
+                    cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                    if (cache_x.shape[2] < 2 and feat_cache[idx] is not None and
+                            feat_cache[idx] != "Rep"):
+                        # cache last frame of last two chunk
+                        cache_x = torch.cat(
+                            [
+                                feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                                    cache_x.device),
+                                cache_x,
+                            ],
+                            dim=2,
+                        )
+                    if (cache_x.shape[2] < 2 and feat_cache[idx] is not None and
+                            feat_cache[idx] == "Rep"):
+                        cache_x = torch.cat(
+                            [
+                                torch.zeros_like(cache_x).to(cache_x.device),
+                                cache_x
+                            ],
+                            dim=2,
+                        )
+                    if feat_cache[idx] == "Rep":
+                        x = self.time_conv(x)
+                    else:
+                        x = self.time_conv(x, feat_cache[idx])
+                    feat_cache[idx] = cache_x
+                    feat_idx[0] += 1
+                    x = x.reshape(b, 2, c, t, h, w)
+                    x = torch.stack((x[:, 0, :, :, :, :], x[:, 1, :, :, :, :]),
+                                    3)
+                    x = x.reshape(b, c, t * 2, h, w)
+        t = x.shape[2]
+        x = rearrange(x, "b c t h w -> (b t) c h w")
+        x = self.resample(x)
+        x = rearrange(x, "(b t) c h w -> b c t h w", t=t)
+
+        if self.mode == "downsample3d":
+            if feat_cache is not None:
+                idx = feat_idx[0]
+                if feat_cache[idx] is None:
+                    feat_cache[idx] = x.clone()
+                    feat_idx[0] += 1
+                else:
+                    cache_x = x[:, :, -1:, :, :].clone()
+                    x = self.time_conv(
+                        torch.cat([feat_cache[idx][:, :, -1:, :, :], x], 2))
+                    feat_cache[idx] = cache_x
+                    feat_idx[0] += 1
+        return x
+
+
+class ResidualBlock(nn.Module):
+
+    def __init__(self, in_dim, out_dim, dropout=0.0):
+        super().__init__()
+        self.in_dim = in_dim
+        self.out_dim = out_dim
+
+        # layers
+        self.residual = nn.Sequential(
+            RMS_norm(in_dim, images=False),
+            nn.SiLU(),
+            CausalConv3d(in_dim, out_dim, 3, padding=1),
+            RMS_norm(out_dim, images=False),
+            nn.SiLU(),
+            nn.Dropout(dropout),
+            CausalConv3d(out_dim, out_dim, 3, padding=1),
+        )
+        self.shortcut = (
+            CausalConv3d(in_dim, out_dim, 1)
+            if in_dim != out_dim else nn.Identity())
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        old_x = x
+        for layer in self.residual:
+            if isinstance(layer, CausalConv3d) and feat_cache is not None:
+                idx = feat_idx[0]
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                    # cache last frame of last two chunk
+                    cache_x = torch.cat(
+                        [
+                            feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                                cache_x.device),
+                            cache_x,
+                        ],
+                        dim=2,
+                    )
+                x = layer(x, cache_list=feat_cache, cache_idx=idx)
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+            else:
+                x = layer(x)
+        return x + self.shortcut(old_x)
+
+
+def patchify(x, patch_size):
+    if patch_size == 1:
+        return x
+    if x.dim() == 4:
+        x = rearrange(
+            x, "b c (h q) (w r) -> b (c r q) h w", q=patch_size, r=patch_size)
+    elif x.dim() == 5:
+        x = rearrange(
+            x,
+            "b c f (h q) (w r) -> b (c r q) f h w",
+            q=patch_size,
+            r=patch_size,
+        )
+    else:
+        raise ValueError(f"Invalid input shape: {x.shape}")
+
+    return x
+
+
+def unpatchify(x, patch_size):
+    if patch_size == 1:
+        return x
+
+    if x.dim() == 4:
+        x = rearrange(
+            x, "b (c r q) h w -> b c (h q) (w r)", q=patch_size, r=patch_size)
+    elif x.dim() == 5:
+        x = rearrange(
+            x,
+            "b (c r q) f h w -> b c f (h q) (w r)",
+            q=patch_size,
+            r=patch_size,
+        )
+    return x
+
+
+class AvgDown3D(nn.Module):
+
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        factor_t,
+        factor_s=1,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.factor_t = factor_t
+        self.factor_s = factor_s
+        self.factor = self.factor_t * self.factor_s * self.factor_s
+
+        assert in_channels * self.factor % out_channels == 0
+        self.group_size = in_channels * self.factor // out_channels
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        pad_t = (self.factor_t - x.shape[2] % self.factor_t) % self.factor_t
+        pad = (0, 0, 0, 0, pad_t, 0)
+        x = F.pad(x, pad)
+        B, C, T, H, W = x.shape
+        x = x.view(
+            B,
+            C,
+            T // self.factor_t,
+            self.factor_t,
+            H // self.factor_s,
+            self.factor_s,
+            W // self.factor_s,
+            self.factor_s,
+        )
+        x = x.permute(0, 1, 3, 5, 7, 2, 4, 6).contiguous()
+        x = x.view(
+            B,
+            C * self.factor,
+            T // self.factor_t,
+            H // self.factor_s,
+            W // self.factor_s,
+        )
+        x = x.view(
+            B,
+            self.out_channels,
+            self.group_size,
+            T // self.factor_t,
+            H // self.factor_s,
+            W // self.factor_s,
+        )
+        x = x.mean(dim=2)
+        return x
+
+
+class DupUp3D(nn.Module):
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        factor_t,
+        factor_s=1,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+
+        self.factor_t = factor_t
+        self.factor_s = factor_s
+        self.factor = self.factor_t * self.factor_s * self.factor_s
+
+        assert out_channels * self.factor % in_channels == 0
+        self.repeats = out_channels * self.factor // in_channels
+
+    def forward(self, x: torch.Tensor, first_chunk=False) -> torch.Tensor:
+        x = x.repeat_interleave(self.repeats, dim=1)
+        x = x.view(
+            x.size(0),
+            self.out_channels,
+            self.factor_t,
+            self.factor_s,
+            self.factor_s,
+            x.size(2),
+            x.size(3),
+            x.size(4),
+        )
+        x = x.permute(0, 1, 5, 2, 6, 3, 7, 4).contiguous()
+        x = x.view(
+            x.size(0),
+            self.out_channels,
+            x.size(2) * self.factor_t,
+            x.size(4) * self.factor_s,
+            x.size(6) * self.factor_s,
+        )
+        if first_chunk:
+            x = x[:, :, self.factor_t - 1:, :, :]
+        return x
+
+
+class Down_ResidualBlock(nn.Module):
+
+    def __init__(self,
+                 in_dim,
+                 out_dim,
+                 dropout,
+                 mult,
+                 temperal_downsample=False,
+                 down_flag=False):
+        super().__init__()
+
+        # Shortcut path with downsample
+        self.avg_shortcut = AvgDown3D(
+            in_dim,
+            out_dim,
+            factor_t=2 if temperal_downsample else 1,
+            factor_s=2 if down_flag else 1,
+        )
+
+        # Main path with residual blocks and downsample
+        downsamples = []
+        for _ in range(mult):
+            downsamples.append(ResidualBlock(in_dim, out_dim, dropout))
+            in_dim = out_dim
+
+        # Add the final downsample block
+        if down_flag:
+            mode = "downsample3d" if temperal_downsample else "downsample2d"
+            downsamples.append(Resample(out_dim, mode=mode))
+
+        self.downsamples = nn.Sequential(*downsamples)
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        x_copy = x
+        for module in self.downsamples:
+            x = module(x, feat_cache, feat_idx)
+
+        return x + self.avg_shortcut(x_copy)
+
+
+class Up_ResidualBlock(nn.Module):
+
+    def __init__(self,
+                 in_dim,
+                 out_dim,
+                 dropout,
+                 mult,
+                 temperal_upsample=False,
+                 up_flag=False):
+        super().__init__()
+        # Shortcut path with upsample
+        if up_flag:
+            self.avg_shortcut = DupUp3D(
+                in_dim,
+                out_dim,
+                factor_t=2 if temperal_upsample else 1,
+                factor_s=2 if up_flag else 1,
+            )
+        else:
+            self.avg_shortcut = None
+
+        # Main path with residual blocks and upsample
+        upsamples = []
+        for _ in range(mult):
+            upsamples.append(ResidualBlock(in_dim, out_dim, dropout))
+            in_dim = out_dim
+
+        # Add the final upsample block
+        if up_flag:
+            mode = "upsample3d" if temperal_upsample else "upsample2d"
+            upsamples.append(Resample(out_dim, mode=mode))
+
+        self.upsamples = nn.Sequential(*upsamples)
+
+    def forward(self, x, feat_cache=None, feat_idx=[0], first_chunk=False):
+        x_main = x
+        for module in self.upsamples:
+            x_main = module(x_main, feat_cache, feat_idx)
+        if self.avg_shortcut is not None:
+            x_shortcut = self.avg_shortcut(x, first_chunk)
+            return x_main + x_shortcut
+        else:
+            return x_main
+
+
+class Encoder3d(nn.Module):
+
+    def __init__(
+        self,
+        dim=128,
+        z_dim=4,
+        dim_mult=[1, 2, 4, 4],
+        num_res_blocks=2,
+        attn_scales=[],
+        temperal_downsample=[True, True, False],
+        dropout=0.0,
+    ):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_downsample = temperal_downsample
+
+        # dimensions
+        dims = [dim * u for u in [1] + dim_mult]
+        scale = 1.0
+
+        # init block
+        self.conv1 = CausalConv3d(12, dims[0], 3, padding=1)
+
+        # downsample blocks
+        downsamples = []
+        for i, (in_dim, out_dim) in enumerate(zip(dims[:-1], dims[1:])):
+            t_down_flag = (
+                temperal_downsample[i]
+                if i < len(temperal_downsample) else False)
+            downsamples.append(
+                Down_ResidualBlock(
+                    in_dim=in_dim,
+                    out_dim=out_dim,
+                    dropout=dropout,
+                    mult=num_res_blocks,
+                    temperal_downsample=t_down_flag,
+                    down_flag=i != len(dim_mult) - 1,
+                ))
+            scale /= 2.0
+        self.downsamples = nn.Sequential(*downsamples)
+
+        # middle blocks
+        self.middle = nn.Sequential(
+            ResidualBlock(out_dim, out_dim, dropout),
+            AttentionBlock(out_dim),
+            ResidualBlock(out_dim, out_dim, dropout),
+        )
+
+        # # output blocks
+        self.head = nn.Sequential(
+            RMS_norm(out_dim, images=False),
+            nn.SiLU(),
+            CausalConv3d(out_dim, z_dim, 3, padding=1),
+        )
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+
+        if feat_cache is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                cache_x = torch.cat(
+                    [
+                        feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                            cache_x.device),
+                        cache_x,
+                    ],
+                    dim=2,
+                )
+            x = self.conv1(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            x = self.conv1(x)
+
+        ## downsamples
+        for layer in self.downsamples:
+            if feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## middle
+        for layer in self.middle:
+            if isinstance(layer, ResidualBlock) and feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## head
+        for layer in self.head:
+            if isinstance(layer, CausalConv3d) and feat_cache is not None:
+                idx = feat_idx[0]
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                    cache_x = torch.cat(
+                        [
+                            feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                                cache_x.device),
+                            cache_x,
+                        ],
+                        dim=2,
+                    )
+                x = layer(x, feat_cache[idx])
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+            else:
+                x = layer(x)
+
+        return x
+
+
+class Decoder3d(nn.Module):
+
+    def __init__(
+        self,
+        dim=128,
+        z_dim=4,
+        dim_mult=[1, 2, 4, 4],
+        num_res_blocks=2,
+        attn_scales=[],
+        temperal_upsample=[False, True, True],
+        dropout=0.0,
+    ):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_upsample = temperal_upsample
+
+        # dimensions
+        dims = [dim * u for u in [dim_mult[-1]] + dim_mult[::-1]]
+        # init block
+        self.conv1 = CausalConv3d(z_dim, dims[0], 3, padding=1)
+
+        # middle blocks
+        self.middle = nn.Sequential(
+            ResidualBlock(dims[0], dims[0], dropout),
+            AttentionBlock(dims[0]),
+            ResidualBlock(dims[0], dims[0], dropout),
+        )
+
+        # upsample blocks
+        upsamples = []
+        for i, (in_dim, out_dim) in enumerate(zip(dims[:-1], dims[1:])):
+            t_up_flag = temperal_upsample[i] if i < len(
+                temperal_upsample) else False
+            upsamples.append(
+                Up_ResidualBlock(
+                    in_dim=in_dim,
+                    out_dim=out_dim,
+                    dropout=dropout,
+                    mult=num_res_blocks + 1,
+                    temperal_upsample=t_up_flag,
+                    up_flag=i != len(dim_mult) - 1,
+                ))
+        self.upsamples = nn.Sequential(*upsamples)
+
+        # output blocks
+        self.head = nn.Sequential(
+            RMS_norm(out_dim, images=False),
+            nn.SiLU(),
+            CausalConv3d(out_dim, 12, 3, padding=1),
+        )
+
+    def forward(self, x, feat_cache=None, feat_idx=[0], first_chunk=False):
+        if feat_cache is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                cache_x = torch.cat(
+                    [
+                        feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                            cache_x.device),
+                        cache_x,
+                    ],
+                    dim=2,
+                )
+            x = self.conv1(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            x = self.conv1(x)
+
+        for layer in self.middle:
+            if isinstance(layer, ResidualBlock) and feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## upsamples
+        for layer in self.upsamples:
+            if feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx, first_chunk)
+            else:
+                x = layer(x)
+
+        ## head
+        for layer in self.head:
+            if isinstance(layer, CausalConv3d) and feat_cache is not None:
+                idx = feat_idx[0]
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                    cache_x = torch.cat(
+                        [
+                            feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                                cache_x.device),
+                            cache_x,
+                        ],
+                        dim=2,
+                    )
+                x = layer(x, feat_cache[idx])
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+            else:
+                x = layer(x)
+        return x
+
+
+def count_conv3d(model):
+    count = 0
+    for m in model.modules():
+        if isinstance(m, CausalConv3d):
+            count += 1
+    return count
+
+
+class WanVAE(nn.Module):
+
+    def __init__(
+        self,
+        dim=160,
+        dec_dim=256,
+        z_dim=16,
+        dim_mult=[1, 2, 4, 4],
+        num_res_blocks=2,
+        attn_scales=[],
+        temperal_downsample=[True, True, False],
+        dropout=0.0,
+    ):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_downsample = temperal_downsample
+        self.temperal_upsample = temperal_downsample[::-1]
+
+        # modules
+        self.encoder = Encoder3d(
+            dim,
+            z_dim * 2,
+            dim_mult,
+            num_res_blocks,
+            attn_scales,
+            self.temperal_downsample,
+            dropout,
+        )
+        self.conv1 = CausalConv3d(z_dim * 2, z_dim * 2, 1)
+        self.conv2 = CausalConv3d(z_dim, z_dim, 1)
+        self.decoder = Decoder3d(
+            dec_dim,
+            z_dim,
+            dim_mult,
+            num_res_blocks,
+            attn_scales,
+            self.temperal_upsample,
+            dropout,
+        )
+
+    def encode(self, x):
+        conv_idx = [0]
+        feat_map = [None] * count_conv3d(self.encoder)
+        x = patchify(x, patch_size=2)
+        t = x.shape[2]
+        iter_ = 1 + (t - 1) // 4
+        for i in range(iter_):
+            conv_idx = [0]
+            if i == 0:
+                out = self.encoder(
+                    x[:, :, :1, :, :],
+                    feat_cache=feat_map,
+                    feat_idx=conv_idx,
+                )
+            else:
+                out_ = self.encoder(
+                    x[:, :, 1 + 4 * (i - 1):1 + 4 * i, :, :],
+                    feat_cache=feat_map,
+                    feat_idx=conv_idx,
+                )
+                out = torch.cat([out, out_], 2)
+        mu, log_var = self.conv1(out).chunk(2, dim=1)
+        return mu
+
+    def decode(self, z):
+        conv_idx = [0]
+        feat_map = [None] * count_conv3d(self.decoder)
+        iter_ = z.shape[2]
+        x = self.conv2(z)
+        for i in range(iter_):
+            conv_idx = [0]
+            if i == 0:
+                out = self.decoder(
+                    x[:, :, i:i + 1, :, :],
+                    feat_cache=feat_map,
+                    feat_idx=conv_idx,
+                    first_chunk=True,
+                )
+            else:
+                out_ = self.decoder(
+                    x[:, :, i:i + 1, :, :],
+                    feat_cache=feat_map,
+                    feat_idx=conv_idx,
+                )
+                out = torch.cat([out, out_], 2)
+        out = unpatchify(out, patch_size=2)
+        return out
+
+    def reparameterize(self, mu, log_var):
+        std = torch.exp(0.5 * log_var)
+        eps = torch.randn_like(std)
+        return eps * std + mu
+
+    def sample(self, imgs, deterministic=False):
+        mu, log_var = self.encode(imgs)
+        if deterministic:
+            return mu
+        std = torch.exp(0.5 * log_var.clamp(-30.0, 20.0))
+        return mu + std * torch.randn_like(std)