nn

backend/nn/t5.py

@@ -193,10 +193,10 @@ class T5(torch.nn.Module):
         self.encoder = T5Stack(self.num_layers, model_dim, model_dim, config["d_ff"], config["dense_act_fn"], config["is_gated_act"], config["num_heads"], config["model_type"] != "umt5")
         self.shared = torch.nn.Embedding(config["vocab_size"], model_dim)
 
-    def forward(self, input_ids, *args, **kwargs):
+    def forward(self, input_ids, attention_mask=None, *args, **kwargs):
         x = self.shared(input_ids)
         x = torch.nan_to_num(x)
-        return self.encoder(x, *args, **kwargs)
+        return self.encoder(x, attention_mask=attention_mask, *args, **kwargs)
 
 
 class IntegratedT5(torch.nn.Module):

backend/nn/vae.py

@@ -299,17 +299,14 @@ class IntegratedAutoencoderKL(nn.Module, ConfigMixin):
         if not isinstance(self.shift_factor, float):
             self.shift_factor = 0.0
 
-    def encode(self, x, regulation=None):
+    def encode(self, x):
         z = self.encoder(x)
 
         if self.quant_conv is not None:
             z = self.quant_conv(z)
 
         posterior = DiagonalGaussianDistribution(z)
-        if regulation is not None:
-            return regulation(posterior)
-        else:
-            return posterior.sample()
+        return posterior.sample()
 
     def decode(self, z):
         if self.post_quant_conv is not None:

backend/nn/wan.py

@@ -0,0 +1,482 @@
+# original version: https://github.com/Wan-Video/Wan2.1/blob/main/wan/modules/model.py
+# Copyright 2024-2025 The Alibaba Wan Team Authors. All rights reserved.
+import math
+
+import torch
+import torch.nn as nn
+from einops import repeat
+
+from backend.attention import attention_function as optimized_attention
+from backend.memory_management import cast_to_device
+from backend.nn.flux import EmbedND, apply_rope
+from backend.utils import pad_to_patch_size
+
+
+def sinusoidal_embedding_1d(dim, position):
+    # preprocess
+    assert dim % 2 == 0
+    half = dim // 2
+    position = position.type(torch.float32)
+
+    # calculation
+    sinusoid = torch.outer(position, torch.pow(10000, -torch.arange(half).to(position).div(half)))
+    x = torch.cat([torch.cos(sinusoid), torch.sin(sinusoid)], dim=1)
+    return x
+
+
+class WanSelfAttention(nn.Module):
+
+    def __init__(self, dim, num_heads, window_size=(-1, -1), qk_norm=True, eps=1e-6):
+        assert dim % num_heads == 0
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+        self.window_size = window_size
+        self.qk_norm = qk_norm
+        self.eps = eps
+
+        # layers
+        self.q = nn.Linear(dim, dim)
+        self.k = nn.Linear(dim, dim)
+        self.v = nn.Linear(dim, dim)
+        self.o = nn.Linear(dim, dim)
+        self.norm_q = nn.RMSNorm(dim, eps=eps) if qk_norm else nn.Identity()
+        self.norm_k = nn.RMSNorm(dim, eps=eps) if qk_norm else nn.Identity()
+
+    def forward(self, x, freqs):
+        r"""
+        Args:
+            x(Tensor): Shape [B, L, num_heads, C / num_heads]
+            freqs(Tensor): Rope freqs, shape [1024, C / num_heads / 2]
+        """
+        b, s, n, d = *x.shape[:2], self.num_heads, self.head_dim
+
+        # query, key, value function
+        def qkv_fn(x):
+            q = self.norm_q(self.q(x)).view(b, s, n, d)
+            k = self.norm_k(self.k(x)).view(b, s, n, d)
+            v = self.v(x).view(b, s, n * d)
+            return q, k, v
+
+        q, k, v = qkv_fn(x)
+        q, k = apply_rope(q, k, freqs)
+
+        x = optimized_attention(
+            q.view(b, s, n * d),
+            k.view(b, s, n * d),
+            v,
+            heads=self.num_heads,
+        )
+
+        x = self.o(x)
+        return x
+
+
+class WanT2VCrossAttention(WanSelfAttention):
+
+    def forward(self, x, context, **kwargs):
+        r"""
+        Args:
+            x(Tensor): Shape [B, L1, C]
+            context(Tensor): Shape [B, L2, C]
+        """
+        # compute query, key, value
+        q = self.norm_q(self.q(x))
+        k = self.norm_k(self.k(context))
+        v = self.v(context)
+
+        # compute attention
+        x = optimized_attention(q, k, v, heads=self.num_heads)
+
+        x = self.o(x)
+        return x
+
+
+class WanI2VCrossAttention(WanSelfAttention):
+
+    def __init__(self, dim, num_heads, window_size=(-1, -1), qk_norm=True, eps=1e-6):
+        super().__init__(dim, num_heads, window_size, qk_norm, eps)
+
+        self.k_img = nn.Linear(dim, dim)
+        self.v_img = nn.Linear(dim, dim)
+        # self.alpha = nn.Parameter(torch.zeros((1, )))
+        self.norm_k_img = nn.RMSNorm(dim, eps=eps) if qk_norm else nn.Identity()
+
+    def forward(self, x, context, context_img_len):
+        r"""
+        Args:
+            x(Tensor): Shape [B, L1, C]
+            context(Tensor): Shape [B, L2, C]
+        """
+        context_img = context[:, :context_img_len]
+        context = context[:, context_img_len:]
+
+        # compute query, key, value
+        q = self.norm_q(self.q(x))
+        k = self.norm_k(self.k(context))
+        v = self.v(context)
+        k_img = self.norm_k_img(self.k_img(context_img))
+        v_img = self.v_img(context_img)
+        img_x = optimized_attention(q, k_img, v_img, heads=self.num_heads)
+        # compute attention
+        x = optimized_attention(q, k, v, heads=self.num_heads)
+
+        # output
+        x = x + img_x
+        x = self.o(x)
+        return x
+
+
+WAN_CROSSATTENTION_CLASSES = {
+    "t2v_cross_attn": WanT2VCrossAttention,
+    "i2v_cross_attn": WanI2VCrossAttention,
+}
+
+
+def repeat_e(e, x):
+    repeats = 1
+    if e.shape[1] > 1:
+        repeats = x.shape[1] // e.shape[1]
+    if repeats == 1:
+        return e
+    return torch.repeat_interleave(e, repeats, dim=1)
+
+
+class WanAttentionBlock(nn.Module):
+
+    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):
+        super().__init__()
+        self.dim = dim
+        self.ffn_dim = ffn_dim
+        self.num_heads = num_heads
+        self.window_size = window_size
+        self.qk_norm = qk_norm
+        self.cross_attn_norm = cross_attn_norm
+        self.eps = eps
+
+        # layers
+        self.norm1 = nn.LayerNorm(dim, eps, elementwise_affine=False)
+        self.self_attn = WanSelfAttention(dim, num_heads, window_size, qk_norm, eps)
+        self.norm3 = nn.LayerNorm(dim, eps, elementwise_affine=True) if cross_attn_norm else nn.Identity()
+        self.cross_attn = WAN_CROSSATTENTION_CLASSES[cross_attn_type](dim, num_heads, (-1, -1), qk_norm, eps)
+        self.norm2 = nn.LayerNorm(dim, eps, elementwise_affine=False)
+        self.ffn = nn.Sequential(nn.Linear(dim, ffn_dim), nn.GELU(approximate="tanh"), nn.Linear(ffn_dim, dim))
+
+        # modulation
+        self.modulation = nn.Parameter(torch.empty(1, 6, dim))
+
+    def forward(
+        self,
+        x,
+        e,
+        freqs,
+        context,
+        context_img_len=257,
+    ):
+        r"""
+        Args:
+            x(Tensor): Shape [B, L, C]
+            e(Tensor): Shape [B, 6, C]
+            freqs(Tensor): Rope freqs, shape [1024, C / num_heads / 2]
+        """
+        # assert e.dtype == torch.float32
+
+        if e.ndim < 4:
+            e = (cast_to_device(self.modulation, dtype=x.dtype, device=x.device) + e).chunk(6, dim=1)
+        else:
+            e = (cast_to_device(self.modulation, dtype=x.dtype, device=x.device).unsqueeze(0) + e).unbind(2)
+        # assert e[0].dtype == torch.float32
+
+        # self-attention
+        y = self.self_attn(self.norm1(x) * (1 + repeat_e(e[1], x)) + repeat_e(e[0], x), freqs)
+
+        x = x + y * repeat_e(e[2], x)
+
+        # cross-attention & ffn
+        x = x + self.cross_attn(self.norm3(x), context, context_img_len=context_img_len)
+        y = self.ffn(self.norm2(x) * (1 + repeat_e(e[4], x)) + repeat_e(e[3], x))
+        x = x + y * repeat_e(e[5], x)
+        return x
+
+
+class Head(nn.Module):
+
+    def __init__(self, dim, out_dim, patch_size, eps=1e-6):
+        super().__init__()
+        self.dim = dim
+        self.out_dim = out_dim
+        self.patch_size = patch_size
+        self.eps = eps
+
+        # layers
+        out_dim = math.prod(patch_size) * out_dim
+        self.norm = nn.LayerNorm(dim, eps, elementwise_affine=False)
+        self.head = nn.Linear(dim, out_dim)
+
+        # modulation
+        self.modulation = nn.Parameter(torch.empty(1, 2, dim))
+
+    def forward(self, x, e):
+        r"""
+        Args:
+            x(Tensor): Shape [B, L1, C]
+            e(Tensor): Shape [B, C]
+        """
+        # assert e.dtype == torch.float32
+        if e.ndim < 3:
+            e = (cast_to_device(self.modulation, dtype=x.dtype, device=x.device) + e.unsqueeze(1)).chunk(2, dim=1)
+        else:
+            e = (cast_to_device(self.modulation, dtype=x.dtype, device=x.device).unsqueeze(0) + e.unsqueeze(2)).unbind(2)
+
+        x = self.head(self.norm(x) * (1 + repeat_e(e[1], x)) + repeat_e(e[0], x))
+        return x
+
+
+class MLPProj(nn.Module):
+
+    def __init__(self, in_dim, out_dim, flf_pos_embed_token_number=None):
+        super().__init__()
+
+        self.proj = nn.Sequential(nn.LayerNorm(in_dim), nn.Linear(in_dim, in_dim), nn.GELU(), nn.Linear(in_dim, out_dim), nn.LayerNorm(out_dim))
+
+        if flf_pos_embed_token_number is not None:
+            self.emb_pos = nn.Parameter(torch.empty((1, flf_pos_embed_token_number, in_dim)))
+        else:
+            self.emb_pos = None
+
+    def forward(self, image_embeds):
+        if self.emb_pos is not None:
+            image_embeds = image_embeds[:, : self.emb_pos.shape[1]] + cast_to_device(self.emb_pos[:, : image_embeds.shape[1]], dtype=image_embeds.dtype, device=image_embeds.device)
+
+        clip_extra_context_tokens = self.proj(image_embeds)
+        return clip_extra_context_tokens
+
+
+class WanModel(nn.Module):
+    r"""
+    Wan diffusion backbone supporting both text-to-video and image-to-video.
+    """
+
+    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,
+        flf_pos_embed_token_number=None,
+    ):
+        r"""
+        Initialize the diffusion model backbone.
+
+        Args:
+            model_type (`str`, *optional*, defaults to 't2v'):
+                Model variant - 't2v' (text-to-video) or 'i2v' (image-to-video)
+            patch_size (`tuple`, *optional*, defaults to (1, 2, 2)):
+                3D patch dimensions for video embedding (t_patch, h_patch, w_patch)
+            text_len (`int`, *optional*, defaults to 512):
+                Fixed length for text embeddings
+            in_dim (`int`, *optional*, defaults to 16):
+                Input video channels (C_in)
+            dim (`int`, *optional*, defaults to 2048):
+                Hidden dimension of the transformer
+            ffn_dim (`int`, *optional*, defaults to 8192):
+                Intermediate dimension in feed-forward network
+            freq_dim (`int`, *optional*, defaults to 256):
+                Dimension for sinusoidal time embeddings
+            text_dim (`int`, *optional*, defaults to 4096):
+                Input dimension for text embeddings
+            out_dim (`int`, *optional*, defaults to 16):
+                Output video channels (C_out)
+            num_heads (`int`, *optional*, defaults to 16):
+                Number of attention heads
+            num_layers (`int`, *optional*, defaults to 32):
+                Number of transformer blocks
+            window_size (`tuple`, *optional*, defaults to (-1, -1)):
+                Window size for local attention (-1 indicates global attention)
+            qk_norm (`bool`, *optional*, defaults to True):
+                Enable query/key normalization
+            cross_attn_norm (`bool`, *optional*, defaults to False):
+                Enable cross-attention normalization
+            eps (`float`, *optional*, defaults to 1e-6):
+                Epsilon value for normalization layers
+        """
+
+        super().__init__()
+
+        assert model_type in ["t2v", "i2v"]
+        self.model_type = model_type
+
+        self.patch_size = patch_size
+        self.text_len = text_len
+        self.in_dim = in_dim
+        self.dim = dim
+        self.ffn_dim = ffn_dim
+        self.freq_dim = freq_dim
+        self.text_dim = text_dim
+        self.out_dim = out_dim
+        self.num_heads = num_heads
+        self.num_layers = num_layers
+        self.window_size = window_size
+        self.qk_norm = qk_norm
+        self.cross_attn_norm = cross_attn_norm
+        self.eps = eps
+
+        # embeddings
+        self.patch_embedding = nn.Conv3d(in_dim, dim, kernel_size=patch_size, stride=patch_size)
+        self.text_embedding = nn.Sequential(nn.Linear(text_dim, dim), nn.GELU(approximate="tanh"), nn.Linear(dim, dim))
+
+        self.time_embedding = nn.Sequential(nn.Linear(freq_dim, dim), nn.SiLU(), nn.Linear(dim, dim))
+        self.time_projection = nn.Sequential(nn.SiLU(), nn.Linear(dim, dim * 6))
+
+        # blocks
+        cross_attn_type = "t2v_cross_attn" if model_type == "t2v" else "i2v_cross_attn"
+        self.blocks = nn.ModuleList([WanAttentionBlock(cross_attn_type, dim, ffn_dim, num_heads, window_size, qk_norm, cross_attn_norm, eps) for _ in range(num_layers)])
+
+        # head
+        self.head = Head(dim, out_dim, patch_size, eps)
+
+        d = dim // num_heads
+        self.rope_embedder = EmbedND(dim=d, theta=10000.0, axes_dim=[d - 4 * (d // 6), 2 * (d // 6), 2 * (d // 6)])
+
+        if model_type == "i2v":
+            self.img_emb = MLPProj(1280, dim, flf_pos_embed_token_number=flf_pos_embed_token_number)
+        else:
+            self.img_emb = None
+
+    def forward_orig(
+        self,
+        x,
+        t,
+        context,
+        clip_fea=None,
+        freqs=None,
+        transformer_options={},
+        **kwargs,
+    ):
+        r"""
+        Forward pass through the diffusion model
+
+        Args:
+            x (Tensor):
+                List of input video tensors with shape [B, C_in, F, H, W]
+            t (Tensor):
+                Diffusion timesteps tensor of shape [B]
+            context (List[Tensor]):
+                List of text embeddings each with shape [B, L, C]
+            seq_len (`int`):
+                Maximum sequence length for positional encoding
+            clip_fea (Tensor, *optional*):
+                CLIP image features for image-to-video mode
+            y (List[Tensor], *optional*):
+                Conditional video inputs for image-to-video mode, same shape as x
+
+        Returns:
+            List[Tensor]:
+                List of denoised video tensors with original input shapes [C_out, F, H / 8, W / 8]
+        """
+        # embeddings
+        x = self.patch_embedding(x).to(x.dtype)
+        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))
+
+        # 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", {})
+        for i, block in enumerate(self.blocks):
+            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)
+                    return out
+
+                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": e0, "pe": freqs}, {"original_block": block_wrap})
+                x = out["img"]
+            else:
+                x = block(x, e=e0, freqs=freqs, context=context, context_img_len=context_img_len)
+
+        # head
+        x = self.head(x, e)
+
+        # unpatchify
+        x = self.unpatchify(x, grid_sizes)
+        return x
+
+    def forward(self, x, timestep, context, clip_fea=None, time_dim_concat=None, transformer_options={}, **kwargs):
+        bs, c, t, h, w = x.shape
+
+        if c < self.in_dim:
+            assert hasattr(self, "z")
+            x = torch.cat((x, self.z.to(x)), dim=1)
+            bs, c, t, h, w = x.shape
+
+        x = pad_to_patch_size(x, self.patch_size)
+
+        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 time_dim_concat is not None:
+            time_dim_concat = pad_to_patch_size(time_dim_concat, self.patch_size)
+            x = torch.cat([x, time_dim_concat], dim=2)
+            t_len = (x.shape[2] + (patch_size[0] // 2)) // patch_size[0]
+
+        img_ids = torch.zeros((t_len, h_len, w_len, 3), device=x.device, dtype=x.dtype)
+        img_ids[:, :, :, 0] = img_ids[:, :, :, 0] + torch.linspace(0, t_len - 1, steps=t_len, device=x.device, dtype=x.dtype).reshape(-1, 1, 1)
+        img_ids[:, :, :, 1] = img_ids[:, :, :, 1] + torch.linspace(0, h_len - 1, steps=h_len, device=x.device, dtype=x.dtype).reshape(1, -1, 1)
+        img_ids[:, :, :, 2] = img_ids[:, :, :, 2] + torch.linspace(0, w_len - 1, steps=w_len, device=x.device, dtype=x.dtype).reshape(1, 1, -1)
+        img_ids = repeat(img_ids, "t h w c -> b (t h w) c", b=bs)
+
+        freqs = self.rope_embedder(img_ids).movedim(1, 2)
+        return self.forward_orig(x, timestep, context, clip_fea=clip_fea, freqs=freqs, transformer_options=transformer_options, **kwargs)[:, :, :t, :h, :w]
+
+    def unpatchify(self, x, grid_sizes):
+        r"""
+        Reconstruct video tensors from patch embeddings.
+
+        Args:
+            x (List[Tensor]):
+                List of patchified features, each with shape [L, C_out * prod(patch_size)]
+            grid_sizes (Tensor):
+                Original spatial-temporal grid dimensions before patching,
+                    shape [B, 3] (3 dimensions correspond to F_patches, H_patches, W_patches)
+
+        Returns:
+            List[Tensor]:
+                Reconstructed video tensors with shape [L, C_out, F, H / 8, W / 8]
+        """
+
+        c = self.out_dim
+        u = x
+        b = u.shape[0]
+        u = u[:, : math.prod(grid_sizes)].view(b, *grid_sizes, *self.patch_size, c)
+        u = torch.einsum("bfhwpqrc->bcfphqwr", u)
+        u = u.reshape(b, c, *[i * j for i, j in zip(grid_sizes, self.patch_size)])
+        return u

backend/nn/wan_vae.py

@@ -0,0 +1,438 @@
+# 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 diffusers.configuration_utils import ConfigMixin, register_to_config
+from einops import rearrange
+
+from backend.attention import attention_function_single_head_spatial as vae_attention
+from backend.operations import ForgeOperations as ops
+
+CACHE_T = 2
+
+
+class CausalConv3d(ops.Conv3d):
+    """
+    Causal 3d convolusion.
+    """
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self._padding = (self.padding[2], self.padding[2], self.padding[1], self.padding[1], 2 * self.padding[0], 0)
+        self.padding = (0, 0, 0)
+
+    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
+
+        padding = list(self._padding)
+        if cache_x is not None and self._padding[4] > 0:
+            cache_x = cache_x.to(x.device)
+            x = torch.cat([cache_x, x], dim=2)
+            padding[4] -= cache_x.shape[2]
+            del cache_x
+        x = F.pad(x, padding)
+
+        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=(2.0, 2.0), mode="nearest-exact"), ops.Conv2d(dim, dim // 2, 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 // 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()
+                    # 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)
+
+                    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)
+
+
+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):
+        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, dim_mult=[1, 2, 4, 4], num_res_blocks=2, attn_scales=[], temporal_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.temporal_downsample = temporal_downsample
+
+        # dimensions
+        dims = [dim * u for u in [1] + dim_mult]
+        scale = 1.0
+
+        # init block
+        self.conv1 = CausalConv3d(3, 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 temporal_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]):
+        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 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 = 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 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, 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=[], temporal_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.temporal_upsample = temporal_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 temporal_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, 3, 3, padding=1))
+
+    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:, :, :].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 = 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 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)
+            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 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, 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, ConfigMixin):
+    config_name = "config.json"
+
+    @register_to_config
+    def __init__(self, base_dim=128, z_dim=4, dim_mult=[1, 2, 4, 4], num_res_blocks=2, attn_scales=[], temporal_downsample=[True, True, False], dropout=0.0):
+        super().__init__()
+        self.dim = base_dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temporal_downsample = temporal_downsample
+        self.temporal_upsample = temporal_downsample[::-1]
+
+        # modules
+        self.encoder = Encoder3d(base_dim, z_dim * 2, dim_mult, num_res_blocks, attn_scales, self.temporal_downsample, dropout)
+        self.conv1 = CausalConv3d(z_dim * 2, z_dim * 2, 1)
+        self.conv2 = CausalConv3d(z_dim, z_dim, 1)
+        self.decoder = Decoder3d(base_dim, z_dim, dim_mult, num_res_blocks, attn_scales, self.temporal_upsample, dropout)
+
+        self.latent_format = None
+
+    def encode(self, x):
+        self.clear_cache()
+        ## cache
+        t = x.shape[2]
+        iter_ = 1 + (t - 1) // 4
+        ## split frames into (1, 4, 4, 4...)
+        for i in range(iter_):
+            self._enc_conv_idx = [0]
+            if i == 0:
+                out = self.encoder(x[:, :, :1, :, :], feat_cache=self._enc_feat_map, feat_idx=self._enc_conv_idx)
+            else:
+                out_ = self.encoder(x[:, :, 1 + 4 * (i - 1) : 1 + 4 * i, :, :], feat_cache=self._enc_feat_map, feat_idx=self._enc_conv_idx)
+                out = torch.cat([out, out_], 2)
+        mu, _ = self.conv1(out).chunk(2, dim=1)
+        self.clear_cache()
+        return mu
+
+    def decode(self, z):
+        self.clear_cache()
+        # z: [b, c, t, h, w]
+
+        iter_ = z.shape[2]
+        x = self.conv2(z)
+        for i in range(iter_):
+            self._conv_idx = [0]
+            if i == 0:
+                out = self.decoder(x[:, :, i : i + 1, :, :], feat_cache=self._feat_map, feat_idx=self._conv_idx)
+            else:
+                out_ = self.decoder(x[:, :, i : i + 1, :, :], feat_cache=self._feat_map, feat_idx=self._conv_idx)
+                out = torch.cat([out, out_], 2)
+        self.clear_cache()
+        return out
+
+    def clear_cache(self):
+        self._conv_num = count_conv3d(self.decoder)
+        self._conv_idx = [0]
+        self._feat_map = [None] * self._conv_num
+        # cache encode
+        self._enc_conv_num = count_conv3d(self.encoder)
+        self._enc_conv_idx = [0]
+        self._enc_feat_map = [None] * self._enc_conv_num
+
+    def process_in(self, latent):
+        return self.latent_format.process_in(latent)
+
+    def process_out(self, latent):
+        return self.latent_format.process_out(latent)