Track bundled vendor runtime sources

vendor/ComfyUI/comfy/ldm/cogvideo/model.py

@@ -0,0 +1,573 @@
+# CogVideoX 3D Transformer - ported to ComfyUI native ops
+# Architecture reference: diffusers CogVideoXTransformer3DModel
+# Style reference: comfy/ldm/wan/model.py
+
+import math
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from comfy.ldm.modules.attention import optimized_attention
+import comfy.patcher_extension
+import comfy.ldm.common_dit
+
+
+def _get_1d_rotary_pos_embed(dim, pos, theta=10000.0):
+    """Returns (cos, sin) each with shape [seq_len, dim].
+
+    Frequencies are computed at dim//2 resolution then repeat_interleaved
+    to full dim, matching CogVideoX's interleaved (real, imag) pair format.
+    """
+    freqs = 1.0 / (theta ** (torch.arange(0, dim, 2, dtype=torch.float32, device=pos.device) / dim))
+    angles = torch.outer(pos.float(), freqs.float())
+    cos = angles.cos().repeat_interleave(2, dim=-1).float()
+    sin = angles.sin().repeat_interleave(2, dim=-1).float()
+    return (cos, sin)
+
+
+def apply_rotary_emb(x, freqs_cos_sin):
+    """Apply CogVideoX rotary embedding to query or key tensor.
+
+    x: [B, heads, seq_len, head_dim]
+    freqs_cos_sin: (cos, sin) each [seq_len, head_dim//2]
+
+    Uses interleaved pair rotation (same as diffusers CogVideoX/Flux).
+    head_dim is reshaped to (-1, 2) pairs, rotated, then flattened back.
+    """
+    cos, sin = freqs_cos_sin
+    cos = cos[None, None, :, :].to(x.device)
+    sin = sin[None, None, :, :].to(x.device)
+
+    # Interleaved pairs: [B, H, S, D] -> [B, H, S, D//2, 2] -> (real, imag)
+    x_real, x_imag = x.reshape(*x.shape[:-1], -1, 2).unbind(-1)
+    x_rotated = torch.stack([-x_imag, x_real], dim=-1).flatten(3)
+
+    return (x.float() * cos + x_rotated.float() * sin).to(x.dtype)
+
+
+def get_timestep_embedding(timesteps, dim, flip_sin_to_cos=True, downscale_freq_shift=0, scale=1, max_period=10000):
+    half = dim // 2
+    freqs = torch.exp(-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32, device=timesteps.device) / half)
+    args = timesteps[:, None].float() * freqs[None] * scale
+    embedding = torch.cat([torch.sin(args), torch.cos(args)], dim=-1)
+    if flip_sin_to_cos:
+        embedding = torch.cat([embedding[:, half:], embedding[:, :half]], dim=-1)
+    if dim % 2:
+        embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
+    return embedding
+
+
+def get_3d_sincos_pos_embed(embed_dim, spatial_size, temporal_size, spatial_interpolation_scale=1.0, temporal_interpolation_scale=1.0, device=None):
+    if isinstance(spatial_size, int):
+        spatial_size = (spatial_size, spatial_size)
+
+    grid_w = torch.arange(spatial_size[0], dtype=torch.float32, device=device) / spatial_interpolation_scale
+    grid_h = torch.arange(spatial_size[1], dtype=torch.float32, device=device) / spatial_interpolation_scale
+    grid_t = torch.arange(temporal_size, dtype=torch.float32, device=device) / temporal_interpolation_scale
+
+    grid_t, grid_h, grid_w = torch.meshgrid(grid_t, grid_h, grid_w, indexing="ij")
+
+    embed_dim_spatial = 2 * (embed_dim // 3)
+    embed_dim_temporal = embed_dim // 3
+
+    pos_embed_spatial = _get_2d_sincos_pos_embed(embed_dim_spatial, grid_h, grid_w, device=device)
+    pos_embed_temporal = _get_1d_sincos_pos_embed(embed_dim_temporal, grid_t[:, 0, 0], device=device)
+
+    T, H, W = grid_t.shape
+    pos_embed_temporal = pos_embed_temporal.unsqueeze(1).unsqueeze(1).expand(-1, H, W, -1)
+    pos_embed = torch.cat([pos_embed_temporal, pos_embed_spatial], dim=-1)
+
+    return pos_embed
+
+
+def _get_2d_sincos_pos_embed(embed_dim, grid_h, grid_w, device=None):
+    T, H, W = grid_h.shape
+    half_dim = embed_dim // 2
+    pos_h = _get_1d_sincos_pos_embed(half_dim, grid_h.reshape(-1), device=device).reshape(T, H, W, half_dim)
+    pos_w = _get_1d_sincos_pos_embed(half_dim, grid_w.reshape(-1), device=device).reshape(T, H, W, half_dim)
+    return torch.cat([pos_h, pos_w], dim=-1)
+
+
+def _get_1d_sincos_pos_embed(embed_dim, pos, device=None):
+    half = embed_dim // 2
+    freqs = torch.exp(-math.log(10000.0) * torch.arange(start=0, end=half, dtype=torch.float32, device=device) / half)
+    args = pos.float().reshape(-1)[:, None] * freqs[None]
+    embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
+    if embed_dim % 2:
+        embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
+    return embedding
+
+
+
+class CogVideoXPatchEmbed(nn.Module):
+    def __init__(self, patch_size=2, patch_size_t=None, in_channels=16, dim=1920,
+                 text_dim=4096, bias=True, sample_width=90, sample_height=60,
+                 sample_frames=49, temporal_compression_ratio=4,
+                 max_text_seq_length=226, spatial_interpolation_scale=1.875,
+                 temporal_interpolation_scale=1.0, use_positional_embeddings=True,
+                 use_learned_positional_embeddings=True,
+                 device=None, dtype=None, operations=None):
+        super().__init__()
+        self.patch_size = patch_size
+        self.patch_size_t = patch_size_t
+        self.dim = dim
+        self.sample_height = sample_height
+        self.sample_width = sample_width
+        self.sample_frames = sample_frames
+        self.temporal_compression_ratio = temporal_compression_ratio
+        self.max_text_seq_length = max_text_seq_length
+        self.spatial_interpolation_scale = spatial_interpolation_scale
+        self.temporal_interpolation_scale = temporal_interpolation_scale
+        self.use_positional_embeddings = use_positional_embeddings
+        self.use_learned_positional_embeddings = use_learned_positional_embeddings
+
+        if patch_size_t is None:
+            self.proj = operations.Conv2d(in_channels, dim, kernel_size=patch_size, stride=patch_size, bias=bias, device=device, dtype=dtype)
+        else:
+            self.proj = operations.Linear(in_channels * patch_size * patch_size * patch_size_t, dim, device=device, dtype=dtype)
+
+        self.text_proj = operations.Linear(text_dim, dim, device=device, dtype=dtype)
+
+        if use_positional_embeddings or use_learned_positional_embeddings:
+            persistent = use_learned_positional_embeddings
+            pos_embedding = self._get_positional_embeddings(sample_height, sample_width, sample_frames)
+            self.register_buffer("pos_embedding", pos_embedding, persistent=persistent)
+
+    def _get_positional_embeddings(self, sample_height, sample_width, sample_frames, device=None):
+        post_patch_height = sample_height // self.patch_size
+        post_patch_width = sample_width // self.patch_size
+        post_time_compression_frames = (sample_frames - 1) // self.temporal_compression_ratio + 1
+        if self.patch_size_t is not None:
+            post_time_compression_frames = post_time_compression_frames // self.patch_size_t
+        num_patches = post_patch_height * post_patch_width * post_time_compression_frames
+
+        pos_embedding = get_3d_sincos_pos_embed(
+            self.dim,
+            (post_patch_width, post_patch_height),
+            post_time_compression_frames,
+            self.spatial_interpolation_scale,
+            self.temporal_interpolation_scale,
+            device=device,
+        )
+        pos_embedding = pos_embedding.reshape(-1, self.dim)
+        joint_pos_embedding = pos_embedding.new_zeros(
+            1, self.max_text_seq_length + num_patches, self.dim, requires_grad=False
+        )
+        joint_pos_embedding.data[:, self.max_text_seq_length:].copy_(pos_embedding)
+        return joint_pos_embedding
+
+    def forward(self, text_embeds, image_embeds):
+        input_dtype = text_embeds.dtype
+        text_embeds = self.text_proj(text_embeds.to(self.text_proj.weight.dtype)).to(input_dtype)
+        batch_size, num_frames, channels, height, width = image_embeds.shape
+
+        proj_dtype = self.proj.weight.dtype
+        if self.patch_size_t is None:
+            image_embeds = image_embeds.reshape(-1, channels, height, width)
+            image_embeds = self.proj(image_embeds.to(proj_dtype)).to(input_dtype)
+            image_embeds = image_embeds.view(batch_size, num_frames, *image_embeds.shape[1:])
+            image_embeds = image_embeds.flatten(3).transpose(2, 3)
+            image_embeds = image_embeds.flatten(1, 2)
+        else:
+            p = self.patch_size
+            p_t = self.patch_size_t
+            image_embeds = image_embeds.permute(0, 1, 3, 4, 2)
+            image_embeds = image_embeds.reshape(
+                batch_size, num_frames // p_t, p_t, height // p, p, width // p, p, channels
+            )
+            image_embeds = image_embeds.permute(0, 1, 3, 5, 7, 2, 4, 6).flatten(4, 7).flatten(1, 3)
+            image_embeds = self.proj(image_embeds.to(proj_dtype)).to(input_dtype)
+
+        embeds = torch.cat([text_embeds, image_embeds], dim=1).contiguous()
+
+        if self.use_positional_embeddings or self.use_learned_positional_embeddings:
+            text_seq_length = text_embeds.shape[1]
+            num_image_patches = image_embeds.shape[1]
+
+            if self.use_learned_positional_embeddings:
+                image_pos = self.pos_embedding[
+                    :, self.max_text_seq_length:self.max_text_seq_length + num_image_patches
+                ].to(device=embeds.device, dtype=embeds.dtype)
+            else:
+                image_pos = get_3d_sincos_pos_embed(
+                    self.dim,
+                    (width // self.patch_size, height // self.patch_size),
+                    num_image_patches // ((height // self.patch_size) * (width // self.patch_size)),
+                    self.spatial_interpolation_scale,
+                    self.temporal_interpolation_scale,
+                    device=embeds.device,
+                ).reshape(1, num_image_patches, self.dim).to(dtype=embeds.dtype)
+
+            # Build joint: zeros for text + sincos for image
+            joint_pos = torch.zeros(1, text_seq_length + num_image_patches, self.dim, device=embeds.device, dtype=embeds.dtype)
+            joint_pos[:, text_seq_length:] = image_pos
+            embeds = embeds + joint_pos
+
+        return embeds
+
+
+class CogVideoXLayerNormZero(nn.Module):
+    def __init__(self, time_dim, dim, elementwise_affine=True, eps=1e-5, bias=True,
+                 device=None, dtype=None, operations=None):
+        super().__init__()
+        self.silu = nn.SiLU()
+        self.linear = operations.Linear(time_dim, 6 * dim, bias=bias, device=device, dtype=dtype)
+        self.norm = operations.LayerNorm(dim, eps=eps, elementwise_affine=elementwise_affine, device=device, dtype=dtype)
+
+    def forward(self, hidden_states, encoder_hidden_states, temb):
+        shift, scale, gate, enc_shift, enc_scale, enc_gate = self.linear(self.silu(temb)).chunk(6, dim=1)
+        hidden_states = self.norm(hidden_states) * (1 + scale)[:, None, :] + shift[:, None, :]
+        encoder_hidden_states = self.norm(encoder_hidden_states) * (1 + enc_scale)[:, None, :] + enc_shift[:, None, :]
+        return hidden_states, encoder_hidden_states, gate[:, None, :], enc_gate[:, None, :]
+
+
+class CogVideoXAdaLayerNorm(nn.Module):
+    def __init__(self, time_dim, dim, elementwise_affine=True, eps=1e-5,
+                 device=None, dtype=None, operations=None):
+        super().__init__()
+        self.silu = nn.SiLU()
+        self.linear = operations.Linear(time_dim, 2 * dim, device=device, dtype=dtype)
+        self.norm = operations.LayerNorm(dim, eps=eps, elementwise_affine=elementwise_affine, device=device, dtype=dtype)
+
+    def forward(self, x, temb):
+        temb = self.linear(self.silu(temb))
+        shift, scale = temb.chunk(2, dim=1)
+        x = self.norm(x) * (1 + scale)[:, None, :] + shift[:, None, :]
+        return x
+
+
+class CogVideoXBlock(nn.Module):
+    def __init__(self, dim, num_heads, head_dim, time_dim,
+                 eps=1e-5, ff_inner_dim=None, ff_bias=True,
+                 device=None, dtype=None, operations=None):
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.head_dim = head_dim
+
+        self.norm1 = CogVideoXLayerNormZero(time_dim, dim, eps=eps, device=device, dtype=dtype, operations=operations)
+
+        # Self-attention (joint text + latent)
+        self.q = operations.Linear(dim, dim, bias=True, device=device, dtype=dtype)
+        self.k = operations.Linear(dim, dim, bias=True, device=device, dtype=dtype)
+        self.v = operations.Linear(dim, dim, bias=True, device=device, dtype=dtype)
+        self.norm_q = operations.LayerNorm(head_dim, eps=1e-6, elementwise_affine=True, device=device, dtype=dtype)
+        self.norm_k = operations.LayerNorm(head_dim, eps=1e-6, elementwise_affine=True, device=device, dtype=dtype)
+        self.attn_out = operations.Linear(dim, dim, bias=True, device=device, dtype=dtype)
+
+        self.norm2 = CogVideoXLayerNormZero(time_dim, dim, eps=eps, device=device, dtype=dtype, operations=operations)
+
+        # Feed-forward (GELU approximate)
+        inner_dim = ff_inner_dim or dim * 4
+        self.ff_proj = operations.Linear(dim, inner_dim, bias=ff_bias, device=device, dtype=dtype)
+        self.ff_out = operations.Linear(inner_dim, dim, bias=ff_bias, device=device, dtype=dtype)
+
+    def forward(self, hidden_states, encoder_hidden_states, temb, image_rotary_emb=None, transformer_options=None):
+        if transformer_options is None:
+            transformer_options = {}
+        text_seq_length = encoder_hidden_states.size(1)
+
+        # Norm & modulate
+        norm_hidden, norm_encoder, gate_msa, enc_gate_msa = self.norm1(hidden_states, encoder_hidden_states, temb)
+
+        # Joint self-attention
+        qkv_input = torch.cat([norm_encoder, norm_hidden], dim=1)
+        b, s, _ = qkv_input.shape
+        n, d = self.num_heads, self.head_dim
+
+        q = self.q(qkv_input).view(b, s, n, d)
+        k = self.k(qkv_input).view(b, s, n, d)
+        v = self.v(qkv_input)
+
+        q = self.norm_q(q).view(b, s, n, d)
+        k = self.norm_k(k).view(b, s, n, d)
+
+        # Apply rotary embeddings to image tokens only (diffusers format: [B, heads, seq, head_dim])
+        if image_rotary_emb is not None:
+            q_img = q[:, text_seq_length:].transpose(1, 2)  # [B, heads, img_seq, head_dim]
+            k_img = k[:, text_seq_length:].transpose(1, 2)
+            q_img = apply_rotary_emb(q_img, image_rotary_emb)
+            k_img = apply_rotary_emb(k_img, image_rotary_emb)
+            q = torch.cat([q[:, :text_seq_length], q_img.transpose(1, 2)], dim=1)
+            k = torch.cat([k[:, :text_seq_length], k_img.transpose(1, 2)], dim=1)
+
+        attn_out = optimized_attention(
+            q.reshape(b, s, n * d),
+            k.reshape(b, s, n * d),
+            v,
+            heads=self.num_heads,
+            transformer_options=transformer_options,
+        )
+
+        attn_out = self.attn_out(attn_out)
+
+        attn_encoder, attn_hidden = attn_out.split([text_seq_length, s - text_seq_length], dim=1)
+
+        hidden_states = hidden_states + gate_msa * attn_hidden
+        encoder_hidden_states = encoder_hidden_states + enc_gate_msa * attn_encoder
+
+        # Norm & modulate for FF
+        norm_hidden, norm_encoder, gate_ff, enc_gate_ff = self.norm2(hidden_states, encoder_hidden_states, temb)
+
+        # Feed-forward (GELU on concatenated text + latent)
+        ff_input = torch.cat([norm_encoder, norm_hidden], dim=1)
+        ff_output = self.ff_out(F.gelu(self.ff_proj(ff_input), approximate="tanh"))
+
+        hidden_states = hidden_states + gate_ff * ff_output[:, text_seq_length:]
+        encoder_hidden_states = encoder_hidden_states + enc_gate_ff * ff_output[:, :text_seq_length]
+
+        return hidden_states, encoder_hidden_states
+
+
+class CogVideoXTransformer3DModel(nn.Module):
+    def __init__(self,
+                 num_attention_heads=30,
+                 attention_head_dim=64,
+                 in_channels=16,
+                 out_channels=16,
+                 flip_sin_to_cos=True,
+                 freq_shift=0,
+                 time_embed_dim=512,
+                 ofs_embed_dim=None,
+                 text_embed_dim=4096,
+                 num_layers=30,
+                 dropout=0.0,
+                 attention_bias=True,
+                 sample_width=90,
+                 sample_height=60,
+                 sample_frames=49,
+                 patch_size=2,
+                 patch_size_t=None,
+                 temporal_compression_ratio=4,
+                 max_text_seq_length=226,
+                 spatial_interpolation_scale=1.875,
+                 temporal_interpolation_scale=1.0,
+                 use_rotary_positional_embeddings=False,
+                 use_learned_positional_embeddings=False,
+                 patch_bias=True,
+                 image_model=None,
+                 device=None,
+                 dtype=None,
+                 operations=None,
+                 ):
+        super().__init__()
+        self.dtype = dtype
+        dim = num_attention_heads * attention_head_dim
+        self.dim = dim
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.patch_size = patch_size
+        self.patch_size_t = patch_size_t
+        self.max_text_seq_length = max_text_seq_length
+        self.use_rotary_positional_embeddings = use_rotary_positional_embeddings
+
+        # 1. Patch embedding
+        self.patch_embed = CogVideoXPatchEmbed(
+            patch_size=patch_size,
+            patch_size_t=patch_size_t,
+            in_channels=in_channels,
+            dim=dim,
+            text_dim=text_embed_dim,
+            bias=patch_bias,
+            sample_width=sample_width,
+            sample_height=sample_height,
+            sample_frames=sample_frames,
+            temporal_compression_ratio=temporal_compression_ratio,
+            max_text_seq_length=max_text_seq_length,
+            spatial_interpolation_scale=spatial_interpolation_scale,
+            temporal_interpolation_scale=temporal_interpolation_scale,
+            use_positional_embeddings=not use_rotary_positional_embeddings,
+            use_learned_positional_embeddings=use_learned_positional_embeddings,
+            device=device, dtype=torch.float32, operations=operations,
+        )
+
+        # 2. Time embedding
+        self.time_proj_dim = dim
+        self.time_proj_flip = flip_sin_to_cos
+        self.time_proj_shift = freq_shift
+        self.time_embedding_linear_1 = operations.Linear(dim, time_embed_dim, device=device, dtype=dtype)
+        self.time_embedding_act = nn.SiLU()
+        self.time_embedding_linear_2 = operations.Linear(time_embed_dim, time_embed_dim, device=device, dtype=dtype)
+
+        # Optional OFS embedding (CogVideoX 1.5 I2V)
+        self.ofs_proj_dim = ofs_embed_dim
+        if ofs_embed_dim:
+            self.ofs_embedding_linear_1 = operations.Linear(ofs_embed_dim, ofs_embed_dim, device=device, dtype=dtype)
+            self.ofs_embedding_act = nn.SiLU()
+            self.ofs_embedding_linear_2 = operations.Linear(ofs_embed_dim, ofs_embed_dim, device=device, dtype=dtype)
+        else:
+            self.ofs_embedding_linear_1 = None
+
+        # 3. Transformer blocks
+        self.blocks = nn.ModuleList([
+            CogVideoXBlock(
+                dim=dim,
+                num_heads=num_attention_heads,
+                head_dim=attention_head_dim,
+                time_dim=time_embed_dim,
+                eps=1e-5,
+                device=device, dtype=dtype, operations=operations,
+            )
+            for _ in range(num_layers)
+        ])
+
+        self.norm_final = operations.LayerNorm(dim, eps=1e-5, elementwise_affine=True, device=device, dtype=dtype)
+
+        # 4. Output
+        self.norm_out = CogVideoXAdaLayerNorm(
+            time_dim=time_embed_dim, dim=dim, eps=1e-5,
+            device=device, dtype=dtype, operations=operations,
+        )
+
+        if patch_size_t is None:
+            output_dim = patch_size * patch_size * out_channels
+        else:
+            output_dim = patch_size * patch_size * patch_size_t * out_channels
+
+        self.proj_out = operations.Linear(dim, output_dim, device=device, dtype=dtype)
+
+        self.spatial_interpolation_scale = spatial_interpolation_scale
+        self.temporal_interpolation_scale = temporal_interpolation_scale
+        self.temporal_compression_ratio = temporal_compression_ratio
+
+    def forward(self, x, timestep, context, ofs=None, transformer_options=None, **kwargs):
+        if transformer_options is None:
+            transformer_options = {}
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options)
+        ).execute(x, timestep, context, ofs, transformer_options, **kwargs)
+
+    def _forward(self, x, timestep, context, ofs=None, transformer_options=None, **kwargs):
+        if transformer_options is None:
+            transformer_options = {}
+        # ComfyUI passes [B, C, T, H, W]
+        batch_size, channels, t, h, w = x.shape
+
+        # Pad to patch size (temporal + spatial), same pattern as WAN
+        p_t = self.patch_size_t if self.patch_size_t is not None else 1
+        x = comfy.ldm.common_dit.pad_to_patch_size(x, (p_t, self.patch_size, self.patch_size))
+
+        # CogVideoX expects [B, T, C, H, W]
+        x = x.permute(0, 2, 1, 3, 4)
+        batch_size, num_frames, channels, height, width = x.shape
+
+        # Time embedding
+        t_emb = get_timestep_embedding(timestep, self.time_proj_dim, self.time_proj_flip, self.time_proj_shift)
+        t_emb = t_emb.to(dtype=x.dtype)
+        emb = self.time_embedding_linear_2(self.time_embedding_act(self.time_embedding_linear_1(t_emb)))
+
+        if self.ofs_embedding_linear_1 is not None and ofs is not None:
+            ofs_emb = get_timestep_embedding(ofs, self.ofs_proj_dim, self.time_proj_flip, self.time_proj_shift)
+            ofs_emb = ofs_emb.to(dtype=x.dtype)
+            ofs_emb = self.ofs_embedding_linear_2(self.ofs_embedding_act(self.ofs_embedding_linear_1(ofs_emb)))
+            emb = emb + ofs_emb
+
+        # Patch embedding
+        hidden_states = self.patch_embed(context, x)
+
+        text_seq_length = context.shape[1]
+        encoder_hidden_states = hidden_states[:, :text_seq_length]
+        hidden_states = hidden_states[:, text_seq_length:]
+
+        # Rotary embeddings (if used)
+        image_rotary_emb = None
+        if self.use_rotary_positional_embeddings:
+            post_patch_height = height // self.patch_size
+            post_patch_width = width // self.patch_size
+            if self.patch_size_t is None:
+                post_time = num_frames
+            else:
+                post_time = num_frames // self.patch_size_t
+            image_rotary_emb = self._get_rotary_emb(post_patch_height, post_patch_width, post_time, device=x.device)
+
+        # Transformer blocks
+        for i, block in enumerate(self.blocks):
+            hidden_states, encoder_hidden_states = block(
+                hidden_states=hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                temb=emb,
+                image_rotary_emb=image_rotary_emb,
+                transformer_options=transformer_options,
+            )
+
+        hidden_states = self.norm_final(hidden_states)
+
+        # Output projection
+        hidden_states = self.norm_out(hidden_states, temb=emb)
+        hidden_states = self.proj_out(hidden_states)
+
+        # Unpatchify
+        p = self.patch_size
+        p_t = self.patch_size_t
+
+        if p_t is None:
+            output = hidden_states.reshape(batch_size, num_frames, height // p, width // p, -1, p, p)
+            output = output.permute(0, 1, 4, 2, 5, 3, 6).flatten(5, 6).flatten(3, 4)
+        else:
+            output = hidden_states.reshape(
+                batch_size, (num_frames + p_t - 1) // p_t, height // p, width // p, -1, p_t, p, p
+            )
+            output = output.permute(0, 1, 5, 4, 2, 6, 3, 7).flatten(6, 7).flatten(4, 5).flatten(1, 2)
+
+        # Back to ComfyUI format [B, C, T, H, W] and crop padding
+        output = output.permute(0, 2, 1, 3, 4)[:, :, :t, :h, :w]
+        return output
+
+    def _get_rotary_emb(self, h, w, t, device):
+        """Compute CogVideoX 3D rotary positional embeddings.
+
+        For CogVideoX 1.5 (patch_size_t != None): uses "slice" mode — grid positions
+        are integer arange computed at max_size, then sliced to actual size.
+        For CogVideoX 1.0 (patch_size_t == None): uses "linspace" mode with crop coords
+        scaled by spatial_interpolation_scale.
+        """
+        d = self.attention_head_dim
+        dim_t = d // 4
+        dim_h = d // 8 * 3
+        dim_w = d // 8 * 3
+
+        if self.patch_size_t is not None:
+            # CogVideoX 1.5: "slice" mode — positions are simple integer indices
+            # Compute at max(sample_size, actual_size) then slice to actual
+            base_h = self.patch_embed.sample_height // self.patch_size
+            base_w = self.patch_embed.sample_width // self.patch_size
+            max_h = max(base_h, h)
+            max_w = max(base_w, w)
+
+            grid_h = torch.arange(max_h, device=device, dtype=torch.float32)
+            grid_w = torch.arange(max_w, device=device, dtype=torch.float32)
+            grid_t = torch.arange(t, device=device, dtype=torch.float32)
+        else:
+            # CogVideoX 1.0: "linspace" mode with interpolation scale
+            grid_h = torch.linspace(0, h - 1, h, device=device, dtype=torch.float32) * self.spatial_interpolation_scale
+            grid_w = torch.linspace(0, w - 1, w, device=device, dtype=torch.float32) * self.spatial_interpolation_scale
+            grid_t = torch.arange(t, device=device, dtype=torch.float32)
+
+        freqs_t = _get_1d_rotary_pos_embed(dim_t, grid_t)
+        freqs_h = _get_1d_rotary_pos_embed(dim_h, grid_h)
+        freqs_w = _get_1d_rotary_pos_embed(dim_w, grid_w)
+
+        t_cos, t_sin = freqs_t
+        h_cos, h_sin = freqs_h
+        w_cos, w_sin = freqs_w
+
+        # Slice to actual size (for "slice" mode where grids may be larger)
+        t_cos, t_sin = t_cos[:t], t_sin[:t]
+        h_cos, h_sin = h_cos[:h], h_sin[:h]
+        w_cos, w_sin = w_cos[:w], w_sin[:w]
+
+        # Broadcast and concatenate into [T*H*W, head_dim]
+        t_cos = t_cos[:, None, None, :].expand(-1, h, w, -1)
+        t_sin = t_sin[:, None, None, :].expand(-1, h, w, -1)
+        h_cos = h_cos[None, :, None, :].expand(t, -1, w, -1)
+        h_sin = h_sin[None, :, None, :].expand(t, -1, w, -1)
+        w_cos = w_cos[None, None, :, :].expand(t, h, -1, -1)
+        w_sin = w_sin[None, None, :, :].expand(t, h, -1, -1)
+
+        cos = torch.cat([t_cos, h_cos, w_cos], dim=-1).reshape(t * h * w, -1)
+        sin = torch.cat([t_sin, h_sin, w_sin], dim=-1).reshape(t * h * w, -1)
+        return (cos, sin)

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

@@ -0,0 +1,566 @@
+# CogVideoX VAE - ported to ComfyUI native ops
+# Architecture reference: diffusers AutoencoderKLCogVideoX
+# Style reference: comfy/ldm/wan/vae.py
+
+import numpy as np
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+import comfy.ops
+ops = comfy.ops.disable_weight_init
+
+
+class CausalConv3d(nn.Module):
+    """Causal 3D convolution with temporal padding.
+
+    Uses comfy.ops.Conv3d with autopad='causal_zero' fast path: when input has
+    a single temporal frame and no cache, the 3D conv weight is sliced to act
+    as a 2D conv, avoiding computation on zero-padded temporal dimensions.
+    """
+    def __init__(self, in_channels, out_channels, kernel_size, stride=1, dilation=1, pad_mode="constant"):
+        super().__init__()
+        if isinstance(kernel_size, int):
+            kernel_size = (kernel_size,) * 3
+
+        time_kernel, height_kernel, width_kernel = kernel_size
+        self.time_kernel_size = time_kernel
+        self.pad_mode = pad_mode
+
+        height_pad = (height_kernel - 1) // 2
+        width_pad = (width_kernel - 1) // 2
+        self.time_causal_padding = (width_pad, width_pad, height_pad, height_pad, time_kernel - 1, 0)
+
+        stride = stride if isinstance(stride, tuple) else (stride, 1, 1)
+        dilation = (dilation, 1, 1)
+        self.conv = ops.Conv3d(
+            in_channels, out_channels, kernel_size,
+            stride=stride, dilation=dilation,
+            padding=(0, height_pad, width_pad),
+        )
+
+    def forward(self, x, conv_cache=None):
+        if self.pad_mode == "replicate":
+            x = F.pad(x, self.time_causal_padding, mode="replicate")
+            conv_cache = None
+        else:
+            kernel_t = self.time_kernel_size
+            if kernel_t > 1:
+                if conv_cache is None and x.shape[2] == 1:
+                    # Fast path: single frame, no cache. All temporal padding
+                    # frames are copies of the input (replicate-style), so the
+                    # 3D conv reduces to a 2D conv with summed temporal kernel.
+                    w = comfy.ops.cast_to_input(self.conv.weight, x)
+                    b = comfy.ops.cast_to_input(self.conv.bias, x) if self.conv.bias is not None else None
+                    w2d = w.sum(dim=2, keepdim=True)
+                    out = F.conv3d(x, w2d, b,
+                                   self.conv.stride, self.conv.padding,
+                                   self.conv.dilation, self.conv.groups)
+                    return out, None
+                cached = [conv_cache] if conv_cache is not None else [x[:, :, :1]] * (kernel_t - 1)
+                x = torch.cat(cached + [x], dim=2)
+            conv_cache = x[:, :, -self.time_kernel_size + 1:].clone() if self.time_kernel_size > 1 else None
+
+        out = self.conv(x)
+        return out, conv_cache
+
+
+def _interpolate_zq(zq, target_size):
+    """Interpolate latent z to target (T, H, W), matching CogVideoX's first-frame-special handling."""
+    t = target_size[0]
+    if t > 1 and t % 2 == 1:
+        z_first = F.interpolate(zq[:, :, :1], size=(1, target_size[1], target_size[2]))
+        z_rest = F.interpolate(zq[:, :, 1:], size=(t - 1, target_size[1], target_size[2]))
+        return torch.cat([z_first, z_rest], dim=2)
+    return F.interpolate(zq, size=target_size)
+
+
+class SpatialNorm3D(nn.Module):
+    """Spatially conditioned normalization."""
+    def __init__(self, f_channels, zq_channels, groups=32):
+        super().__init__()
+        self.norm_layer = ops.GroupNorm(num_channels=f_channels, num_groups=groups, eps=1e-6, affine=True)
+        self.conv_y = CausalConv3d(zq_channels, f_channels, kernel_size=1, stride=1)
+        self.conv_b = CausalConv3d(zq_channels, f_channels, kernel_size=1, stride=1)
+
+    def forward(self, f, zq, conv_cache=None):
+        new_cache = {}
+        conv_cache = conv_cache or {}
+
+        if zq.shape[-3:] != f.shape[-3:]:
+            zq = _interpolate_zq(zq, f.shape[-3:])
+
+        conv_y, new_cache["conv_y"] = self.conv_y(zq, conv_cache=conv_cache.get("conv_y"))
+        conv_b, new_cache["conv_b"] = self.conv_b(zq, conv_cache=conv_cache.get("conv_b"))
+
+        return self.norm_layer(f) * conv_y + conv_b, new_cache
+
+
+class ResnetBlock3D(nn.Module):
+    """3D ResNet block with optional spatial norm."""
+    def __init__(self, in_channels, out_channels=None, temb_channels=512, groups=32,
+                 eps=1e-6, act_fn="silu", spatial_norm_dim=None, pad_mode="first"):
+        super().__init__()
+        out_channels = out_channels or in_channels
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.spatial_norm_dim = spatial_norm_dim
+
+        if act_fn == "silu":
+            self.nonlinearity = nn.SiLU()
+        elif act_fn == "swish":
+            self.nonlinearity = nn.SiLU()
+        else:
+            self.nonlinearity = nn.SiLU()
+
+        if spatial_norm_dim is None:
+            self.norm1 = ops.GroupNorm(num_channels=in_channels, num_groups=groups, eps=eps)
+            self.norm2 = ops.GroupNorm(num_channels=out_channels, num_groups=groups, eps=eps)
+        else:
+            self.norm1 = SpatialNorm3D(in_channels, spatial_norm_dim, groups=groups)
+            self.norm2 = SpatialNorm3D(out_channels, spatial_norm_dim, groups=groups)
+
+        self.conv1 = CausalConv3d(in_channels, out_channels, kernel_size=3, pad_mode=pad_mode)
+
+        if temb_channels > 0:
+            self.temb_proj = ops.Linear(temb_channels, out_channels)
+
+        self.conv2 = CausalConv3d(out_channels, out_channels, kernel_size=3, pad_mode=pad_mode)
+
+        if in_channels != out_channels:
+            self.conv_shortcut = ops.Conv3d(in_channels, out_channels, kernel_size=1, stride=1, padding=0)
+        else:
+            self.conv_shortcut = None
+
+    def forward(self, x, temb=None, zq=None, conv_cache=None):
+        new_cache = {}
+        conv_cache = conv_cache or {}
+        residual = x
+
+        if zq is not None:
+            x, new_cache["norm1"] = self.norm1(x, zq, conv_cache=conv_cache.get("norm1"))
+        else:
+            x = self.norm1(x)
+
+        x = self.nonlinearity(x)
+        x, new_cache["conv1"] = self.conv1(x, conv_cache=conv_cache.get("conv1"))
+
+        if temb is not None and hasattr(self, "temb_proj"):
+            x = x + self.temb_proj(self.nonlinearity(temb))[:, :, None, None, None]
+
+        if zq is not None:
+            x, new_cache["norm2"] = self.norm2(x, zq, conv_cache=conv_cache.get("norm2"))
+        else:
+            x = self.norm2(x)
+
+        x = self.nonlinearity(x)
+        x, new_cache["conv2"] = self.conv2(x, conv_cache=conv_cache.get("conv2"))
+
+        if self.conv_shortcut is not None:
+            residual = self.conv_shortcut(residual)
+
+        return x + residual, new_cache
+
+
+class Downsample3D(nn.Module):
+    """3D downsampling with optional temporal compression."""
+    def __init__(self, in_channels, out_channels, kernel_size=3, stride=2, padding=0, compress_time=False):
+        super().__init__()
+        self.conv = ops.Conv2d(in_channels, out_channels, kernel_size=kernel_size, stride=stride, padding=padding)
+        self.compress_time = compress_time
+
+    def forward(self, x):
+        if self.compress_time:
+            b, c, t, h, w = x.shape
+            x = x.permute(0, 3, 4, 1, 2).reshape(b * h * w, c, t)
+            if t % 2 == 1:
+                x_first, x_rest = x[..., 0], x[..., 1:]
+                if x_rest.shape[-1] > 0:
+                    x_rest = F.avg_pool1d(x_rest, kernel_size=2, stride=2)
+                x = torch.cat([x_first[..., None], x_rest], dim=-1)
+                x = x.reshape(b, h, w, c, x.shape[-1]).permute(0, 3, 4, 1, 2)
+            else:
+                x = F.avg_pool1d(x, kernel_size=2, stride=2)
+                x = x.reshape(b, h, w, c, x.shape[-1]).permute(0, 3, 4, 1, 2)
+
+        pad = (0, 1, 0, 1)
+        x = F.pad(x, pad, mode="constant", value=0)
+        b, c, t, h, w = x.shape
+        x = x.permute(0, 2, 1, 3, 4).reshape(b * t, c, h, w)
+        x = self.conv(x)
+        x = x.reshape(b, t, x.shape[1], x.shape[2], x.shape[3]).permute(0, 2, 1, 3, 4)
+        return x
+
+
+class Upsample3D(nn.Module):
+    """3D upsampling with optional temporal decompression."""
+    def __init__(self, in_channels, out_channels, kernel_size=3, stride=1, padding=1, compress_time=False):
+        super().__init__()
+        self.conv = ops.Conv2d(in_channels, out_channels, kernel_size=kernel_size, stride=stride, padding=padding)
+        self.compress_time = compress_time
+
+    def forward(self, x):
+        if self.compress_time:
+            if x.shape[2] > 1 and x.shape[2] % 2 == 1:
+                x_first, x_rest = x[:, :, 0], x[:, :, 1:]
+                x_first = F.interpolate(x_first, scale_factor=2.0)
+                x_rest = F.interpolate(x_rest, scale_factor=2.0)
+                x = torch.cat([x_first[:, :, None, :, :], x_rest], dim=2)
+            elif x.shape[2] > 1:
+                x = F.interpolate(x, scale_factor=2.0)
+            else:
+                x = x.squeeze(2)
+                x = F.interpolate(x, scale_factor=2.0)
+                x = x[:, :, None, :, :]
+        else:
+            b, c, t, h, w = x.shape
+            x = x.permute(0, 2, 1, 3, 4).reshape(b * t, c, h, w)
+            x = F.interpolate(x, scale_factor=2.0)
+            x = x.reshape(b, t, c, *x.shape[2:]).permute(0, 2, 1, 3, 4)
+
+        b, c, t, h, w = x.shape
+        x = x.permute(0, 2, 1, 3, 4).reshape(b * t, c, h, w)
+        x = self.conv(x)
+        x = x.reshape(b, t, *x.shape[1:]).permute(0, 2, 1, 3, 4)
+        return x
+
+
+class DownBlock3D(nn.Module):
+    def __init__(self, in_channels, out_channels, temb_channels=0, num_layers=1,
+                 eps=1e-6, act_fn="silu", groups=32, add_downsample=True,
+                 compress_time=False, pad_mode="first"):
+        super().__init__()
+        self.resnets = nn.ModuleList([
+            ResnetBlock3D(
+                in_channels=in_channels if i == 0 else out_channels,
+                out_channels=out_channels,
+                temb_channels=temb_channels,
+                groups=groups, eps=eps, act_fn=act_fn, pad_mode=pad_mode,
+            )
+            for i in range(num_layers)
+        ])
+        self.downsamplers = nn.ModuleList([Downsample3D(out_channels, out_channels, compress_time=compress_time)]) if add_downsample else None
+
+    def forward(self, x, temb=None, zq=None, conv_cache=None):
+        new_cache = {}
+        conv_cache = conv_cache or {}
+        for i, resnet in enumerate(self.resnets):
+            x, new_cache[f"resnet_{i}"] = resnet(x, temb, zq, conv_cache=conv_cache.get(f"resnet_{i}"))
+        if self.downsamplers is not None:
+            for ds in self.downsamplers:
+                x = ds(x)
+        return x, new_cache
+
+
+class MidBlock3D(nn.Module):
+    def __init__(self, in_channels, temb_channels=0, num_layers=1,
+                 eps=1e-6, act_fn="silu", groups=32, spatial_norm_dim=None, pad_mode="first"):
+        super().__init__()
+        self.resnets = nn.ModuleList([
+            ResnetBlock3D(
+                in_channels=in_channels, out_channels=in_channels,
+                temb_channels=temb_channels, groups=groups, eps=eps,
+                act_fn=act_fn, spatial_norm_dim=spatial_norm_dim, pad_mode=pad_mode,
+            )
+            for _ in range(num_layers)
+        ])
+
+    def forward(self, x, temb=None, zq=None, conv_cache=None):
+        new_cache = {}
+        conv_cache = conv_cache or {}
+        for i, resnet in enumerate(self.resnets):
+            x, new_cache[f"resnet_{i}"] = resnet(x, temb, zq, conv_cache=conv_cache.get(f"resnet_{i}"))
+        return x, new_cache
+
+
+class UpBlock3D(nn.Module):
+    def __init__(self, in_channels, out_channels, temb_channels=0, num_layers=1,
+                 eps=1e-6, act_fn="silu", groups=32, spatial_norm_dim=16,
+                 add_upsample=True, compress_time=False, pad_mode="first"):
+        super().__init__()
+        self.resnets = nn.ModuleList([
+            ResnetBlock3D(
+                in_channels=in_channels if i == 0 else out_channels,
+                out_channels=out_channels,
+                temb_channels=temb_channels, groups=groups, eps=eps,
+                act_fn=act_fn, spatial_norm_dim=spatial_norm_dim, pad_mode=pad_mode,
+            )
+            for i in range(num_layers)
+        ])
+        self.upsamplers = nn.ModuleList([Upsample3D(out_channels, out_channels, compress_time=compress_time)]) if add_upsample else None
+
+    def forward(self, x, temb=None, zq=None, conv_cache=None):
+        new_cache = {}
+        conv_cache = conv_cache or {}
+        for i, resnet in enumerate(self.resnets):
+            x, new_cache[f"resnet_{i}"] = resnet(x, temb, zq, conv_cache=conv_cache.get(f"resnet_{i}"))
+        if self.upsamplers is not None:
+            for us in self.upsamplers:
+                x = us(x)
+        return x, new_cache
+
+
+class Encoder3D(nn.Module):
+    def __init__(self, in_channels=3, out_channels=16,
+                 block_out_channels=(128, 256, 256, 512),
+                 layers_per_block=3, act_fn="silu",
+                 eps=1e-6, groups=32, pad_mode="first",
+                 temporal_compression_ratio=4):
+        super().__init__()
+        temporal_compress_level = int(np.log2(temporal_compression_ratio))
+
+        self.conv_in = CausalConv3d(in_channels, block_out_channels[0], kernel_size=3, pad_mode=pad_mode)
+
+        self.down_blocks = nn.ModuleList()
+        output_channel = block_out_channels[0]
+        for i in range(len(block_out_channels)):
+            input_channel = output_channel
+            output_channel = block_out_channels[i]
+            is_final = i == len(block_out_channels) - 1
+            compress_time = i < temporal_compress_level
+
+            self.down_blocks.append(DownBlock3D(
+                in_channels=input_channel, out_channels=output_channel,
+                temb_channels=0, num_layers=layers_per_block,
+                eps=eps, act_fn=act_fn, groups=groups,
+                add_downsample=not is_final, compress_time=compress_time,
+            ))
+
+        self.mid_block = MidBlock3D(
+            in_channels=block_out_channels[-1], temb_channels=0,
+            num_layers=2, eps=eps, act_fn=act_fn, groups=groups, pad_mode=pad_mode,
+        )
+
+        self.norm_out = ops.GroupNorm(groups, block_out_channels[-1], eps=1e-6)
+        self.conv_act = nn.SiLU()
+        self.conv_out = CausalConv3d(block_out_channels[-1], 2 * out_channels, kernel_size=3, pad_mode=pad_mode)
+
+    def forward(self, x, conv_cache=None):
+        new_cache = {}
+        conv_cache = conv_cache or {}
+
+        x, new_cache["conv_in"] = self.conv_in(x, conv_cache=conv_cache.get("conv_in"))
+
+        for i, block in enumerate(self.down_blocks):
+            key = f"down_block_{i}"
+            x, new_cache[key] = block(x, None, None, conv_cache.get(key))
+
+        x, new_cache["mid_block"] = self.mid_block(x, None, None, conv_cache=conv_cache.get("mid_block"))
+
+        x = self.norm_out(x)
+        x = self.conv_act(x)
+        x, new_cache["conv_out"] = self.conv_out(x, conv_cache=conv_cache.get("conv_out"))
+
+        return x, new_cache
+
+
+class Decoder3D(nn.Module):
+    def __init__(self, in_channels=16, out_channels=3,
+                 block_out_channels=(128, 256, 256, 512),
+                 layers_per_block=3, act_fn="silu",
+                 eps=1e-6, groups=32, pad_mode="first",
+                 temporal_compression_ratio=4):
+        super().__init__()
+        reversed_channels = list(reversed(block_out_channels))
+        temporal_compress_level = int(np.log2(temporal_compression_ratio))
+
+        self.conv_in = CausalConv3d(in_channels, reversed_channels[0], kernel_size=3, pad_mode=pad_mode)
+
+        self.mid_block = MidBlock3D(
+            in_channels=reversed_channels[0], temb_channels=0,
+            num_layers=2, eps=eps, act_fn=act_fn, groups=groups,
+            spatial_norm_dim=in_channels, pad_mode=pad_mode,
+        )
+
+        self.up_blocks = nn.ModuleList()
+        output_channel = reversed_channels[0]
+        for i in range(len(block_out_channels)):
+            prev_channel = output_channel
+            output_channel = reversed_channels[i]
+            is_final = i == len(block_out_channels) - 1
+            compress_time = i < temporal_compress_level
+
+            self.up_blocks.append(UpBlock3D(
+                in_channels=prev_channel, out_channels=output_channel,
+                temb_channels=0, num_layers=layers_per_block + 1,
+                eps=eps, act_fn=act_fn, groups=groups,
+                spatial_norm_dim=in_channels,
+                add_upsample=not is_final, compress_time=compress_time,
+            ))
+
+        self.norm_out = SpatialNorm3D(reversed_channels[-1], in_channels, groups=groups)
+        self.conv_act = nn.SiLU()
+        self.conv_out = CausalConv3d(reversed_channels[-1], out_channels, kernel_size=3, pad_mode=pad_mode)
+
+    def forward(self, sample, conv_cache=None):
+        new_cache = {}
+        conv_cache = conv_cache or {}
+
+        x, new_cache["conv_in"] = self.conv_in(sample, conv_cache=conv_cache.get("conv_in"))
+
+        x, new_cache["mid_block"] = self.mid_block(x, None, sample, conv_cache=conv_cache.get("mid_block"))
+
+        for i, block in enumerate(self.up_blocks):
+            key = f"up_block_{i}"
+            x, new_cache[key] = block(x, None, sample, conv_cache=conv_cache.get(key))
+
+        x, new_cache["norm_out"] = self.norm_out(x, sample, conv_cache=conv_cache.get("norm_out"))
+        x = self.conv_act(x)
+        x, new_cache["conv_out"] = self.conv_out(x, conv_cache=conv_cache.get("conv_out"))
+
+        return x, new_cache
+
+
+
+class AutoencoderKLCogVideoX(nn.Module):
+    """CogVideoX VAE. Spatial tiling/slicing handled by ComfyUI's VAE wrapper.
+
+    Uses rolling temporal decode: conv_in + mid_block + temporal up_blocks run
+    on the full (low-res) tensor, then the expensive spatial-only up_blocks +
+    norm_out + conv_out are processed in small temporal chunks with conv_cache
+    carrying causal state between chunks. This keeps peak VRAM proportional to
+    chunk_size rather than total frame count.
+    """
+
+    def __init__(self,
+                 in_channels=3, out_channels=3,
+                 block_out_channels=(128, 256, 256, 512),
+                 latent_channels=16, layers_per_block=3,
+                 act_fn="silu", eps=1e-6, groups=32,
+                 temporal_compression_ratio=4,
+                 ):
+        super().__init__()
+        self.latent_channels = latent_channels
+        self.temporal_compression_ratio = temporal_compression_ratio
+
+        self.encoder = Encoder3D(
+            in_channels=in_channels, out_channels=latent_channels,
+            block_out_channels=block_out_channels, layers_per_block=layers_per_block,
+            act_fn=act_fn, eps=eps, groups=groups,
+            temporal_compression_ratio=temporal_compression_ratio,
+        )
+        self.decoder = Decoder3D(
+            in_channels=latent_channels, out_channels=out_channels,
+            block_out_channels=block_out_channels, layers_per_block=layers_per_block,
+            act_fn=act_fn, eps=eps, groups=groups,
+            temporal_compression_ratio=temporal_compression_ratio,
+        )
+
+        self.num_latent_frames_batch_size = 2
+        self.num_sample_frames_batch_size = 8
+
+    def encode(self, x):
+        t = x.shape[2]
+        frame_batch = self.num_sample_frames_batch_size
+        remainder = t % frame_batch
+        conv_cache = None
+        enc = []
+
+        # Process remainder frames first so only the first chunk can have an
+        # odd temporal dimension — where Downsample3D's first-frame-special
+        # handling in temporal compression is actually correct.
+        if remainder > 0:
+            chunk, conv_cache = self.encoder(x[:, :, :remainder], conv_cache=conv_cache)
+            enc.append(chunk.to(x.device))
+
+        for start in range(remainder, t, frame_batch):
+            chunk, conv_cache = self.encoder(x[:, :, start:start + frame_batch], conv_cache=conv_cache)
+            enc.append(chunk.to(x.device))
+
+        enc = torch.cat(enc, dim=2)
+        mean, _ = enc.chunk(2, dim=1)
+        return mean
+
+    def decode(self, z):
+        return self._decode_rolling(z)
+
+    def _decode_batched(self, z):
+        """Original batched decode - processes 2 latent frames through full decoder."""
+        t = z.shape[2]
+        frame_batch = self.num_latent_frames_batch_size
+        num_batches = max(t // frame_batch, 1)
+        conv_cache = None
+        dec = []
+        for i in range(num_batches):
+            remaining = t % frame_batch
+            start = frame_batch * i + (0 if i == 0 else remaining)
+            end = frame_batch * (i + 1) + remaining
+            chunk, conv_cache = self.decoder(z[:, :, start:end], conv_cache=conv_cache)
+            dec.append(chunk.cpu())
+        return torch.cat(dec, dim=2).to(z.device)
+
+    def _decode_rolling(self, z):
+        """Rolling decode - processes low-res layers on full tensor, then rolls
+        through expensive high-res layers in temporal chunks."""
+        decoder = self.decoder
+        device = z.device
+
+        # Determine which up_blocks have temporal upsample vs spatial-only.
+        # Temporal up_blocks are cheap (low res), spatial-only are expensive.
+        temporal_compress_level = int(np.log2(self.temporal_compression_ratio))
+        split_at = temporal_compress_level  # first N up_blocks do temporal upsample
+
+        # Phase 1: conv_in + mid_block + temporal up_blocks on full tensor (low/medium res)
+        x, _ = decoder.conv_in(z)
+        x, _ = decoder.mid_block(x, None, z)
+
+        for i in range(split_at):
+            x, _ = decoder.up_blocks[i](x, None, z)
+
+        # Phase 2: remaining spatial-only up_blocks + norm_out + conv_out in temporal chunks
+        remaining_blocks = list(range(split_at, len(decoder.up_blocks)))
+        chunk_size = 4  # pixel frames per chunk through high-res layers
+        t_expanded = x.shape[2]
+
+        if t_expanded <= chunk_size or len(remaining_blocks) == 0:
+            # Small enough to process in one go
+            for i in remaining_blocks:
+                x, _ = decoder.up_blocks[i](x, None, z)
+            x, _ = decoder.norm_out(x, z)
+            x = decoder.conv_act(x)
+            x, _ = decoder.conv_out(x)
+            return x
+
+        # Expand z temporally once to match Phase 2's time dimension.
+        # z stays at latent spatial resolution so this is small (~16 MB vs ~1.3 GB
+        # for the old approach of pre-interpolating to every pixel resolution).
+        z_time_expanded = _interpolate_zq(z, (t_expanded, z.shape[3], z.shape[4]))
+
+        # Process in temporal chunks, interpolating spatially per-chunk to avoid
+        # allocating full [B, C, t_expanded, H, W] tensors at each resolution.
+        dec_out = []
+        conv_caches = {}
+
+        for chunk_start in range(0, t_expanded, chunk_size):
+            chunk_end = min(chunk_start + chunk_size, t_expanded)
+            x_chunk = x[:, :, chunk_start:chunk_end]
+            z_t_chunk = z_time_expanded[:, :, chunk_start:chunk_end]
+            z_spatial_cache = {}
+
+            for i in remaining_blocks:
+                block = decoder.up_blocks[i]
+                cache_key = f"up_block_{i}"
+                hw_key = (x_chunk.shape[3], x_chunk.shape[4])
+                if hw_key not in z_spatial_cache:
+                    if z_t_chunk.shape[3] == hw_key[0] and z_t_chunk.shape[4] == hw_key[1]:
+                        z_spatial_cache[hw_key] = z_t_chunk
+                    else:
+                        z_spatial_cache[hw_key] = F.interpolate(z_t_chunk, size=(z_t_chunk.shape[2], hw_key[0], hw_key[1]))
+                x_chunk, new_cache = block(x_chunk, None, z_spatial_cache[hw_key], conv_cache=conv_caches.get(cache_key))
+                conv_caches[cache_key] = new_cache
+
+            hw_key = (x_chunk.shape[3], x_chunk.shape[4])
+            if hw_key not in z_spatial_cache:
+                z_spatial_cache[hw_key] = F.interpolate(z_t_chunk, size=(z_t_chunk.shape[2], hw_key[0], hw_key[1]))
+            x_chunk, new_cache = decoder.norm_out(x_chunk, z_spatial_cache[hw_key], conv_cache=conv_caches.get("norm_out"))
+            conv_caches["norm_out"] = new_cache
+            x_chunk = decoder.conv_act(x_chunk)
+            x_chunk, new_cache = decoder.conv_out(x_chunk, conv_cache=conv_caches.get("conv_out"))
+            conv_caches["conv_out"] = new_cache
+
+            dec_out.append(x_chunk.cpu())
+            del z_spatial_cache
+
+        del x, z_time_expanded
+        return torch.cat(dec_out, dim=2).to(device)