feat: Auto-regressive video generation (CORE-25) (#13082)

Author: Talmaj

comfy/k_diffusion/sampling.py

@@ -1810,3 +1810,102 @@ def sample_sa_solver(model, x, sigmas, extra_args=None, callback=None, disable=F
 def sample_sa_solver_pece(model, x, sigmas, extra_args=None, callback=None, disable=False, tau_func=None, s_noise=1.0, noise_sampler=None, predictor_order=3, corrector_order=4, simple_order_2=False):
     """Stochastic Adams Solver with PECE (Predict–Evaluate–Correct–Evaluate) mode (NeurIPS 2023)."""
     return sample_sa_solver(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, tau_func=tau_func, s_noise=s_noise, noise_sampler=noise_sampler, predictor_order=predictor_order, corrector_order=corrector_order, use_pece=True, simple_order_2=simple_order_2)
+
+
+@torch.no_grad()
+def sample_ar_video(model, x, sigmas, extra_args=None, callback=None, disable=None,
+                    num_frame_per_block=1):
+    """
+    Autoregressive video sampler: block-by-block denoising with KV cache
+    and flow-match re-noising for Causal Forcing / Self-Forcing models.
+
+    Requires a Causal-WAN compatible model (diffusion_model must expose
+    init_kv_caches / init_crossattn_caches) and 5-D latents [B,C,T,H,W].
+
+    All AR-loop parameters are passed via the SamplerARVideo node, not read
+    from the checkpoint or transformer_options.
+    """
+    extra_args = {} if extra_args is None else extra_args
+    model_options = extra_args.get("model_options", {})
+    transformer_options = model_options.get("transformer_options", {})
+
+    if x.ndim != 5:
+        raise ValueError(
+            f"ar_video sampler requires 5-D video latents [B,C,T,H,W], got {x.ndim}-D tensor with shape {x.shape}. "
+            "This sampler is only compatible with autoregressive video models (e.g. Causal-WAN)."
+        )
+
+    inner_model = model.inner_model.inner_model
+    causal_model = inner_model.diffusion_model
+
+    if not (hasattr(causal_model, "init_kv_caches") and hasattr(causal_model, "init_crossattn_caches")):
+        raise TypeError(
+            "ar_video sampler requires a Causal-WAN compatible model whose diffusion_model "
+            "exposes init_kv_caches() and init_crossattn_caches(). The loaded checkpoint "
+            "does not support this interface — choose a different sampler."
+        )
+
+    seed = extra_args.get("seed", 0)
+
+    bs, c, lat_t, lat_h, lat_w = x.shape
+    frame_seq_len = -(-lat_h // 2) * -(-lat_w // 2) # ceiling division
+    num_blocks = -(-lat_t // num_frame_per_block)   # ceiling division
+    device = x.device
+    model_dtype = inner_model.get_dtype()
+
+    kv_caches = causal_model.init_kv_caches(bs, lat_t * frame_seq_len, device, model_dtype)
+    crossattn_caches = causal_model.init_crossattn_caches(bs, device, model_dtype)
+
+    output = torch.zeros_like(x)
+    s_in = x.new_ones([x.shape[0]])
+    current_start_frame = 0
+    num_sigma_steps = len(sigmas) - 1
+    total_real_steps = num_blocks * num_sigma_steps
+    step_count = 0
+
+    try:
+        for block_idx in trange(num_blocks, disable=disable):
+            bf = min(num_frame_per_block, lat_t - current_start_frame)
+            fs, fe = current_start_frame, current_start_frame + bf
+            noisy_input = x[:, :, fs:fe]
+
+            ar_state = {
+                "start_frame": current_start_frame,
+                "kv_caches": kv_caches,
+                "crossattn_caches": crossattn_caches,
+            }
+            transformer_options["ar_state"] = ar_state
+
+            for i in range(num_sigma_steps):
+                denoised = model(noisy_input, sigmas[i] * s_in, **extra_args)
+
+                if callback is not None:
+                    scaled_i = step_count * num_sigma_steps // total_real_steps
+                    callback({"x": noisy_input, "i": scaled_i, "sigma": sigmas[i],
+                              "sigma_hat": sigmas[i], "denoised": denoised})
+
+                if sigmas[i + 1] == 0:
+                    noisy_input = denoised
+                else:
+                    sigma_next = sigmas[i + 1]
+                    torch.manual_seed(seed + block_idx * 1000 + i)
+                    fresh_noise = torch.randn_like(denoised)
+                    noisy_input = (1.0 - sigma_next) * denoised + sigma_next * fresh_noise
+
+                    for cache in kv_caches:
+                        cache["end"] -= bf * frame_seq_len
+
+                step_count += 1
+
+            output[:, :, fs:fe] = noisy_input
+
+            for cache in kv_caches:
+                cache["end"] -= bf * frame_seq_len
+            zero_sigma = sigmas.new_zeros([1])
+            _ = model(noisy_input, zero_sigma * s_in, **extra_args)
+
+            current_start_frame += bf
+    finally:
+        transformer_options.pop("ar_state", None)
+
+    return output

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)