[FEAT]Add FastVideo VSA backend for Wan2.2 (#4820)

Signed-off-by: bjf-frz <frz123db@gmail.com>
Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>

vllm_omni/diffusion/attention/backends/fastvideo_vsa.py

@@ -0,0 +1,431 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
+from __future__ import annotations
+
+import functools
+import math
+from collections.abc import Mapping
+from typing import Any
+
+import torch
+from vllm.logger import init_logger
+
+from vllm_omni.diffusion.attention.backends.abstract import (
+    AttentionBackend,
+    AttentionImpl,
+    AttentionMetadata,
+)
+from vllm_omni.diffusion.attention.backends.sdpa import SDPAImpl
+from vllm_omni.diffusion.forward_context import get_forward_context, is_forward_context_available
+
+logger = init_logger(__name__)
+
+
+# Keep the external FastVideo pybind/CUDA kernel opaque to torch.compile.
+# This mirrors the SageAttention3 backend pattern: tracing the raw extension
+# through Dynamo can reach Inductor scheduling with unstable internal op names
+# (e.g. KeyError: "op12").  The custom op gives Dynamo a single Tensor->Tensor
+# boundary and lets Inductor schedule the surrounding Wan block normally.
+if not hasattr(torch.ops.vllm_omni, "fastvideo_vsa_bshd"):
+
+    @torch.library.custom_op("vllm_omni::fastvideo_vsa_bshd", mutates_args=())
+    def _fastvideo_vsa_bshd_op(
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        variable_block_sizes: torch.Tensor,
+        q_variable_block_sizes: torch.Tensor,
+        compress_attn_weight: torch.Tensor,
+        topk: int,
+        block_t: int,
+        block_h: int,
+        block_w: int,
+    ) -> torch.Tensor:
+        from fastvideo_kernel import video_sparse_attn_bshd
+
+        return video_sparse_attn_bshd(
+            query,
+            key,
+            value,
+            variable_block_sizes=variable_block_sizes,
+            q_variable_block_sizes=q_variable_block_sizes,
+            topk=topk,
+            block_size=(block_t, block_h, block_w),
+            compress_attn_weight=compress_attn_weight if compress_attn_weight.numel() else None,
+        )
+
+    @_fastvideo_vsa_bshd_op.register_fake
+    def _(
+        query,
+        key,
+        value,
+        variable_block_sizes,
+        q_variable_block_sizes,
+        compress_attn_weight,
+        topk,
+        block_t,
+        block_h,
+        block_w,
+    ):
+        del (
+            key,
+            value,
+            variable_block_sizes,
+            q_variable_block_sizes,
+            compress_attn_weight,
+            topk,
+            block_t,
+            block_h,
+            block_w,
+        )
+        return torch.empty_like(query)
+
+
+_fastvideo_vsa_bshd_op = torch.ops.vllm_omni.fastvideo_vsa_bshd
+
+
+@functools.lru_cache(maxsize=32)
+def _get_tile_partition_indices(
+    dit_seq_shape: tuple[int, int, int],
+    tile_size: tuple[int, int, int],
+    device: torch.device,
+) -> torch.Tensor:
+    t_size, h_size, w_size = dit_seq_shape
+    tile_t, tile_h, tile_w = tile_size
+    indices = torch.arange(t_size * h_size * w_size, device=device, dtype=torch.long).reshape(t_size, h_size, w_size)
+    tiles = []
+    for tile_t_idx in range(math.ceil(t_size / tile_t)):
+        for tile_h_idx in range(math.ceil(h_size / tile_h)):
+            for tile_w_idx in range(math.ceil(w_size / tile_w)):
+                tiles.append(
+                    indices[
+                        tile_t_idx * tile_t : min((tile_t_idx + 1) * tile_t, t_size),
+                        tile_h_idx * tile_h : min((tile_h_idx + 1) * tile_h, h_size),
+                        tile_w_idx * tile_w : min((tile_w_idx + 1) * tile_w, w_size),
+                    ].flatten()
+                )
+    return torch.cat(tiles, dim=0)
+
+
+@functools.lru_cache(maxsize=32)
+def _construct_variable_block_sizes(
+    dit_seq_shape: tuple[int, int, int],
+    tile_size: tuple[int, int, int],
+    device: torch.device,
+) -> torch.Tensor:
+    num_tiles = tuple(math.ceil(seq_dim / tile_dim) for seq_dim, tile_dim in zip(dit_seq_shape, tile_size))
+
+    def _sizes(dim_len: int, tile: int, n_tiles: int) -> torch.Tensor:
+        sizes = torch.full((n_tiles,), tile, dtype=torch.int32, device=device)
+        remainder = dim_len - (n_tiles - 1) * tile
+        sizes[-1] = remainder if remainder > 0 else tile
+        return sizes
+
+    t_sizes = _sizes(dit_seq_shape[0], tile_size[0], num_tiles[0])
+    h_sizes = _sizes(dit_seq_shape[1], tile_size[1], num_tiles[1])
+    w_sizes = _sizes(dit_seq_shape[2], tile_size[2], num_tiles[2])
+    return (t_sizes[:, None, None] * h_sizes[None, :, None] * w_sizes[None, None, :]).reshape(-1)
+
+
+@functools.lru_cache(maxsize=32)
+def _get_non_pad_index(variable_block_sizes: torch.Tensor, max_block_size: int) -> torch.Tensor:
+    num_blocks = variable_block_sizes.shape[0]
+    device = variable_block_sizes.device
+    starts = torch.arange(num_blocks, device=device) * max_block_size
+    padded_index = starts[:, None] + torch.arange(max_block_size, device=device)[None, :]
+    valid = torch.arange(max_block_size, device=device)[None, :] < variable_block_sizes[:, None]
+    return padded_index[valid]
+
+
+@torch.compiler.disable
+def _get_tile_metadata(
+    dit_seq_shape: tuple[int, int, int],
+    tile_size: tuple[int, int, int],
+    block_elements: int,
+    device: torch.device,
+) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+    tile_partition_indices = _get_tile_partition_indices(dit_seq_shape, tile_size, device)
+    variable_block_sizes = _construct_variable_block_sizes(dit_seq_shape, tile_size, device)
+    non_pad_index = _get_non_pad_index(variable_block_sizes, block_elements)
+    untile_combined_index = non_pad_index[torch.argsort(tile_partition_indices)]
+    return tile_partition_indices, variable_block_sizes, non_pad_index, untile_combined_index
+
+
+def _get_vsa_dit_seq_shape(attn_metadata: AttentionMetadata | None) -> tuple[int, int, int] | None:
+    if attn_metadata is None:
+        return None
+    value = attn_metadata.extra.get("vsa_dit_seq_shape")
+    if value is None:
+        return None
+    if not isinstance(value, (list, tuple)) or len(value) != 3:
+        return None
+    return (int(value[0]), int(value[1]), int(value[2]))
+
+
+def _get_gate_compress(attn_metadata: AttentionMetadata | None) -> torch.Tensor | None:
+    if attn_metadata is None:
+        return None
+    value = attn_metadata.extra.get("gate_compress")
+    return value if isinstance(value, torch.Tensor) else None
+
+
+def _preserve_vsa_all_blocks(attn_metadata: AttentionMetadata | None) -> bool:
+    if attn_metadata is None:
+        return False
+    return attn_metadata.extra.get("preserve_vsa_all_blocks") is True
+
+
+class FastVideoVSABackend(AttentionBackend):
+    accept_output_buffer: bool = True
+
+    @staticmethod
+    def get_supported_head_sizes() -> list[int]:
+        # FastVideo VSA is intended for video DiT head sizes such as 64/128.
+        # Keep this permissive and let the runtime fallback handle unsupported
+        # cases from the installed fastvideo-kernel build.
+        return []
+
+    @staticmethod
+    def get_name() -> str:
+        return "FASTVIDEO_VSA"
+
+    @staticmethod
+    def get_impl_cls() -> type[FastVideoVSAImpl]:
+        return FastVideoVSAImpl
+
+
+class FastVideoVSAImpl(AttentionImpl):
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        softmax_scale: float,
+        causal: bool = False,
+        num_kv_heads: int | None = None,
+        prefix: str = "",
+        qkv_layout: str | None = None,
+        backend_kwargs: Mapping[str, Any] | None = None,
+        **extra_impl_args,
+    ) -> None:
+        backend_kwargs = backend_kwargs or {}
+        self.num_heads = num_heads
+        self.num_kv_heads = num_kv_heads if num_kv_heads is not None else num_heads
+        self.head_size = head_size
+        self.softmax_scale = softmax_scale
+        self.causal = causal
+        self.qkv_layout = qkv_layout
+
+        self.topk = int(backend_kwargs.get("topk", 64))
+        self.block_size = self._parse_block_size(backend_kwargs.get("block_size", (4, 8, 8)))
+        self.block_elements = self.block_size[0] * self.block_size[1] * self.block_size[2]
+        self.min_seq_len = int(backend_kwargs.get("min_seq_len", self.block_elements * 2))
+        self.fallback_on_error = bool(backend_kwargs.get("fallback_on_error", True))
+        self.disable_when_sp_active = bool(backend_kwargs.get("disable_when_sp_active", True))
+
+        self.sdpa_fallback = SDPAImpl(
+            num_heads=num_heads,
+            head_size=head_size,
+            softmax_scale=softmax_scale,
+            causal=causal,
+            num_kv_heads=num_kv_heads,
+            qkv_layout=qkv_layout,
+        )
+
+        if self.block_elements != 256:
+            logger.warning(
+                "FASTVIDEO_VSA currently uses fastvideo_kernel.video_sparse_attn_bshd, "
+                "which supports only 256-token blocks. Configured block_size=%s "
+                "(product=%d) will fall back to SDPA.",
+                self.block_size,
+                self.block_elements,
+            )
+
+    @staticmethod
+    def _parse_block_size(value: Any) -> tuple[int, int, int]:
+        if isinstance(value, int):
+            return (value, value, value)
+        if isinstance(value, (list, tuple)) and len(value) == 3:
+            return (int(value[0]), int(value[1]), int(value[2]))
+        raise ValueError(f"FASTVIDEO_VSA block_size must be an int or length-3 tuple/list, got {value!r}")
+
+    def _fallback(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        attn_metadata: AttentionMetadata | None,
+        reason: str,
+    ) -> torch.Tensor:
+        logger.warning_once("FASTVIDEO_VSA falling back to SDPA: %s", reason)
+        return self.sdpa_fallback.forward(query, key, value, attn_metadata)
+
+    def _fallback_reason(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        attn_metadata: AttentionMetadata | None,
+    ) -> str | None:
+        if self.causal:
+            return "causal attention is not supported"
+        if self.block_elements != 256:
+            return f"block_elements must be 256, got {self.block_elements}"
+        if self.topk <= 0:
+            return f"topk must be positive, got {self.topk}"
+        if query.ndim != 4 or key.ndim != 4 or value.ndim != 4:
+            return f"expected [B, S, H, D] tensors, got {query.shape}, {key.shape}, {value.shape}"
+        if query.shape[0] != key.shape[0] or query.shape[0] != value.shape[0]:
+            return "batch dimensions must match"
+        if query.shape[2:] != key.shape[2:] or query.shape[2:] != value.shape[2:]:
+            return "head/head_dim dimensions must match"
+        if query.shape[1] != key.shape[1] or query.shape[1] != value.shape[1]:
+            return "initial VSA backend supports self-attention with Sq == Skv only"
+        if query.shape[1] < self.min_seq_len:
+            return f"sequence length {query.shape[1]} is below min_seq_len {self.min_seq_len}"
+        dit_seq_shape = _get_vsa_dit_seq_shape(attn_metadata)
+        if dit_seq_shape is None:
+            return "vsa_dit_seq_shape metadata is required"
+        if math.prod(dit_seq_shape) != query.shape[1]:
+            return f"vsa_dit_seq_shape product {math.prod(dit_seq_shape)} != sequence length {query.shape[1]}"
+        num_blocks = math.prod(
+            math.ceil(seq_dim / tile_dim) for seq_dim, tile_dim in zip(dit_seq_shape, self.block_size)
+        )
+        if self.topk > num_blocks:
+            return f"topk {self.topk} > num_blocks {num_blocks}"
+        if query.dtype not in (torch.float16, torch.bfloat16):
+            return f"dtype {query.dtype} is not supported"
+        if key.dtype != query.dtype or value.dtype != query.dtype:
+            return "q/k/v dtypes must match"
+        if query.device.type != "cuda" or key.device.type != "cuda" or value.device.type != "cuda":
+            return "q/k/v must be CUDA tensors"
+        expected_scale = self.head_size**-0.5
+        if abs(float(self.softmax_scale) - float(expected_scale)) > 1e-6:
+            return f"softmax_scale {self.softmax_scale} differs from FastVideo VSA scale {expected_scale}"
+        if attn_metadata is not None and attn_metadata.attn_mask is not None:
+            return "attention masks are not supported"
+        if attn_metadata is not None and attn_metadata.full_attn_spans is not None:
+            return "piecewise/full attention spans are not supported"
+        if self.num_heads != self.num_kv_heads:
+            return "GQA/MQA is not supported"
+        if self.disable_when_sp_active and is_forward_context_available():
+            ctx = get_forward_context()
+            if getattr(ctx, "sp_active", False):
+                return "sequence parallel context is active"
+        return None
+
+    def forward_cuda(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        attn_metadata: AttentionMetadata | None = None,
+    ) -> torch.Tensor:
+        reason = self._fallback_reason(query, key, value, attn_metadata)
+        if reason is not None:
+            return self._fallback(query, key, value, attn_metadata, reason)
+
+        seq_len = query.shape[1]
+        dit_seq_shape = _get_vsa_dit_seq_shape(attn_metadata)
+        assert dit_seq_shape is not None
+        num_blocks = math.prod(
+            math.ceil(seq_dim / tile_dim) for seq_dim, tile_dim in zip(dit_seq_shape, self.block_size)
+        )
+        preserve_all_blocks = _preserve_vsa_all_blocks(attn_metadata)
+        use_native_sdpa = self.topk == num_blocks and not preserve_all_blocks
+        route = "SDPA" if use_native_sdpa else "VSA_ALL_BLOCKS" if self.topk == num_blocks else "VSA"
+        checkpoint_mode = "fastvideo_dmd" if preserve_all_blocks else "native"
+        logger.info_once(
+            "FASTVIDEO_VSA routing: seq_len=%d, dit_seq_shape=%s, block_size=%s, num_blocks=%d, "
+            "topk=%d, keep_ratio=%.1f%%, checkpoint_mode=%s, route=%s",
+            seq_len,
+            dit_seq_shape,
+            self.block_size,
+            num_blocks,
+            self.topk,
+            100.0 * self.topk / num_blocks,
+            checkpoint_mode,
+            route,
+        )
+        if use_native_sdpa:
+            return self._fallback(
+                query,
+                key,
+                value,
+                attn_metadata,
+                f"topk {self.topk} selects all blocks for a native checkpoint",
+            )
+
+        try:
+            tile_partition_indices, variable_block_sizes, non_pad_index, untile_combined_index = _get_tile_metadata(
+                dit_seq_shape,
+                self.block_size,
+                self.block_elements,
+                query.device,
+            )
+
+            padded_len = variable_block_sizes.numel() * self.block_elements
+            target_shape = (query.shape[0], padded_len, query.shape[2], query.shape[3])
+            query_tiled = torch.zeros(target_shape, device=query.device, dtype=query.dtype)
+            key_tiled = torch.zeros_like(query_tiled)
+            value_tiled = torch.zeros_like(query_tiled)
+            query_tiled[:, non_pad_index] = query[:, tile_partition_indices]
+            key_tiled[:, non_pad_index] = key[:, tile_partition_indices]
+            value_tiled[:, non_pad_index] = value[:, tile_partition_indices]
+            # Gate behavior is checkpoint-driven, not user-configured.
+            # Wan VSA layers always provide a gate projection. Its zero
+            # initialization makes checkpoints without gate weights sparse-only;
+            # checkpoints containing to_gate_compress weights use the learned gate.
+            gate_compress = _get_gate_compress(attn_metadata)
+            if gate_compress is None:
+                gate_compress = torch.zeros_like(query)
+            elif gate_compress.shape != query.shape:
+                raise ValueError(f"gate_compress shape {gate_compress.shape} must match query shape {query.shape}")
+            gate_tiled = torch.zeros_like(query_tiled)
+            gate_tiled[:, non_pad_index] = gate_compress[:, tile_partition_indices]
+            compress_attn_weight = gate_tiled
+
+            output = _fastvideo_vsa_bshd_op(
+                query_tiled.contiguous(),
+                key_tiled.contiguous(),
+                value_tiled.contiguous(),
+                variable_block_sizes,
+                variable_block_sizes,
+                compress_attn_weight,
+                self.topk,
+                self.block_size[0],
+                self.block_size[1],
+                self.block_size[2],
+            )
+            return output[:, untile_combined_index].contiguous()
+        except Exception as exc:
+            if not self.fallback_on_error:
+                raise
+            return self._fallback(query, key, value, attn_metadata, f"VSA kernel failed: {exc}")
+
+    def forward_npu(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        attn_metadata: AttentionMetadata | None = None,
+    ) -> torch.Tensor:
+        return self.sdpa_fallback.forward_npu(query, key, value, attn_metadata)
+
+    def forward_xpu(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        attn_metadata: AttentionMetadata | None = None,
+    ) -> torch.Tensor:
+        return self.sdpa_fallback.forward_xpu(query, key, value, attn_metadata)
+
+    def forward_musa(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        attn_metadata: AttentionMetadata | None = None,
+    ) -> torch.Tensor:
+        return self.sdpa_fallback.forward_musa(query, key, value, attn_metadata)

vllm_omni/diffusion/attention/backends/registry.py

@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """Diffusion attention backend registry.
 
 This module provides an enum-based registry for diffusion attention backends,
@@ -23,7 +23,7 @@ logger = init_logger(__name__)
 class _DiffusionBackendEnumMeta(EnumMeta):
     """Metaclass for DiffusionAttentionBackendEnum to provide better error messages."""
 
-    def __getitem__(cls, name: str) -> "DiffusionAttentionBackendEnum":
+    def __getitem__(cls, name: str) -> "DiffusionAttentionBackendEnum":  # type: ignore[override]
         """Get backend by name with helpful error messages."""
         try:
             return super().__getitem__(name)  # type: ignore[return-value]
@@ -62,6 +62,7 @@ class DiffusionAttentionBackendEnum(Enum, metaclass=_DiffusionBackendEnumMeta):
     SAGE_ATTN_3 = "vllm_omni.diffusion.attention.backends.sage_attn3.SageAttention3Backend"
     CUDNN_ATTN = "vllm_omni.diffusion.attention.backends.cudnn_attn.CuDNNAttentionBackend"
     FLASHINFER_ATTN = "vllm_omni.diffusion.attention.backends.flashinfer_attn.FlashInferAttentionBackend"
+    FASTVIDEO_VSA = "vllm_omni.diffusion.attention.backends.fastvideo_vsa.FastVideoVSABackend"
     FLASH_ATTN_HUB = "vllm_omni.diffusion.attention.backends.flash_attn_hub.FlashAttentionHubBackend"
     FLASH_ATTN_3_HUB = "vllm_omni.diffusion.attention.backends.flash_attn_hub.FlashAttention3HubBackend"
     TRTLLM_ATTN = "vllm_omni.diffusion.attention.backends.trtllm_attn.TrtllmAttentionBackend"

vllm_omni/diffusion/data.py

@@ -1,6 +1,6 @@
 # adapted from sglang and fastvideo
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 import copy
 import math
 import os
@@ -77,11 +77,13 @@ def normalize_omni_diffusion_kwargs(kwargs: Mapping[str, Any]) -> dict[str, Any]
     # Handle "diffusion_attention_backend" shorthand: merge into
     # diffusion_attention_config before field filtering.
     diffusion_attn_backend = normalized.pop("diffusion_attention_backend", None)
-    if diffusion_attn_backend is not None:
+    fastvideo_vsa_topk = normalized.pop("fastvideo_vsa_topk", None)
+    if diffusion_attn_backend is not None or fastvideo_vsa_topk is not None:
         existing = normalized.get("diffusion_attention_config")
         normalized["diffusion_attention_config"] = parse_attention_config(
             existing,
             attention_backend=diffusion_attn_backend,
+            fastvideo_vsa_topk=fastvideo_vsa_topk,
         )
 
     # Check environment variable as fallback for cache_backend.
@@ -104,6 +106,7 @@ def parse_kv_cache_skip_selector(
     """Parse a non-negative index selector such as "0-9,20,25-30"."""
     if selector is None:
         return None
+    values: set[int]
     if isinstance(selector, set):
         values = selector
     elif isinstance(selector, (list, tuple)):
@@ -112,7 +115,7 @@ def parse_kv_cache_skip_selector(
         text = selector.strip()
         if not text:
             return None
-        values: set[int] = set()
+        values = set()
         for chunk in text.split(","):
             token = chunk.strip()
             if not token:
@@ -245,6 +248,7 @@ class DiffusionParallelConfig:
         assert self.pipeline_parallel_size > 0, "Pipeline parallel size must be > 0"
         assert self.data_parallel_size is None or self.data_parallel_size > 0, "Data parallel size must be > 0"
         assert self.tensor_parallel_size > 0, "Tensor parallel size must be > 0"
+        assert self.sequence_parallel_size is not None
         assert self.sequence_parallel_size > 0, "Sequence parallel size must be > 0"
         assert self.ulysses_degree > 0, "Ulysses degree must be > 0"
         assert self.ring_degree > 0, "Ring degree must be > 0"
@@ -362,6 +366,7 @@ class DiffusionParallelConfig:
             self.world_size = world_size
             return 1
 
+        assert self.sequence_parallel_size is not None
         non_dp_size = (
             self.pipeline_parallel_size
             * self.tensor_parallel_size
@@ -1063,6 +1068,7 @@ class OmniDiffusionConfig:
 
         if self.diffusion_compile_granularity == "full":
             incompatible_features = []
+            assert self.parallel_config.sequence_parallel_size is not None
             if self.parallel_config.use_hsdp:
                 incompatible_features.append("HSDP")
             if self.parallel_config.sequence_parallel_size > 1:
@@ -1262,6 +1268,7 @@ class OmniDiffusionConfig:
         if self.model_class_name is None and self.diffusion_load_format == "diffusers":
             self.model_class_name = "DiffusersAdapterPipeline"
 
+        assert self.model is not None
         try:
             config_dict = get_diffusion_model_index(
                 self.model,
@@ -1338,6 +1345,7 @@ class OmniDiffusionConfig:
 
                 from vllm_omni.diffusion.utils.hf_utils import _looks_like_hidream_o1
 
+                assert self.model is not None
                 is_hidream_o1 = _looks_like_hidream_o1(self.model, cfg)
                 if self.model_class_name == "HiDreamO1ImagePipeline" and not is_hidream_o1:
                     raise ValueError(f"Checkpoint {self.model} does not have the HiDream-O1 signature")
@@ -1395,6 +1403,7 @@ class OmniDiffusionConfig:
                 elif model_type == "vla":
                     from vllm_omni.diffusion.utils.hf_utils import _looks_like_dreamzero
 
+                    assert self.model is not None
                     if _looks_like_dreamzero(self.model):
                         self.model_class_name = "DreamZeroPipeline"
                         self.set_tf_model_config(TransformerConfig())
@@ -1593,10 +1602,11 @@ class AttnQuantSpec:
         for name, v in (("dtype_qk", self.dtype_qk), ("dtype_vo", self.dtype_vo)):
             if v is not None and v not in self._VALID_DTYPES:
                 raise ValueError(f"quant.{name}={v!r} unsupported; use one of {sorted(self._VALID_DTYPES)}.")
-        for name, v in (("q_block_size", self.q_block_size), ("k_block_size", self.k_block_size)):
-            if v not in self._VALID_BLOCK_SIZES:
+        for block_name, block_value in (("q_block_size", self.q_block_size), ("k_block_size", self.k_block_size)):
+            if block_value not in self._VALID_BLOCK_SIZES:
                 raise ValueError(
-                    f"quant.{name}={v!r} unsupported; kernels exist only for {sorted(self._VALID_BLOCK_SIZES)}."
+                    f"quant.{block_name}={block_value!r} unsupported; "
+                    f"kernels exist only for {sorted(self._VALID_BLOCK_SIZES)}."
                 )
 
     @property
@@ -1638,6 +1648,7 @@ class AttentionSpec:
     backend: str
     skip_softmax: SkipSoftmaxSpec | None = None
     quant: AttnQuantSpec | None = None
+    fastvideo_vsa_topk: int | None = None
     block_sparse: BlockSparseSpec | None = None
     skip_calibration: dict | None = field(default=None, repr=False)
 
@@ -1657,6 +1668,11 @@ class AttentionSpec:
                 f"quant is only supported by the TRTLLM_ATTN and FLASHINFER_ATTN backends, but "
                 f"backend={self.backend!r}. Remove quant or set a supported backend."
             )
+        if self.fastvideo_vsa_topk is not None:
+            if self.backend.upper() != "FASTVIDEO_VSA":
+                raise ValueError("fastvideo_vsa_topk is only supported by the FASTVIDEO_VSA backend.")
+            if self.fastvideo_vsa_topk <= 0:
+                raise ValueError("fastvideo_vsa_topk must be positive.")
         if self.backend.upper() in BLOCK_SPARSE_BACKENDS:
             # Selecting the backend is the opt-in; without an explicit block the
             # defaults apply rather than silently running dense.
@@ -1698,12 +1714,15 @@ class AttentionSpec:
             if q.flashinfer_backend is not None:
                 quant_kw["flashinfer_backend"] = q.flashinfer_backend
             kw["quant"] = quant_kw
+        if self.fastvideo_vsa_topk is not None:
+            kw["topk"] = self.fastvideo_vsa_topk
         if self.block_sparse is not None:
             bs = self.block_sparse
             kw["sparsity"] = bs.sparsity
             kw["start_step"] = bs.start_step
             if bs.skip_layer_indices:
                 kw["skip_layers"] = sorted(bs.skip_layer_indices)
+
         return kw or None
 
 
@@ -1773,7 +1792,7 @@ class AttentionConfig:
             normalized[role] = node
             return
 
-        spec_keys = {"backend", "skip_softmax", "quant", "block_sparse"}
+        spec_keys = {"backend", "skip_softmax", "quant", "fastvideo_vsa_topk", "block_sparse"}
         node_dict = dict(node)
         node_keys = set(node_dict)
         if node_keys & spec_keys:
@@ -1812,6 +1831,7 @@ def parse_attention_config(
     attention_config: AttentionConfig | Mapping[str, Any] | None = None,
     *,
     attention_backend: str | None = None,
+    fastvideo_vsa_topk: int | None = None,
 ) -> AttentionConfig:
     """Pure type-conversion: coerce *attention_config* to an AttentionConfig.
 
@@ -1839,6 +1859,14 @@ def parse_attention_config(
         if attention_backend.lower() != "auto":
             normalized.default = AttentionSpec(backend=attention_backend)
 
+    if fastvideo_vsa_topk is not None:
+        if normalized.default is None:
+            raise ValueError("--fastvideo-vsa-topk requires --diffusion-attention-backend FASTVIDEO_VSA.")
+        if normalized.default.backend.upper() != "FASTVIDEO_VSA":
+            raise ValueError("--fastvideo-vsa-topk is only valid with the FASTVIDEO_VSA backend.")
+        normalized.default.fastvideo_vsa_topk = fastvideo_vsa_topk
+        normalized.default.__post_init__()
+
     return normalized
 
 

vllm_omni/diffusion/models/wan2_2/pipeline_wan2_2.py

@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 
 from __future__ import annotations
 
@@ -33,7 +33,7 @@ from vllm_omni.diffusion.models.interface import SupportsComponentDiscovery
 from vllm_omni.diffusion.models.progress_bar import ProgressBarMixin, _is_rank_zero
 from vllm_omni.diffusion.models.schedulers import FlowUniPCMultistepScheduler
 from vllm_omni.diffusion.models.wan2_2.scheduling_wan_euler import WanEulerScheduler
-from vllm_omni.diffusion.models.wan2_2.wan2_2_transformer import WanTransformer3DModel
+from vllm_omni.diffusion.models.wan2_2.wan2_2_transformer import WanSelfAttention, WanTransformer3DModel
 from vllm_omni.diffusion.postprocess import interpolate_video_tensor
 from vllm_omni.diffusion.profiler.diffusion_pipeline_profiler import DiffusionPipelineProfilerMixin
 from vllm_omni.diffusion.request import OmniDiffusionRequest
@@ -44,6 +44,8 @@ from vllm_omni.platforms import current_omni_platform
 logger = logging.getLogger(__name__)
 DEBUG_PERF = False
 WAN_SAMPLE_SOLVER_CHOICES = {"unipc", "euler"}
+FASTWAN_DMD_TIMESTEPS = (1000.0, 757.0, 522.0)
+FASTWAN_DMD_SCHEDULER_SHIFT = 8.0
 
 
 def build_wan_scheduler(sample_solver: str, flow_shift: float) -> Any:
@@ -64,6 +66,30 @@ def build_wan_scheduler(sample_solver: str, flow_shift: float) -> Any:
     )
 
 
+def load_wan_weights_with_optional_gate(model: nn.Module, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
+    """Load Wan weights and discard optional VSA gates absent from the checkpoint."""
+    gate_param_names = {name for name, _ in model.named_parameters() if ".to_gate_compress." in name}
+    has_gate_compress_weights = False
+
+    def tracked_weights():
+        nonlocal has_gate_compress_weights
+        for name, weight in weights:
+            if ".to_gate_compress." in name:
+                has_gate_compress_weights = True
+            yield name, weight
+
+    loaded_weights = AutoWeightsLoader(model).load_weights(tracked_weights())
+    setattr(model, "has_gate_compress_weights", has_gate_compress_weights)
+    if not has_gate_compress_weights:
+        for module in model.modules():
+            if isinstance(module, WanSelfAttention):
+                module.to_gate_compress = None
+
+    # Optional gate parameters are absent from public FullAttn/DMD checkpoints.
+    loaded_weights.update(gate_param_names)
+    return loaded_weights
+
+
 def resolve_wan_sample_solver(req: OmniDiffusionRequest, default: str = "unipc") -> str:
     extra_args = getattr(req.sampling_params, "extra_args", {}) or {}
     raw = extra_args.get("sample_solver", default)
@@ -314,6 +340,7 @@ class Wan22Pipeline(
 
         # Read model_index.json to detect expand_timesteps mode (for TI2V-5B)
         self.expand_timesteps = False
+        self.is_dmd = False
         self.has_transformer_2 = False
         if local_files_only:
             model_index_path = os.path.join(model, "model_index.json")
@@ -321,6 +348,7 @@ class Wan22Pipeline(
                 with open(model_index_path) as f:
                     model_index = json.load(f)
                     self.expand_timesteps = model_index.get("expand_timesteps", False)
+                    self.is_dmd = model_index.get("_class_name") == "WanDMDPipeline"
             # Check if this is a two-stage model (MoE with transformer_2)
             transformer_2_path = os.path.join(model, "transformer_2")
             self.has_transformer_2 = os.path.exists(transformer_2_path)
@@ -333,6 +361,7 @@ class Wan22Pipeline(
                 with open(model_index_path) as f:
                     model_index = json.load(f)
                     self.expand_timesteps = model_index.get("expand_timesteps", False)
+                    self.is_dmd = model_index.get("_class_name") == "WanDMDPipeline"
                     # Check transformer_2 from model_index
                     transformer_2_info = model_index.get("transformer_2", [None, None])
                     self.has_transformer_2 = transformer_2_info[0] is not None
@@ -422,6 +451,10 @@ class Wan22Pipeline(
         else:
             self.transformer_2 = None
 
+        for transformer in (self.transformer, self.transformer_2):
+            if transformer is not None:
+                transformer.preserve_vsa_all_blocks = self.is_dmd
+
         # Store the active transformer config
         if load_transformer:
             self.transformer_config = self.transformer.config
@@ -430,8 +463,14 @@ class Wan22Pipeline(
         else:
             raise RuntimeError("No transformer loaded")
 
-        self._sample_solver = "unipc"
-        self._flow_shift = od_config.flow_shift if od_config.flow_shift is not None else 5.0
+        self._sample_solver = "euler" if self.is_dmd else "unipc"
+        self._flow_shift = (
+            FASTWAN_DMD_SCHEDULER_SHIFT
+            if self.is_dmd
+            else od_config.flow_shift
+            if od_config.flow_shift is not None
+            else 5.0
+        )
         self.scheduler = build_wan_scheduler(self._sample_solver, self._flow_shift)
 
         self.vae_scale_factor_temporal = self.vae.config.scale_factor_temporal if getattr(self, "vae", None) else 4
@@ -480,6 +519,7 @@ class Wan22Pipeline(
         attention_kwargs: dict[str, Any],
         latent_condition: torch.Tensor | None = None,
         first_frame_mask: torch.Tensor | None = None,
+        generator: torch.Generator | None = None,
     ) -> torch.Tensor | AsyncLatents:
         if attention_kwargs is None:
             attention_kwargs = {}
@@ -561,7 +601,20 @@ class Wan22Pipeline(
                     cfg_normalize=False,
                 )
 
-                latents = self.scheduler_step_maybe_with_cfg(noise_pred, t, latents, do_true_cfg)
+                if self.is_dmd:
+                    pred_clean = self.scheduler.predict_clean(noise_pred, latents, t).to(noise_pred.dtype)
+                    if step_idx + 1 < len(timesteps):
+                        noise = randn_tensor(
+                            latents.shape,
+                            generator=generator,
+                            device=latents.device,
+                            dtype=pred_clean.dtype,
+                        )
+                        latents = self.scheduler.add_noise(pred_clean, noise, timesteps[step_idx + 1])
+                    else:
+                        latents = pred_clean
+                else:
+                    latents = self.scheduler_step_maybe_with_cfg(noise_pred, t, latents, do_true_cfg)
                 pbar.update()
 
         return latents
@@ -600,13 +653,21 @@ class Wan22Pipeline(
         mod_value = self.vae_scale_factor_spatial * patch_size[1]  # 16*2=32 for TI2V, 8*2=16 for I2V
         height = (height // mod_value) * mod_value
         width = (width // mod_value) * mod_value
-        num_steps = 40 if common.num_inference_steps is None else common.num_inference_steps
+        if self.is_dmd:
+            # The checkpoint was distilled for these three transitions. Ignore
+            # request-level step counts, including the engine's 1-step warmup.
+            num_steps = len(FASTWAN_DMD_TIMESTEPS)
+        else:
+            num_steps = 40 if common.num_inference_steps is None else common.num_inference_steps
 
         output_type = common.output_type or "np"
         num_outputs_per_prompt = common.num_outputs_per_prompt or 1
         attention_kwargs: dict | None = None
 
-        guidance_low, guidance_high = resolve_wan_guidance_scales(common, default_guidance_scale=4.0)
+        guidance_low, guidance_high = resolve_wan_guidance_scales(
+            common,
+            default_guidance_scale=1.0 if self.is_dmd else 4.0,
+        )
 
         # record guidance for properties
         self._guidance_scale = guidance_low
@@ -685,17 +746,19 @@ class Wan22Pipeline(
             current_omni_platform.synchronize()
             _t_text_enc_ms = (time.perf_counter() - _t_text_enc_start) * 1000
 
-        first_request = req.requests[0]
-        sample_solver = resolve_wan_sample_solver(first_request, default=self._sample_solver)
-        flow_shift = resolve_wan_flow_shift(first_request, self.od_config)
-        if sample_solver != self._sample_solver or abs(flow_shift - self._flow_shift) > 1e-6:
-            self.scheduler = build_wan_scheduler(sample_solver, flow_shift)
-            self._sample_solver = sample_solver
-            self._flow_shift = flow_shift
+        if self.is_dmd:
+            timesteps = torch.tensor(FASTWAN_DMD_TIMESTEPS, device=device, dtype=torch.float32)
+        else:
+            first_request = req.requests[0]
+            sample_solver = resolve_wan_sample_solver(first_request, default=self._sample_solver)
+            flow_shift = resolve_wan_flow_shift(first_request, self.od_config)
+            if sample_solver != self._sample_solver or abs(flow_shift - self._flow_shift) > 1e-6:
+                self.scheduler = build_wan_scheduler(sample_solver, flow_shift)
+                self._sample_solver = sample_solver
+                self._flow_shift = flow_shift
 
-        # Timesteps
-        self.scheduler.set_timesteps(num_steps, device=device)
-        timesteps = self.scheduler.timesteps
+            self.scheduler.set_timesteps(num_steps, device=device)
+            timesteps = self.scheduler.timesteps
         self._num_timesteps = len(timesteps)
         boundary_timestep = None
         if boundary_ratio is not None:
@@ -817,6 +880,7 @@ class Wan22Pipeline(
             attention_kwargs=attention_kwargs,
             latent_condition=latent_condition,
             first_frame_mask=first_frame_mask,
+            generator=generator,
         )
 
         # Wan2.2 is prone to out of memory errors when predicting large videos
@@ -1003,9 +1067,7 @@ class Wan22Pipeline(
         return latents
 
     def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
-        """Load weights using AutoWeightsLoader for vLLM integration."""
-        loader = AutoWeightsLoader(self)
-        return loader.load_weights(weights)
+        return load_wan_weights_with_optional_gate(self, weights)
 
     def check_inputs(
         self,

vllm_omni/diffusion/models/wan2_2/pipeline_wan2_2_i2v.py

@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 
 from __future__ import annotations
 
@@ -16,7 +16,6 @@ import torchvision.transforms.functional as TF
 from diffusers.utils.torch_utils import randn_tensor
 from torch import nn
 from transformers import AutoTokenizer, CLIPImageProcessor, CLIPVisionModel, UMT5EncoderModel
-from vllm.model_executor.models.utils import AutoWeightsLoader
 from vllm.sequence import IntermediateTensors
 
 from vllm_omni.diffusion.data import DiffusionOutput, OmniDiffusionConfig
@@ -36,6 +35,7 @@ from vllm_omni.diffusion.models.wan2_2.pipeline_wan2_2 import (
     build_wan_scheduler,
     create_transformer_from_config,
     load_transformer_config,
+    load_wan_weights_with_optional_gate,
     resolve_wan_flow_shift,
     resolve_wan_guidance_scales,
     resolve_wan_sample_solver,
@@ -1014,9 +1014,7 @@ class Wan22I2VPipeline(
             raise ValueError("`guidance_scale_2` is only supported when `boundary_ratio` is set.")
 
     def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
-        """Load weights using AutoWeightsLoader for vLLM integration."""
-        loader = AutoWeightsLoader(self)
-        return loader.load_weights(weights)
+        return load_wan_weights_with_optional_gate(self, weights)
 
 
 # ---------------------------------------------------------------------------

vllm_omni/diffusion/models/wan2_2/pipeline_wan2_2_s2v.py

@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 
 """
 Wan2.2 Speech-to-Video (S2V) Pipeline for vLLM-Omni.
@@ -22,7 +22,6 @@ from diffusers.utils.torch_utils import randn_tensor
 from torch import nn
 from torchvision import transforms
 from transformers import AutoTokenizer, UMT5Config, UMT5EncoderModel, Wav2Vec2ForCTC, Wav2Vec2Processor
-from vllm.model_executor.models.utils import AutoWeightsLoader
 from vllm.multimodal.media.audio import load_audio
 
 from vllm_omni.diffusion.data import DiffusionOutput, OmniDiffusionConfig
@@ -36,6 +35,7 @@ from vllm_omni.diffusion.models.progress_bar import ProgressBarMixin
 from vllm_omni.diffusion.models.schedulers import FlowUniPCMultistepScheduler
 from vllm_omni.diffusion.models.wan2_2.pipeline_wan2_2 import (
     load_transformer_config,
+    load_wan_weights_with_optional_gate,
     retrieve_latents,
 )
 from vllm_omni.diffusion.models.wan2_2.wan2_2_s2v_transformer import (
@@ -1475,6 +1475,4 @@ class Wan22S2VPipeline(
         return outputs
 
     def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
-        """Load weights using AutoWeightsLoader for vLLM integration."""
-        loader = AutoWeightsLoader(self)
-        return loader.load_weights(weights)
+        return load_wan_weights_with_optional_gate(self, weights)

vllm_omni/diffusion/models/wan2_2/pipeline_wan2_2_vace.py

@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 
 """
 VACE (Video Creation and Editing) Pipeline for WAN models.
@@ -22,12 +22,12 @@ import PIL.Image
 import torch
 from vllm.logger import init_logger
 from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
-from vllm.model_executor.models.utils import AutoWeightsLoader
 
 from vllm_omni.diffusion.data import DiffusionOutput, OmniDiffusionConfig
 from vllm_omni.diffusion.models.interface import SupportImageInput
 from vllm_omni.diffusion.models.wan2_2.pipeline_wan2_2 import (
     Wan22Pipeline,
+    load_wan_weights_with_optional_gate,
     retrieve_latents,
 )
 from vllm_omni.diffusion.models.wan2_2.pipeline_wan2_2 import (
@@ -747,6 +747,4 @@ class Wan22VACEPipeline(Wan22Pipeline, SupportImageInput):
         )
 
     def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
-        """Load weights using AutoWeightsLoader for vLLM integration."""
-        loader = AutoWeightsLoader(self)
-        return loader.load_weights(weights)
+        return load_wan_weights_with_optional_gate(self, weights)

vllm_omni/diffusion/models/wan2_2/scheduling_wan_euler.py

@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 
 from __future__ import annotations
 
@@ -110,6 +110,42 @@ class WanEulerScheduler:
         self._step_index = None
         self._begin_index = None
 
+    def sigma_for_timestep(
+        self,
+        timestep: float | torch.Tensor,
+        *,
+        device: torch.device | str | None = None,
+        dtype: torch.dtype = torch.float32,
+    ) -> torch.Tensor:
+        """Return the flow sigma nearest to a (possibly distilled) timestep."""
+        timestep_t = torch.as_tensor(timestep, device=self.timesteps.device, dtype=self.timesteps.dtype)
+        if timestep_t.numel() != 1:
+            raise ValueError(f"Expected a scalar timestep, got shape {tuple(timestep_t.shape)}")
+        index = torch.argmin(torch.abs(self.timesteps - timestep_t))
+        return self.sigmas[index].to(device=device or self.timesteps.device, dtype=dtype)
+
+    def predict_clean(
+        self,
+        model_output: torch.Tensor,
+        sample: torch.Tensor,
+        timestep: float | torch.Tensor,
+    ) -> torch.Tensor:
+        """Convert a flow/noise prediction into the corresponding clean latent."""
+        sample_fp32 = sample.to(torch.float32)
+        sigma = self.sigma_for_timestep(timestep, device=sample.device, dtype=torch.float32)
+        return sample_fp32 - sigma * model_output.to(torch.float32)
+
+    def add_noise(
+        self,
+        clean_sample: torch.Tensor,
+        noise: torch.Tensor,
+        timestep: float | torch.Tensor,
+    ) -> torch.Tensor:
+        """Re-noise a clean latent at a distilled flow timestep."""
+        sigma = self.sigma_for_timestep(timestep, device=noise.device, dtype=torch.float32)
+        noisy = (1.0 - sigma) * clean_sample.to(torch.float32) + sigma * noise.to(torch.float32)
+        return noisy.to(noise.dtype)
+
     def scale_model_input(self, sample: torch.Tensor, timestep: int | None = None) -> torch.Tensor:  # noqa: ARG002
         return sample
 

vllm_omni/diffusion/models/wan2_2/wan2_2_transformer.py

@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 
 import math
 from collections.abc import Iterable
@@ -419,6 +419,24 @@ class WanSelfAttention(nn.Module):
             prefix=prefix,
         )
 
+        # FastVideo VSA checkpoints may add a learned projection that gates
+        # the compressed global branch. Zero initialization makes checkpoints
+        # without these weights sparse-only, with no user-facing mode switch.
+        self.to_gate_compress: ColumnParallelLinear | None = None
+        if self.attn.attn_backend.get_name() == "FASTVIDEO_VSA":
+            self.to_gate_compress = ColumnParallelLinear(
+                dim,
+                self.inner_dim,
+                bias=True,
+                gather_output=False,
+                return_bias=False,
+                quant_config=quant_config,
+                prefix=f"{prefix}.to_gate_compress" if prefix else "to_gate_compress",
+            )
+            nn.init.zeros_(self.to_gate_compress.weight)
+            if self.to_gate_compress.bias is not None:
+                nn.init.zeros_(self.to_gate_compress.bias)
+
     def forward(
         self,
         hidden_states: torch.Tensor,
@@ -441,6 +459,15 @@ class WanSelfAttention(nn.Module):
         key = key.unflatten(2, (self.num_kv_heads, self.head_dim))
         value = value.unflatten(2, (self.num_kv_heads, self.head_dim))
 
+        if self.to_gate_compress is not None:
+            gate_result = self.to_gate_compress(hidden_states)
+            gate_compress = gate_result[0] if isinstance(gate_result, tuple) else gate_result
+            gate_compress = gate_compress.unflatten(2, (self.num_heads, self.head_dim))
+            if attn_metadata is None:
+                attn_metadata = AttentionMetadata(extra={"gate_compress": gate_compress})
+            else:
+                attn_metadata.extra["gate_compress"] = gate_compress
+
         # Apply rotary embeddings
         if rotary_emb is not None:
             freqs_cos, freqs_sin = rotary_emb
@@ -709,6 +736,8 @@ class WanTransformerBlock(nn.Module):
         temb: torch.Tensor,
         rotary_emb: tuple[torch.Tensor, torch.Tensor],
         hidden_states_mask: torch.Tensor | None = None,
+        vsa_dit_seq_shape: tuple[int, int, int] | None = None,
+        preserve_vsa_all_blocks: bool = False,
     ) -> torch.Tensor:
         if temb.ndim == 4:
             # temb: batch_size, seq_len, 6, inner_dim (wan2.2 ti2v)
@@ -729,7 +758,12 @@ class WanTransformerBlock(nn.Module):
 
         # 1. Self-attention
         norm_hidden_states = self.norm1(hidden_states, scale_msa, shift_msa).type_as(hidden_states)
-        self_attn_metadata = AttentionMetadata(attn_mask=hidden_states_mask)
+        self_attn_extra = {}
+        if vsa_dit_seq_shape is not None:
+            self_attn_extra["vsa_dit_seq_shape"] = vsa_dit_seq_shape
+        if preserve_vsa_all_blocks:
+            self_attn_extra["preserve_vsa_all_blocks"] = True
+        self_attn_metadata = AttentionMetadata(attn_mask=hidden_states_mask, extra=self_attn_extra)
         attn_output = self.attn1(norm_hidden_states, rotary_emb, self_attn_metadata)
         hidden_states = (hidden_states + attn_output * gate_msa).type_as(hidden_states)
 
@@ -910,6 +944,9 @@ class WanTransformer3DModel(nn.Module):
             pos_embed_seq_len=pos_embed_seq_len,
         )
 
+        # DMD/FastVideo checkpoints retain VSA semantics when top-k selects every block.
+        self.preserve_vsa_all_blocks = False
+
         # 3. Transformer blocks — partitioned across PP stages via vLLM's `make_layers`.
         # It computes the [start_layer, end_layer) slice for this rank and fills the remaining slots
         # with PPMissingLayer so that weight names stay globally consistent.
@@ -1044,8 +1081,19 @@ class WanTransformer3DModel(nn.Module):
             )
 
         # Transformer blocks
+        # Preserve the post-patch (T, H, W) grid so VSA can partition
+        # the flattened DiT sequence into spatiotemporal blocks.
+        vsa_dit_seq_shape = (post_patch_num_frames, post_patch_height, post_patch_width)
         for block in self.blocks[self.start_layer : self.end_layer]:
-            hidden_states = block(hidden_states, encoder_hidden_states, timestep_proj, rotary_emb, hidden_states_mask)
+            hidden_states = block(
+                hidden_states,
+                encoder_hidden_states,
+                timestep_proj,
+                rotary_emb,
+                hidden_states_mask,
+                vsa_dit_seq_shape,
+                self.preserve_vsa_all_blocks,
+            )
 
         if not is_pipeline_last_stage():
             # Non-last PP stage: hand the token sequence to the caller via IntermediateTensors.

vllm_omni/diffusion/models/wan2_2/wan2_2_vace_transformer.py

@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 
 """VACE variant of WanTransformer3DModel for conditional video generation."""
 
@@ -61,6 +61,7 @@ class VaceWanTransformerBlock(WanTransformerBlock):
         temb: torch.Tensor,
         rotary_emb: tuple[torch.Tensor, torch.Tensor],
         hidden_states_mask: torch.Tensor | None = None,
+        vsa_dit_seq_shape: tuple[int, int, int] | None = None,
     ) -> tuple[torch.Tensor, torch.Tensor]:
         if self.proj_in is not None:
             control_hidden_states = self.proj_in(control_hidden_states)
@@ -72,6 +73,7 @@ class VaceWanTransformerBlock(WanTransformerBlock):
             temb,
             rotary_emb,
             hidden_states_mask,
+            vsa_dit_seq_shape,
         )
 
         conditioning_states = self.proj_out(control_hidden_states)
@@ -250,6 +252,7 @@ class WanVACETransformer3DModel(WanTransformer3DModel):
             )
 
         # VACE: embed context and run conditioning blocks
+        vsa_dit_seq_shape = (post_patch_num_frames, post_patch_height, post_patch_width)
         vace_hints = None
         if vace_context is not None and self.vace_blocks is not None:
             full_seq_len = hidden_states.shape[1] * sp_size
@@ -263,6 +266,7 @@ class WanVACETransformer3DModel(WanTransformer3DModel):
                     timestep_proj,
                     rotary_emb,
                     hidden_states_mask,
+                    vsa_dit_seq_shape,
                 )
                 vace_hints.append(conditioning_states)
 
@@ -278,6 +282,7 @@ class WanVACETransformer3DModel(WanTransformer3DModel):
                 timestep_proj,
                 rotary_emb,
                 hidden_states_mask,
+                vsa_dit_seq_shape,
             )
             if vace_hints is not None and self.vace_layers_mapping is not None and i in self.vace_layers_mapping:
                 vace_idx = self.vace_layers_mapping[i]

vllm_omni/diffusion/registry.py

@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 
 import importlib
 
@@ -61,6 +61,11 @@ _DIFFUSION_MODELS = {
         "pipeline_wan2_2",
         "Wan22Pipeline",
     ),
+    "WanDMDPipeline": (
+        "wan2_2",
+        "pipeline_wan2_2",
+        "Wan22Pipeline",
+    ),
     "WanVACEPipeline": (
         "wan2_2",
         "pipeline_wan2_2_vace",
@@ -459,6 +464,7 @@ def _apply_sequence_parallel_if_enabled(model, od_config: OmniDiffusionConfig) -
 
     try:
         sp_size = od_config.parallel_config.sequence_parallel_size
+        assert sp_size is not None
         if sp_size <= 1:
             return
 
@@ -539,6 +545,7 @@ _DIFFUSION_POST_PROCESS_FUNCS = {
     "OvisImagePipeline": "get_ovis_image_post_process_func",
     "BooguImagePipeline": "get_boogu_image_post_process_func",
     "WanPipeline": "get_wan22_post_process_func",
+    "WanDMDPipeline": "get_wan22_post_process_func",
     "WanVACEPipeline": "get_wan22_vace_post_process_func",
     "LTX2Pipeline": "get_ltx2_post_process_func",
     "LTX2TwoStagePipeline": "get_ltx2_post_process_func",
@@ -612,6 +619,7 @@ _DIFFUSION_PRE_PROCESS_FUNCS = {
     "LongCatVideoAvatarPipeline": "get_longcat_video_avatar_pre_process_func",
     "QwenImageLayeredPipeline": "get_qwen_image_layered_pre_process_func",
     "WanPipeline": "get_wan22_pre_process_func",
+    "WanDMDPipeline": "get_wan22_pre_process_func",
     "WanVACEPipeline": "get_wan22_vace_pre_process_func",
     "WanImageToVideoPipeline": "get_wan22_i2v_pre_process_func",
     "WanS2VPipeline": "get_wan22_s2v_pre_process_func",
@@ -703,6 +711,7 @@ def register_diffusion_model(
 
 def _load_process_func(od_config: OmniDiffusionConfig, func_name: str):
     """Load and return a process function from the appropriate module."""
+    assert od_config.model_class_name is not None
     mod_folder, mod_relname, _ = _DIFFUSION_MODELS[od_config.model_class_name]
     if mod_relname == "":
         # Full module path (registered via register_diffusion_model)