sgl-project/sglang · error · ValueError

Sparse Video Gen 2 attention does not support causal attenti

Error message

Sparse Video Gen 2 attention does not support causal attention

What it means

SparseVideoGen2AttentionImpl is a bidirectional (non-causal) attention used in video DiT backbones. Its constructor explicitly rejects causal=True because the SVG2 kernel/metadata only supports full (non-causal) attention over tokens.

Source

Thrown at python/sglang/multimodal_gen/runtime/layers/attention/backends/sparse_video_gen_2_attn.py:195

            frame_size=frame_size,
            cache=cache,
        )


class SparseVideoGen2AttentionImpl(AttentionImpl):

    def __init__(
        self,
        num_heads: int,
        head_size: int,
        causal: bool,
        softmax_scale: float,
        num_kv_heads: int | None = None,
        prefix: str = "",
        **extra_impl_args,
    ) -> None:
        if causal:
            raise ValueError(
                "Sparse Video Gen 2 attention does not support causal attention"
            )
        if not svg2_available:
            raise ImportError(
                "Sparse Video Gen 2 attention backend requires svg package to be installed"
                "Please install it by following the instructions at "
                "https://github.com/svg-project/Sparse-VideoGen"
            )
        self.prefix = prefix
        self.layer_idx = self._get_layer_idx(prefix)

    def _get_layer_idx(self, prefix: str) -> int:
        parts = prefix.split(".")
        if len(parts) < 3:
            raise ValueError(
                f"Invalid prefix for SparseVideoGen2AttentionImpl: {prefix}"
            )
        return int(parts[-3])

View on GitHub (pinned to 0132848349)

Solutions

Pass causal=False (or omit the causal flag) when instantiating this backend — video DiT blocks are bidirectional.
In a shared factory, only select the SVG2 backend for non-causal layers and route causal layers to a flash/fa3-style backend.
Audit the layer config that forwards the causal flag and correct it for the video backbone.

Example fix

# before
impl = SparseVideoGen2AttentionImpl(..., causal=True, ...)

# after
impl = SparseVideoGen2AttentionImpl(..., causal=False, ...)

Defensive patterns

Strategy: validation

Validate before calling

assert not causal, "SVG2 attention is bidirectional only"
impl = SparseVideoGen2AttentionImpl(..., causal=False, ...)

Prevention

In attention factories, route causal layers away from SVG2 explicitly.
Keep video-DiT layer configs separate from text LM configs.
Add a unit test asserting the backend is only created with causal=False.

When it happens

Trigger: Constructing SparseVideoGen2AttentionImpl with causal=True — typically because a generic attention factory passes the model's causal flag through to every backend, or a config copied from a causal LM/text model.

Common situations: Wiring the SVG2 backend into a shared attention-creation path that defaults causal=True for most layers; using a text-model attention config template for a video DiT; misunderstanding that DiT double/single-stream blocks use bidirectional attention.

Related errors

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/86c4f4d05d229bff. Report an issue: GitHub.