sgl-project/sglang · error · ValueError

f"Invalid prefix for SparseVideoGen2AttentionImpl: {prefix}"

Error message

f"Invalid prefix for SparseVideoGen2AttentionImpl: {prefix}"

What it means

SparseVideoGen2AttentionImpl derives its layer index from the prefix string by taking the third-from-last dot-separated component (parts[-3]) and parsing it as an int. If the prefix has fewer than 3 parts (e.g. 'layers.5' or 'blocks'), it raises this ValueError.

Source

Thrown at python/sglang/multimodal_gen/runtime/layers/attention/backends/sparse_video_gen_2_attn.py:210

        **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])

    def kmeans_init(
        self,
        query: torch.Tensor,
        key: torch.Tensor,
        attn_metadata: SparseVideoGen2AttentionMetadata,
    ):
        cfg, num_heads, seq_len, dim = query.size()
        qlabels, qcentroids, qcluster_sizes, qiter = batch_kmeans_Euclid(
            query.reshape(cfg * num_heads, seq_len, dim),
            n_clusters=attn_metadata.num_q_centroids,
            max_iters=attn_metadata.kmeans_iter_init,
        )
        klabels, kcentroids, kcluster_sizes, kiter = batch_kmeans_Euclid(
            key.reshape(cfg * num_heads, seq_len, dim),

View on GitHub (pinned to 0132848349)

Solutions

1. Pass the full hierarchical prefix containing at least 3 dot-separated components where parts[-3] is the numeric layer index, e.g. 'backbone.layers.12.attn'.
2. If you renamed modules, keep a numeric layer component third from the end, or subclass and override _get_layer_idx for custom mapping.
3. Regenerate prefixes from the model's named_modules() rather than hand-writing them.

Example fix

# before
impl = SparseVideoGen2AttentionImpl(prefix="attn", ...)  # too shallow -> ValueError

# after
impl = SparseVideoGen2AttentionImpl(prefix="model.layers.12.attn", ...)  # parts[-3] == '12'

def is_valid_svg2_prefix(prefix: str) -> bool:
    parts = prefix.split(".")
    return len(parts) >= 3 and parts[-3].isdigit()

assert is_valid_svg2_prefix(prefix), f"bad prefix: {prefix}"

def is_valid_svg2_prefix(prefix: str) -> bool:
    parts = prefix.split(".")
    return len(parts) >= 3 and parts[-3].isdigit()

When it happens

Trigger: Constructing the impl with a shallow prefix like 'blocks.3.q_proj', 'model.layers', or a custom name without the expected '...{layer_idx}.{module}.{child}' depth, so len(prefix.split('.')) < 3.

Common situations: Loading per-layer weights under renamed/remapped module paths (e.g. after checkpoint conversion or prefix stripping), or passing a user-defined prefix when manually instantiating the impl instead of letting the model builder generate hierarchical prefixes like 'transformer.blocks.12.attn'.

Related errors

• v_cache must be provided
• q can only be None when only_qv=True
• q must be provided unless qv is provided with only_qv=True
• `dt_bias` must have {HV * K} elements (got {dt_bias.numel()}
• The batch size is expected to be 1 rather than {q.shape[0]}

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/24c90e3d6767f007. Report an issue: GitHub.