sgl-project/sglang · critical · ImportError

Sparse Video Gen 2 attention backend requires svg package to

Error message

Sparse Video Gen 2 attention backend requires svg package to be installedPlease install it by following the instructions at https://github.com/svg-project/Sparse-VideoGen

What it means

The SVG2 backend is guarded by an availability check (svg2_available). If the optional 'svg' package (SparseVideoGen, github.com/svg-project/Sparse-VideoGen) is not installed, the constructor raises ImportError with install instructions.

Source

Thrown at python/sglang/multimodal_gen/runtime/layers/attention/backends/sparse_video_gen_2_attn.py:199

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])

    def kmeans_init(
        self,
        query: torch.Tensor,

View on GitHub (pinned to 0132848349)

Solutions

Install the svg package per https://github.com/svg-project/Sparse-VideoGen (usually pip install from the repo / provided wheel matching your CUDA+Torch version).
If install claims it's present, test `import svg` in the same Python env — a failing import (missing .so, CUDA mismatch) also sets svg2_available=False; fix the underlying build issue.
Fall back to a built-in attention backend (e.g. flashinfer/fa3) that doesn't need svg if you just need to run the model.

Example fix

# before: backend selected but svg missing -> ImportError
# after
pip install git+https://github.com/svg-project/Sparse-VideoGen.git
python -c "import svg"  # verify import works in the serving env

Defensive patterns

Strategy: try-catch

Validate before calling

import importlib.util
svg2_available = importlib.util.find_spec("svg") is not None
if not svg2_available:
    # choose fallback backend or fail with a clear message before launch
    backend = "flashinfer"

Type guard

def svg2_installed() -> bool:
    import importlib.util
    return importlib.util.find_spec("svg") is not None

Try / catch

try:
    impl = SparseVideoGen2AttentionImpl(...)
except ImportError as e:
    if "svg package" in str(e):
        log.warning("svg missing; falling back to flashinfer attention")
        impl = make_flashinfer_impl(...)
    else:
        raise

Prevention

Bake optional deps like svg into your Docker image and verify with `python -c 'import svg'` at container build time.
Reinstall/rebuild svg wheels after CUDA or PyTorch upgrades.
Make backend selection conditional on an availability probe instead of hardcoding svg2.

When it happens

Trigger: Selecting the SVG2 attention backend without having installed the svg package — e.g. --attention-backend svg2 on a fresh sglang install, or in a Docker image that only ships core dependencies.

Common situations: Fresh environments/CI images lacking optional deps; GPU/CUDA-specific svg build not compiled after a PyTorch or CUDA upgrade; the import of svg silently failing (which flips svg2_available to False) due to a missing shared library rather than the package itself.

Related errors

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/1ae8225584ce8aa4. Report an issue: GitHub.