sgl-project/sglang · error · ValueError

st attn not supported

Error message

st attn not supported

What it means

SlidingTileAttentionBackend.__init__ raises ValueError when st_attn_backend_available is False, i.e. the sliding-tile-attention CUDA extension failed to import. The backend cannot run at all without the native kernel.

Source

Thrown at python/sglang/multimodal_gen/runtime/layers/attention/backends/sliding_tile_attn.py:120

        return SlidingTileAttentionMetadata(
            current_timestep=current_timestep, STA_param=param[current_timestep]
        )


class SlidingTileAttentionImpl(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 not st_attn_backend_available:
            raise ValueError("st attn not supported")
        # TODO(will-refactor): for now this is the mask strategy, but maybe we should
        # have a more general config for STA?
        mask_strategy_file_path = (
            get_global_server_args().attention_backend_config.mask_strategy_file_path
        )
        if mask_strategy_file_path is None:
            raise ValueError("SGLANG_DIFFUSION_ATTENTION_CONFIG is not set")

        # TODO(kevin): get mask strategy for different STA modes
        with open(mask_strategy_file_path) as f:
            mask_strategy = json.load(f)
        self.mask_strategy = dict_to_3d_list(mask_strategy)

        self.prefix = prefix
        sp_group = get_sp_group()
        self.sp_size = sp_group.world_size
        # STA config
        self.STA_base_tile_size = [6, 8, 8]

View on GitHub (pinned to 0132848349)

Solutions

1. Install/build the sliding-tile-attention extension for your GPU/CUDA
2. Check the import-time warning to see why the extension failed to load
3. Switch to another attention backend if STA is optional for your model

Example fix

# before
python launch.py --attention-backend sliding_tile   # ValueError: st attn not supported
# after
pip install sliding-tile-attention
python launch.py --attention-backend sliding_tile

from sglang.multimodal_gen.runtime.layers.attention.backends.sliding_tile_attn import st_attn_backend_available
if not st_attn_backend_available:
    raise SystemExit("install sliding-tile-attention to use the STA backend")

def sta_available() -> bool:
    try:
        import st_attn  # noqa
        return True
    except ImportError:
        return False

try:
    backend = SlidingTileAttentionBackend(...)
except ValueError as e:
    if "st attn not supported" in str(e):
        backend = FlashAttentionBackend(...)

When it happens

Trigger: Constructing SlidingTileAttentionBackend in an environment where the st_attn native package is not installed or failed to load.

Common situations: Using the STA backend without building/installing the sliding-tile-attention kernel; GPU arch unsupported by the extension; container missing the compiled wheel.

Related errors

• SGLANG_DIFFUSION_ATTENTION_CONFIG is not set
• Can not import FA3 in sgl_kernel. Please check your installa
• Cannot find NVIDIA Math-DX (cuBLASDx) headers. Install the `
• Cannot find CUTLASS headers required for JIT compilation. Pl
• FlashAttention-4 CUTE is not available. Install flash-attn-4

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/a1939354da98d94e. Report an issue: GitHub.