sgl-project/sglang · error · ImportError

The Helion package is required when a KDA backend is set to

Error message

The Helion package is required when a KDA backend is set to Helion. Install it with: pip install helion==1.4.0

What it means

HelionKDAKernel imports the helion JIT package when either decode or prefill is enabled; if helion is absent, __init__ raises ImportError pinning the required version (1.4.0).

Source

Thrown at python/sglang/srt/layers/attention/linear/kernels/kda_helion.py:41

    def __init__(
        self,
        triton_fallback: TritonKDAKernel | None = None,
        *,
        enable_decode: bool = True,
        enable_prefill: bool = True,
    ) -> None:
        self.supports_packed_decode = enable_decode
        self._packed_decode = None
        self._replayssm_decode = None
        self._chunk_kda = None
        if enable_decode or enable_prefill:
            try:
                import helion  # noqa: F401
            except ModuleNotFoundError as error:
                if error.name != "helion":
                    raise
                raise ImportError(
                    "The Helion package is required when a KDA backend is set to "
                    "Helion. Install it with: pip install helion==1.4.0"
                ) from None
        if enable_decode:
            from sglang.kernels.ops.attention.helion.kda_decode import (
                helion_fused_recurrent_kda_packed_decode,
            )
            from sglang.kernels.ops.attention.helion.kda_replayssm import (
                helion_fused_recurrent_kda_replayssm_decode,
            )

            self._packed_decode = helion_fused_recurrent_kda_packed_decode
            self._replayssm_decode = helion_fused_recurrent_kda_replayssm_decode
        if enable_prefill:
            from sglang.kernels.ops.attention.helion.kda_prefill import chunk_kda

            self._chunk_kda = chunk_kda
        self._triton = triton_fallback or TritonKDAKernel()

View on GitHub (pinned to 0132848349)

Solutions

1. pip install helion==1.4.0
2. Fall back to triton backend if helion cannot be installed
3. Verify the install with python -c 'import helion' before launching the server

Example fix

# before
# helion not installed
# after
pip install helion==1.4.0

try:
    import helion
    helion_ok = True
except ModuleNotFoundError:
    helion_ok = False
if not helion_ok:
    linear_attn_backend = 'triton'

null

try:
    kernel = HelionKDAKernel(enable_decode=True, enable_prefill=True)
except ImportError as e:
    logger.warning('%s; falling back to triton KDA', e)
    kernel = TritonKDAKernel()

When it happens

Trigger: Selecting helion as a KDA backend (decode and/or prefill) in an environment without the helion package installed.

Common situations: New environment or minimal Docker image without helion; version drift where helion is present but the import of the exact module fails with name 'helion'.

Related errors

• The 'flashkda' KDA prefill backend requires the flash_kda mo
• `mixed_qkv` must be a 2D tensor (got ndim={mixed_qkv.ndim}).
• `mixed_qkv` must be contiguous in the last dim.
• `a` and `b` must be 2D tensors (got a.ndim={a.ndim}, b.ndim=
• `a`/`b` must be contiguous in the last dim.

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/d0373910e10ff17b. Report an issue: GitHub.