sgl-project/sglang · error · ValueError

The hpc_ops MoE runner backend runs a plain SiLU-and-mul; it

Error message

The hpc_ops MoE runner backend runs a plain SiLU-and-mul; it does not support gemm1_alpha / gemm1_clamp_limit / swiglu_limit.

What it means

The hpc_ops backend applies a plain SiLU-and-mul activation between the two GEMMs and cannot apply gemm1_alpha, gemm1_clamp_limit, or swiglu_limit (parameters used by e.g. gpt-oss style swiglu with limits). The config check rejects any of these being non-None.

Source

Thrown at python/sglang/srt/layers/moe/moe_runner/hpc_ops.py:121

        raise ValueError(
            "The hpc_ops MoE runner backend does not support fused shared experts."
        )
    if runner_config.apply_router_weight_on_input:
        raise ValueError(
            "The hpc_ops MoE runner backend does not support "
            "apply_router_weight_on_input."
        )
    if runner_config.no_combine:
        raise ValueError(
            "The hpc_ops MoE runner backend does not support no_combine "
            "(the fused kernel always reduces over top-k experts)."
        )
    if (
        runner_config.gemm1_alpha is not None
        or runner_config.gemm1_clamp_limit is not None
        or runner_config.swiglu_limit is not None
    ):
        raise ValueError(
            "The hpc_ops MoE runner backend runs a plain SiLU-and-mul; it does "
            "not support gemm1_alpha / gemm1_clamp_limit / swiglu_limit."
        )


@register_fused_func("none", "hpc_ops")
def fused_experts_none_to_hpc_ops(
    dispatch_output: StandardDispatchOutput,
    quant_info: HpcOpsMoeQuantInfo,
    runner_config: MoeRunnerConfig,
) -> StandardCombineInput:
    import hpc

    from sglang.kernels.ops.quantization.fp8_kernel import (
        scaled_fp8_quant,
        sglang_per_token_group_quant_fp8,
    )
    from sglang.srt.layers.moe.token_dispatcher.standard import StandardCombineInput

View on GitHub (pinned to 0132848349)

Solutions

1. Use the triton or flashinfer backend for this model
2. Remove --moe-runner-backend hpc_ops and let SGLang pick the default
3. Keep hpc_ops only for plain gated-SiLU FP8 models

cfg_keys = ('gemm1_alpha','gemm1_clamp_limit','swiglu_limit')
if server_args.moe_runner_backend == 'hpc_ops' and any(getattr(moe_cfg, k, None) is not None for k in cfg_keys):
    server_args.moe_runner_backend = None

When it happens

Trigger: Loading a model whose MoE config defines gemm1_alpha / gemm1_clamp_limit / swiglu_limit (gpt-oss, some Qwen/Kimi variants) with --moe-runner-backend hpc_ops.

Common situations: Serving gpt-oss or other swiglu-limit models and trying hpc_ops after seeing it benchmark well on DeepSeek; version upgrades that expose these knobs in MoE runner config.

Related errors

• The hpc_ops MoE runner backend does not support fused shared
• The hpc_ops MoE runner backend does not support apply_router
• The hpc_ops MoE runner backend does not support no_combine (
• The hpc_ops MoE runner backend only supports FP8-quantized M
• Unsupported activation: {self.activation}

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/250754eead056d80. Report an issue: GitHub.