sgl-project/sglang · error · ValueError

--mamba-max-states-per-path must be -1 (unlimited) or a posi

Error message

--mamba-max-states-per-path must be -1 (unlimited) or a positive integer, got {cfg.mamba_max_states_per_path}.

What it means

ServerArgs validation raised in _handle_mamba_backend during the resolution pipeline. --mamba-max-states-per-path controls how many Mamba states are retained per request path; the only legal values are -1 (unlimited) or a positive integer. Any value of 0 or below -1 fails fast because it would produce a degenerate/invalid state budget.

Source

Thrown at python/sglang/srt/server_args.py:6752

        if cfg.pre_warm_nccl and not (is_cuda() or is_hip() or is_npu()):
            logger.warning(
                "pre_warm_nccl is only applicable for CUDA or HIP hardware or NPU hardware. "
                "Ignoring pre_warm_nccl setting on current hardware."
            )
            self._declare("_handle_nccl_pre_warm", pre_warm_nccl=False)

    def _handle_grammar_backend(self):
        cfg = resolving_view(self)
        if cfg.grammar_backend is None:
            self._declare("_handle_grammar_backend", grammar_backend="xgrammar")

    def _handle_mamba_backend(self):
        cfg = resolving_view(self)
        if cfg.mamba_cache_philox_rounds < 0:
            raise ValueError("--mamba-cache-philox-rounds must be non-negative.")

        if cfg.mamba_max_states_per_path == 0 or cfg.mamba_max_states_per_path < -1:
            raise ValueError(
                "--mamba-max-states-per-path must be -1 (unlimited) or a positive "
                f"integer, got {cfg.mamba_max_states_per_path}."
            )

        if cfg.enable_mamba_cache_stochastic_rounding:
            if cfg.mamba_ssm_dtype != "float16":
                raise ValueError(
                    "Stochastic rounding for the Mamba SSM cache requires "
                    f"--mamba-ssm-dtype float16, got {cfg.mamba_ssm_dtype!r}. "
                    "Run with --mamba-ssm-dtype float16 or disable "
                    "--enable-mamba-cache-stochastic-rounding."
                )
            if not is_cuda():
                raise ValueError(
                    "Stochastic rounding for the Mamba SSM cache is only "
                    "supported on NVIDIA CUDA platforms. Disable "
                    "--enable-mamba-cache-stochastic-rounding on this platform."
                )

View on GitHub (pinned to 0132848349)

Solutions

1. Set --mamba-max-states-per-path -1 for unlimited states
2. Set it to a positive integer (e.g. 8, 16) to cap states per path
3. Remove the flag entirely if the default value is acceptable

Example fix

# before
python -m sglang.launch_server --mamba-max-states-per-path 0
# after
python -m sglang.launch_server --mamba-max-states-per-path -1

v = args.mamba_max_states_per_path
assert v == -1 or v > 0, f"mamba_max_states_per_path must be -1 or positive, got {v}"

null

When it happens

Trigger: Launching the server with --mamba-max-states-per-path 0 or any value < -1 (e.g. -2, -5). The check runs in ServerArgs._handle_mamba_backend via _run_resolution_pipeline, so it fires at argument parsing/startup time before any model loads.

Common situations: Typos or scripted configs that pass 0 thinking it means 'unlimited', arithmetic that computes a negative cap, or copying a config from another tool where 0 is the unlimited sentinel.

Related errors

• Decode context parallel size (--dcp-size / --decode-context-
• Stochastic rounding for the Mamba SSM cache requires --mamba
• Stochastic rounding for the Mamba SSM cache is only supporte
• --enable-int8-mamba-checkpoint is not supported together wit
• --enable-int8-mamba-checkpoint only supports the built-in ma

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/add59746e337dcf9. Report an issue: GitHub.