sgl-project/sglang · error · TypeError

rank_consensus() got unexpected keyword argument(s): {list(k

Error message

rank_consensus() got unexpected keyword argument(s): {list(kwargs)}

What it means

The rank_consensus decorator factory accepts only positional usage — rank_consensus(func) — and no keyword arguments; any kwargs trigger this TypeError. Configuration is done via the decorator's own named parameters (same_params, same_results), not via call-time kwargs.

Source

Thrown at python/sglang/srt/utils/rank_consensus_checker.py:77

    def foo():
        return 1

    * Assert for part of the results are same.
    @rank_consensus(same_results = ["result.some_field"]
    def foo():
        return SomeObject()

    @rank_consensus(same_results = ["result.field", "len(result.field2)"]
    def foo():
        return SomeObject()

    * Assert the function is called by all ranks and all parameters and results are the same.
    @rank_consensus(same_params = True, same_results = True)
    def foo():
        return 1
    """
    if kwargs:
        raise TypeError(
            f"rank_consensus() got unexpected keyword argument(s): " f"{list(kwargs)}"
        )

    params_selector = _normalize_selector(same_params, "same_params")
    results_selector = _normalize_selector(same_results, "same_results")

    def decorator(func: Callable) -> Callable:
        # This decorator function called at import time.  So it should be zero runtime overhead
        # when the consensus checker is disabled.
        if not envs.SGLANG_ENABLE_RANK_CONSENSUS_CHECKER.get():
            return func

        # Unwrap static/class-method descriptors so we always operate on the
        # raw function. We remember the descriptor type so we can re-wrap the
        # result and the class-body descriptor protocol keeps working.
        if isinstance(func, (classmethod, staticmethod)):
            raw_func = func.__func__
            descriptor_type = type(func)

View on GitHub (pinned to 0132848349)

Solutions

1. Use only the supported parameters: @rank_consensus(same_params=..., same_results=...)
2. If decorating directly without options, write @rank_consensus (no parens)
3. Check the decorator signature in rank_consensus_checker.py for supported knobs

Example fix

# before
@rank_consensus(mode='strict')
def check(): ...
# after
@rank_consensus(same_params=True, same_results=True)
def check(): ...

import inspect
allowed = set(inspect.signature(rank_consensus).parameters) - {'kwargs'}
assert set(kwargs) <= set()

null

try:
    rank_consensus(fn, **opts)
except TypeError as e:
    if 'unexpected keyword' in str(e):
        rank_consensus(fn)  # decorate without options

When it happens

Trigger: Writing @rank_consensus(check=True) or calling rank_consensus(fn, foo=1) — i.e. passing unexpected keyword arguments to the decorator factory itself.

Common situations: Developers assuming the checker takes options at decoration time like other decorators, or copy-pasting from a different consensus helper's API.

Related errors

• {name} must be True / False / str / list[str], got {value!r}
• cannot mix lambda extractor with static kwargs
• must provide either a lambda or static kwargs
• NCCL only supports CUDA, ROCm and MUSA backends.
• Multi-node weight cache daemons (nnodes > 1) require --dist-

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/385b109bc10f7f48. Report an issue: GitHub.