sgl-project/sglang · error · ValueError

Invalid --log-requests-level: {self.log_requests_level=}

Error message

Invalid --log-requests-level: {self.log_requests_level=}

What it means

RequestLogger validates --log-requests-level: with output logging enabled, only levels 0-3 are accepted (2 => max_length 2048, 3 => ~1GB, etc.). Any other integer raises ValueError echoing the value.

Source

Thrown at python/sglang/srt/utils/request_logger.py:230

                out_skip_names = {"text", "output_ids", "embedding"}
            elif self.log_requests_level == 1:
                max_length = 1 << 30
                skip_names = {
                    "text",
                    "input_ids",
                    "input_embeds",
                    "image_data",
                    "audio_data",
                    "video_data",
                    "lora_path",
                }
                out_skip_names = {"text", "output_ids", "embedding"}
            elif self.log_requests_level == 2:
                max_length = 2048
            elif self.log_requests_level == 3:
                max_length = 1 << 30
            else:
                raise ValueError(
                    f"Invalid --log-requests-level: {self.log_requests_level=}"
                )
        return max_length, skip_names, out_skip_names

    def _log(self, msg: str) -> None:
        for target in self.targets:
            target.info(msg)


# TODO unify this w/ `_transform_data_for_logging` if we find performance enough
def _dataclass_to_string_truncated(
    data: Any, max_length: int = 2048, skip_names: Optional[Set[str]] = None
) -> str:
    if skip_names is None:
        skip_names = set()
    if isinstance(data, str):
        if len(data) > max_length:
            half_length = max_length // 2

View on GitHub (pinned to 0132848349)

Solutions

1. Use a valid level: 0 (off/simple), 1, 2, or 3
2. Check the --log-requests-level docs/help for the accepted range
3. If the intent is 'log everything', use level 3

Example fix

# before
--log-requests-level 4
# after
--log-requests-level 3

assert log_requests_level in (0, 1, 2, 3), f'--log-requests-level must be 0-3, got {log_requests_level}'

def is_valid_log_level(v) -> bool:
    return v in (0, 1, 2, 3)

try:
    RequestLogger(...)
except ValueError as e:
    if 'log-requests-level' in str(e):
        level = 1; RequestLogger(...)

When it happens

Trigger: Launching with --log-requests-level 4 (or -1) together with an output target configured; _compute_metadata runs during __init__/configure.

Common situations: Typos in launch flags, scripts incrementing the level beyond 3, or stale configs written against a different level range.

Related errors

• Quantization method specified in the model config ({quant_me
• --mamba-cache-philox-rounds must be non-negative.
• Unknown serve backend {name!r}. Available values: {available
• Multiple distributions register serve backend {name!r}: {pro
• Failed to load serve backend {name!r} from {self._entry_poin

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/84fb0575e42b49f4. Report an issue: GitHub.