vllm-project/vllm · error · ValueError

VLLM_ROCM_QUICK_REDUCE_MIN_SIZE_BYTES_MB must be non-negativ

Error message

VLLM_ROCM_QUICK_REDUCE_MIN_SIZE_BYTES_MB must be non-negative, got {qr_min_size}

What it means

CustomQuickAllReduce._get_qr_min_size reads VLLM_ROCM_QUICK_REDUCE_MIN_SIZE_BYTES_MB and validates it: a negative value raises ValueError because a minimum-size threshold below zero is meaningless. The value is later multiplied by MB to become the byte threshold under which the ROCm quick allreduce path is skipped.

Source

Thrown at vllm/distributed/device_communicators/quick_all_reduce.py:267

            logger.info(
                "Custom quick allreduce: min size override = %d MB",
                qr_min_size // MB,
            )
        if self.qr_quantization_min_size is not None:
            logger.info(
                "Custom quick allreduce: quantization codec threshold = %d KB",
                self.qr_quantization_min_size // KB,
            )
        self.create_shared_buffer()
        self.disabled = False

    @staticmethod
    def _get_qr_min_size(qr_max_size: int | None) -> int | None:
        qr_min_size = envs.VLLM_ROCM_QUICK_REDUCE_MIN_SIZE_BYTES_MB
        if qr_min_size is None:
            return None
        if qr_min_size < 0:
            raise ValueError(
                "VLLM_ROCM_QUICK_REDUCE_MIN_SIZE_BYTES_MB must be non-negative, "
                f"got {qr_min_size}"
            )
        qr_min_size *= MB
        if qr_max_size is not None and qr_min_size > qr_max_size:
            raise ValueError(
                "VLLM_ROCM_QUICK_REDUCE_MIN_SIZE_BYTES_MB must be less than or "
                "equal to the effective QuickReduce max size"
            )
        return qr_min_size

    @staticmethod
    def _get_qr_quantization_min_size() -> int | None:
        quantization_min_size = envs.VLLM_ROCM_QUICK_REDUCE_QUANTIZATION_MIN_SIZE_KB
        if quantization_min_size is None:
            return None
        if quantization_min_size < 0:
            raise ValueError(

View on GitHub (pinned to c794754062)

Solutions

Set the variable to a non-negative value (0 disables the threshold floor) or unset it entirely
Grep deploy scripts/env files for the variable and fix the sign

Example fix

# before
export VLLM_ROCM_QUICK_REDUCE_MIN_SIZE_BYTES_MB=-1

# after
unset VLLM_ROCM_QUICK_REDUCE_MIN_SIZE_BYTES_MB
# or
export VLLM_ROCM_QUICK_REDUCE_MIN_SIZE_BYTES_MB=0

Defensive patterns

Strategy: validation

Validate before calling

v = os.environ.get("VLLM_ROCM_QUICK_REDUCE_MIN_SIZE_BYTES_MB")
if v is not None:
    assert int(v) >= 0, "VLLM_ROCM_QUICK_REDUCE_MIN_SIZE_BYTES_MB must be >= 0"

Prevention

Lint numeric env vars at launcher startup
Use 0 or unset instead of negative sentinels
Keep tuning env vars in one reviewed config file

When it happens

Trigger: Starting vLLM with VLLM_ROCM_QUICK_REDUCE_MIN_SIZE_BYTES_MB set to a negative number (e.g. -1) on an AMD MI300/MI350 (gfx94/gfx95) node where the custom quick allreduce is enabled.

Common situations: Typos in env files (e.g. '-8' vs '8'); attempting to 'disable the threshold' by passing -1 instead of unsetting the variable; CI configs copied from an experiment.

Related errors

AI-assisted analysis of vllm-project/vllm@c794754062 (2026-08-14). Data as JSON: /api/errors/9b7e8811a147b838. Report an issue: GitHub.