vllm.v1.metrics.buckets ¶

def build_1_2_5_buckets(max_value: int) -> list[float]:
    """Build a 1-2-5 series of buckets capped at `max_value`.

    Example:
        >>> build_1_2_5_buckets(100)
        [1, 2, 5, 10, 20, 50, 100]
    """
    return build_buckets([1, 2, 5], max_value)

def build_buckets(mantissa_lst: list[int], max_value: int) -> list[float]:
    """Build buckets from mantissas scaled by increasing powers of 10.

    Args:
        mantissa_lst: Mantissa values multiplied by each power of 10.
        max_value: Largest bucket value to include.

    Returns:
        Bucket values in increasing order, capped at `max_value`.
    """
    exponent = 0
    buckets: list[float] = []
    while True:
        for m in mantissa_lst:
            value = m * 10**exponent
            if value <= max_value:
                buckets.append(value)
            else:
                return buckets
        exponent += 1

def histogram_buckets(
    family: BucketFamilyKey,
    max_model_len: int | None = None,
) -> list[float]:
    """Return the default bucket boundaries for a histogram family.

    Args:
        family: Canonical bucket family key.
        max_model_len: Cap for the token-count series of the
            `request_tokens` family; required for that family and ignored
            otherwise.

    Returns:
        A fresh list of bucket upper bounds; callers may mutate it freely.

    Raises:
        ValueError: If `family` is `request_tokens` and `max_model_len`
            is None.
    """
    if family == "request_tokens":
        if max_model_len is None:
            raise ValueError(
                "max_model_len is required for the 'request_tokens' bucket family"
            )
        return build_1_2_5_buckets(max_model_len)
    return list(_STATIC_FAMILY_DEFAULTS[family])