vLLM v1/metrics/buckets.py - default Prometheus histogram bucket boundaries

The single source of truth for every core vLLM histogram's default bucket bounds, read 2026-10-04. Seven static families with literal tuples: REQUEST_LATENCY_BUCKETS (21 bounds, 0.3 .. 7680.0), TIME_TO_FIRST_TOKEN_BUCKETS (22 bounds, 0.001 .. 2560.0), INTER_TOKEN_LATENCY_BUCKETS (19 bounds, 0.01 .. 80.0), ITERATION_TOKENS_BUCKETS (13 bounds, 1 .. 16384), REQUEST_PARAMS_N_BUCKETS (5 bounds), REQUEST_NUM_PREEMPTIONS_BUCKETS (7 bounds, 1 .. 20), KV_CACHE_RESIDENCY_BUCKETS (21 bounds, 0.001 .. 1800). The request_tokens family is NOT static: histogram_buckets() builds it with build_1_2_5_buckets(max_model_len), so its bounds depend on the served model's context length. BUCKET_FAMILY_KEYS is the set of eight keys --custom-histogram-buckets accepts. The module's own docstrings state the design intent of each family, which is why the boundaries are quotable rather than arbitrary. | 2026-10-05 PATH SWEEP: PATH SWEEP RESULT: NO REPAIR NEEDED. The same agent report that cast doubt on vllm/v1/metrics/loggers.py also implied buckets.py had moved. It has not: vllm/v1/metrics/buckets.py is present in the authenticated recursive tree for vllm@main. The bucket-boundary values quoted in this note are therefore still quotable from the file this record cites.