Function main

benchmarking/vllm/parse_engine_stats.py:122–159

()

def main():
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument("log", type=Path, nargs="+", help="Path(s) to sweep.log")
    parser.add_argument(
        "--by-concurrency",
        action="store_true",
        help="Aggregate across runs per concurrency level",
    )
    args = parser.parse_args()

    all_runs = []
    for log_path in args.log:
        runs = parse_sweep_log(log_path)
        all_runs.extend(runs)

    if not all_runs:
        print("No benchmark runs found in log(s).", file=sys.stderr)
        sys.exit(1)

    df = summarize(all_runs)

    if args.by_concurrency:
        agg = (
            df.groupby("concurrency")
            .agg(
                n_logs=("n_logs", "sum"),
                kv_mean=("kv_mean", "mean"),
                kv_max=("kv_max", "max"),
                running_mean=("running_mean", "mean"),
                running_max=("running_max", "max"),
                waiting_mean=("waiting_mean", "mean"),
                waiting_max=("waiting_max", "max"),
            )
            .round(1)
        )
        print(agg.to_markdown())
    else:
        print(df.round(1).to_markdown(index=False))


if __name__ == "__main__":