diff --git a/benchmarks/multi_turn/benchmark_serving_multi_turn.py b/benchmarks/multi_turn/benchmark_serving_multi_turn.py index e23f6b923f..881039f43f 100644 --- a/benchmarks/multi_turn/benchmark_serving_multi_turn.py +++ b/benchmarks/multi_turn/benchmark_serving_multi_turn.py @@ -1439,6 +1439,12 @@ async def main() -> None: action="store_true", help="Export summary to Excel file (optional)", ) + parser.add_argument( + "--stats-json-output", + type=str, + default=None, + help="Export per-request stats (ttft_ms, tpot_ms, etc.) to a JSON file", + ) parser.add_argument( "-v", "--verbose", @@ -1651,6 +1657,19 @@ async def main() -> None: warmup_runtime_sec=warmup_runtime_sec, ) + if args.stats_json_output is not None: + # Export per-request metrics as a JSON array for downstream analysis. + stats_data = [s._asdict() for s in client_metrics] + logger.info( + f"{Color.GREEN}Writing per-request stats JSON: " + f"{args.stats_json_output}{Color.RESET}" + ) + os.makedirs( + os.path.dirname(os.path.abspath(args.stats_json_output)), exist_ok=True + ) + with open(args.stats_json_output, "w") as f: + json.dump(stats_data, f, indent=2) + if args.output_file is not None: # Write a JSON file with the updated conversations # The "assistant" content will contain the answers from the tested LLM