Function quick_summary

benchmarks/quick_benchmark.py:556–598

快速评估

(quick_result, selected_percentile_metrics, metrics)

def quick_summary(quick_result, selected_percentile_metrics, metrics):
    """
    快速评估
    """

    def process_quick_metric(
        metric_attribute_name: str,
        metric_name: str,
        metric_header: str,
    ):
        if metric_attribute_name not in selected_percentile_metrics:
            return
        print("{s:{c}^{n}}".format(s=metric_header, n=50, c="-"))
        mean_value = getattr(metrics, f"mean_{metric_attribute_name}_ms")
        print("{:<40} {:<10.2f}".format(f"Mean {metric_name} (ms):", mean_value))
        quick_result[f"mean_{metric_attribute_name}_ms"] = mean_value

    def process_quick_length(
        metric_attribute_name: str,
        metric_name: str,
        metric_header: str,
    ):
        if metric_attribute_name not in selected_percentile_metrics:
            return
        print("{s:{c}^{n}}".format(s=metric_header, n=50, c="-"))
        mean_value = getattr(metrics, f"mean_{metric_attribute_name}")
        print("{:<40} {:<10.2f}".format(f"Mean {metric_name}:", mean_value))
        quick_result[f"mean_{metric_attribute_name}"] = mean_value

    print("\n\n\n")
    print("{s:{c}^{n}}".format(s=" Benchmark Quick Summary ", n=50, c="="))
    process_quick_length("s_decode", "Decode", "解码速度(tok/s)")
    process_quick_metric("ttft", "TTFT", "Time to First Token")
    process_quick_metric("s_ttft", "S_TTFT", "Infer Time to First Token")
    process_quick_metric("tpot", "TPOT", "Time per Output Token (excl. 1st token)")
    process_quick_metric("itl", "ITL", "Inter-token Latency")
    process_quick_metric("s_itl", "S_ITL", "Infer Inter-token Latency")
    process_quick_metric("e2el", "E2EL", "End-to-end Latency")
    process_quick_metric("s_e2el", "S_E2EL", "Infer End-to-end Latency")
    process_quick_length("input_len", "Cached Tokens", "Cached Tokens")
    process_quick_length("s_input_len", "Input Length", "Infer Input Length")
    process_quick_length("output_len", "Output Length", "Output Length")
    print("=" * 50)


def check_goodput_args(args):