GLM-5.3-FlashX 冲到 200 tokens/s：真正该测的不只是生成速度

python -m pip install -U openai

export OPENAI_BASE_URL="https://your-endpoint.example/v1"
export OPENAI_API_KEY="replace-with-your-api-key"
export MODEL_NAME="glm-5.3-flashx"

import json
import os
import statistics
import time

from openai import OpenAI

client = OpenAI(
    base_url=os.environ["OPENAI_BASE_URL"],
    api_key=os.environ["OPENAI_API_KEY"],
)
model = os.getenv("MODEL_NAME", "glm-5.3-flashx")

prompt = """请为一个 Python Web 服务编写上线检查清单，覆盖配置、数据库、
可观测性、回滚和安全。输出约 800 个中文字符，使用 Markdown 列表。"""


def run_once():
    started = time.perf_counter()
    first_token_at = None
    completion_tokens = None
    parts = []

    stream = client.chat.completions.create(
        model=model,
        messages=[{"role": "user", "content": prompt}],
        temperature=0,
        max_tokens=1000,
        stream=True,
        # 如果服务不支持该参数，请删除它；此时脚本无法得到精确 token 数。
        stream_options={"include_usage": True},
    )

    for chunk in stream:
        delta = chunk.choices[0].delta if chunk.choices else None
        content = getattr(delta, "content", None) if delta else None
        if content:
            if first_token_at is None:
                first_token_at = time.perf_counter()
            parts.append(content)

        usage = getattr(chunk, "usage", None)
        if usage and usage.completion_tokens is not None:
            completion_tokens = usage.completion_tokens

    finished = time.perf_counter()
    first_token_at = first_token_at or finished
    generation_seconds = max(finished - first_token_at, 1e-9)

    return {
        "ttft_seconds": round(first_token_at - started, 3),
        "total_seconds": round(finished - started, 3),
        "completion_tokens": completion_tokens,
        "generation_tokens_per_second": (
            round(completion_tokens / generation_seconds, 2)
            if completion_tokens is not None
            else None
        ),
        "output_characters": len("".join(parts)),
    }


# 一次预热，减少连接建立和冷启动对结果的干扰。
run_once()
results = [run_once() for _ in range(5)]

print(json.dumps(results, ensure_ascii=False, indent=2))
print(
    "median_ttft_seconds=",
    statistics.median(item["ttft_seconds"] for item in results),
)

speeds = [
    item["generation_tokens_per_second"]
    for item in results
    if item["generation_tokens_per_second"] is not None
]
if speeds:
    print("median_generation_tokens_per_second=", statistics.median(speeds))

python bench_glm.py