再测 DeepSeek 各家 API 真实速度（附：测试脚本）

测试方法

#测试 prompt：给我写一首七言绝句，赞叹祖国的大好河山

代码

import timefrom openai import OpenAIfrom google.colab import userdataimport datetimeimport pytz

def test_provider(provider_config, messages):    """    根据传入的 provider 配置及消息，测试生成过程，并统计各阶段指标。    如果测试过程中出现任何错误，则打印错误信息并跳过当前服务商。    """    provider_name = provider_config.get("name", "Unnamed Provider")    print(f"\n---------------------------")    print(f"开始测试服务商：{provider_name}")    print(f"---------------------------\n")
    try:        api_key = provider_config.get("api_key")        base_url = provider_config.get("base_url")        model = provider_config.get("model")
        # 初始化客户端（请确保你使用的 OpenAI 客户端支持这些参数）        client = OpenAI(api_key=api_key, base_url=base_url)
        # 初始化 token 计数器与文本变量        prompt_tokens = 0        completion_tokens = 0        reasoning_tokens = 0        content_tokens = 0        total_tokens = 0
        reasoning_text = ""        content_text = ""
        # 初始化计时变量        start_time = time.time()        first_token_time = None
        # 用于记录 reasoning 与 content 部分开始与结束的时刻        reasoning_start_time = None        reasoning_end_time = None        content_start_time = None        content_end_time = None
        # 吐槽：各家 usage_content 的更新方式不同，有的是过程中更新，有的是最后展示        usage_content = ""
        # 发起流式请求        response = client.chat.completions.create(            model=model,            messages=messages,            stream=True,            stream_options={"include_usage": True},        )
        # 遍历每个流式响应块        for chunk in response:            # 若 chunk 中没有 choices 信息，则检查是否有 usage 信息打印后继续            if chunk.usage:                usage_content = chunk.usage                # 如果 completion_tokens_details 为 None，则说明没有单独的推理信息，此时设置 reasoning_tokens 为 0                if chunk.usage.completion_tokens_details is None:                    reasoning_tokens = 0                else:                    reasoning_tokens = chunk.usage.completion_tokens_details.reasoning_tokens                prompt_tokens = chunk.usage.prompt_tokens                completion_tokens = chunk.usage.completion_tokens                content_tokens = completion_tokens - reasoning_tokens                total_tokens = chunk.usage.total_tokens 

            # 获取第一个 choice 的 delta            if not chunk.choices:                continue            delta = chunk.choices[0].delta            # 尝试获取 reasoning 与 content 片段（可能为空字符串）            reasoning_piece = getattr(delta, 'reasoning_content', "")            content_piece = getattr(delta, 'content', "")
            # 记录首个 token 到达时间（仅记录一次）            if first_token_time is None and (reasoning_piece or content_piece):                first_token_time = time.time() - start_time
            # 如果有 reasoning 内容            if reasoning_piece:                if reasoning_start_time is None:                    reasoning_start_time = time.time()                reasoning_text += reasoning_piece                reasoning_end_time = time.time()  # 每次更新，最终记录最后一次收到的时刻                print(reasoning_piece, end='', flush=True)
            # 如果有 content 内容            elif content_piece:                if content_start_time is None:                    content_start_time = time.time()                content_text += content_piece                content_end_time = time.time()  # 每次更新                print(content_piece, end='', flush=True)

        total_time = time.time() - start_time        reasoning_time = (reasoning_end_time - reasoning_start_time) if (reasoning_start_time and reasoning_end_time) else 0        content_time = (content_end_time - content_start_time) if (content_start_time and content_end_time) else 0                if(usage_content):            print("\n\n【Usage 信息】")            print(usage_content)                        # 输出测试指标，格式参考如下        print("\n\n【%s】" % provider_name)        if first_token_time is not None:            print(f"首 token 响应时间：{first_token_time:.2f} 秒")        else:            print("未收到 token 响应。")        if reasoning_tokens > 0:            print(f"Reasoning 部分：{len(reasoning_text)} 字符，{reasoning_tokens} tokens, 用时：{reasoning_time:.2f} 秒, 生成速度：{reasoning_tokens / reasoning_time if reasoning_time > 0 else 0:.2f} tokens/s")            print(f"Content 部分：{len(content_text)} 字符，{content_tokens} tokens, 用时：{content_time:.2f} 秒, 生成速度：{content_tokens / content_time if content_time > 0 else 0:.2f} tokens/s")        # 总体生成始终打印        print(f"内容生成：{len(reasoning_text + content_text)} 字符，{completion_tokens} tokens, 总用时：{total_time:.2f} 秒, 生成速度：{completion_tokens / total_time if total_time > 0 else 0:.2f} tokens/s")        print("\n***************************\n")
        return {            "provider": provider_name,            "first_token_time": first_token_time,            "reasoning_tokens": reasoning_tokens,            "reasoning_time": reasoning_time,            "content_tokens": content_tokens,            "content_time": content_time,            "total_tokens": total_tokens,            "total_time": total_time,        }
    except Exception as e:        # 如果出现任何错误，则打印错误信息并跳过该服务商        print(f"服务商 {provider_name} 测试过程中发生错误：{e}")        print("\n---------------------------\n")        return None

if __name__ == "__main__":    # 待测试的对话消息（此处为示例：写一首七言绝句赞美祖国大好河山）    messages = [        {            'role': 'user',            'content': "给我写一首七言绝句，赞叹祖国的大好河山"        }    ]
    # 定义各服务商的配置    providers = [        {            "name": "DeepSeek 官方",  # 这个经常不可用            "api_key": userdata.get('Key_DeepSeek'),  # 请替换为真实 API Key：https://platform.deepseek.com/api_keys            "base_url": "https://api.deepseek.com",            "model": "deepseek-reasoner"        },         {            "name": "阿里云/百炼",            "api_key": userdata.get('Key_Aliyun'),  # 请替换为真实 API Key：https://bailian.console.aliyun.com/?apiKey=1#/api-key            "base_url": "https://dashscope.aliyuncs.com/compatible-mode/v1",            "model": "deepseek-r1"        },         {            "name": "百度千帆",            "api_key": userdata.get('Key_Qianfan'),   # 请替换为真实 API Key：https://console.bce.baidu.com/iam/#/iam/apikey/list            "base_url": "https://qianfan.baidubce.com/v2",            "model": "deepseek-r1"        },        {            "name": "硅基流动",            "api_key": userdata.get('Key_Siliconflow'),  # 请替换为真实 API Key：https://cloud.siliconflow.cn/account/ak            "base_url": "https://api.siliconflow.cn/v1",            "model": "deepseek-ai/DeepSeek-R1"        },        {            "name": "硅基流动Pro",            "api_key": userdata.get('Key_Siliconflow'),  # 请替换为真实 API Key：https://cloud.siliconflow.cn/account/ak            "base_url": "https://api.siliconflow.cn/v1",            "model": "Pro/deepseek-ai/DeepSeek-R1"        },        {            "name": "火山引擎",            "api_key": userdata.get('Key_Volces'),  # 请替换为真实 API Key：https://console.volcengine.com/ark/region:ark+cn-beijing/apiKey?apikey=%7B%7D            "base_url": "https://ark.cn-beijing.volces.com/api/v3",            "model": userdata.get('Endpoint_Volces_R1') # 火山引擎这里叫接入点，在这里创建：https://console.volcengine.com/ark/region:ark+cn-beijing/endpoint?config=%7B%7D        },         {            "name": "腾讯云",            "api_key": userdata.get('Key_Tencentcloud'),  # 请替换为真实 API Key：https://console.cloud.tencent.com/lkeap            "base_url": "https://api.lkeap.cloud.tencent.com/v1",            "model": "deepseek-r1"        },    ]

    print(f"本次测试开始于中国时间：{datetime.datetime.now(pytz.timezone('Asia/Shanghai')).strftime('%Y-%m-%d %H:%M:%S')}")
    # 循环对每个服务商进行测试    for provider in providers:        test_provider(provider, messages)