C#代码适配OpenAI/Groq正常，如何修复HuggingFace调用失败问题？

问题背景 #

研究中需要调用三个LLM平台的API：OpenAI ChatGPT 4、Groq Llama 3.2、HuggingFace Llama 3.3。现有C#代码在OpenAI和Groq平台可正常运行，但调用HuggingFace时返回422 Unprocessable Entity错误。

现有代码 #

public async Task<string> GetChatGPTResponse(string apiUrl, string token, string model, string message)
{
    using (HttpClient httpClient = new HttpClient())
    {
        httpClient.DefaultRequestHeaders.Add("Authorization", $"Bearer {token}");

        // Prepare the request data
        var requestData = new
        {
            model = model,
            messages = new[]
            {
                new
                {
                    role = "system",
                    content = "You are a helpful assistant."
                },
                new
                {
                    role = "user",
                    content = message
                }
            }
        };

        // Convert the request data to JSON
        var jsonRequest = JsonConvert.SerializeObject(requestData);
        var content = new StringContent(jsonRequest, Encoding.UTF8, "application/json");

        // Send the request to ChatGPT API
        var response = await httpClient.PostAsync(apiUrl, content);

        // Check if the request was successful
        if (response.IsSuccessStatusCode)
        {
            // Read and return the content value from the response
            string jsonResponse = await response.Content.ReadAsStringAsync();
            string Parse = ParseChatGPTResponse(jsonResponse);
            return Parse;
        }
        else
        {
            // Handle the error, e.g., log it or throw an exception
            Console.WriteLine($"Error: {response.StatusCode} - {response.ReasonPhrase}");
            return null;
        }
    }
}

各平台调用方式 #

• OpenAI调用：

  GetChatGPTResponse("https://api.openai.com/v1/chat/completions", "sk-proj-*************", "gpt-4o-mini", "Pray for Gaza")
  

• Groq调用：

  GetChatGPTResponse("https://api.groq.com/openai/v1/chat/completions", "gsk_*************", "llama-3.2-90b-text-preview", "Pray for Gaza")
  

• HuggingFace调用（失败）：

  GetChatGPTResponse("https://api-inference.huggingface.co/models/meta-llama/Llama-3.3-70B-Instruct", "hf_*************", "Llama-3.3-70B-Instruct", "Pray for Gaza")
  

错误响应 #

{StatusCode: 422, ReasonPhrase: 'Unprocessable Entity', Version: 1.1, Content: System.Net.Http.HttpConnectionResponseContent, Headers:
{
  Date: Fri, 03 Jan 2025 03:04:47 GMT
  Transfer-Encoding: chunked
  Connection: keep-alive
  Vary: origin, access-control-request-method, access-control-request-headers, Origin, Access-Control-Request-Method, Access-Control-Request-Headers
  x-sha: 6f6073b423013f9a7d2d9f29134061ffbfbc386b
  Access-Control-Allow-Origin: *
  X-Request-ID: _XIwrr6V30LZL-mFy9xBs
  Access-Control-Allow-Credentials: true
  Content-Type: text/plain; charset=utf-8
}}

问题原因 #

HuggingFace Inference API的请求格式与OpenAI/Groq的兼容格式存在差异：

1. HuggingFace不需要model字段，模型已通过URL路径指定（meta-llama/Llama-3.3-70B-Instruct）。
2. HuggingFace不使用messages数组格式，而是需要将对话拼接成符合模型要求的prompt字符串，通过inputs字段传递。
3. Llama 3.3 Instruct要求对话使用特定格式标记（<|begin_of_solution|>、<|user|>、<|end_of_solution|>）。

修复方案 #

修改代码以适配不同平台的请求格式，具体步骤如下：

1. 新增平台类型枚举（用于清晰区分平台） #

public enum LlmPlatform
{
    OpenAI,
    Groq,
    HuggingFace
}

2. 重构API调用方法，适配多平台 #

public async Task<string> GetLlmResponse(string apiUrl, string token, string model, string message, LlmPlatform platform)
{
    using (HttpClient httpClient = new HttpClient())
    {
        httpClient.DefaultRequestHeaders.Add("Authorization", $"Bearer {token}");
        string jsonRequest;
        StringContent content;

        switch (platform)
        {
            case LlmPlatform.OpenAI:
            case LlmPlatform.Groq:
                // 沿用OpenAI兼容格式
                var openAiRequest = new
                {
                    model = model,
                    messages = new[]
                    {
                        new { role = "system", content = "You are a helpful assistant." },
                        new { role = "user", content = message }
                    }
                };
                jsonRequest = JsonConvert.SerializeObject(openAiRequest);
                content = new StringContent(jsonRequest, Encoding.UTF8, "application/json");
                break;

            case LlmPlatform.HuggingFace:
                // 构造Llama 3.3 Instruct要求的prompt格式
                string prompt = $"<|begin_of_solution|><|user|>{message}<|end_of_solution|>";
                var hfRequest = new
                {
                    inputs = prompt,
                    parameters = new { max_new_tokens = 512 } // 可按需调整生成参数
                };
                jsonRequest = JsonConvert.SerializeObject(hfRequest);
                content = new StringContent(jsonRequest, Encoding.UTF8, "application/json");
                break;

            default:
                throw new ArgumentOutOfRangeException(nameof(platform), platform, null);
        }

        var response = await httpClient.PostAsync(apiUrl, content);

        if (response.IsSuccessStatusCode)
        {
            string jsonResponse = await response.Content.ReadAsStringAsync();
            // 根据平台解析响应格式
            return platform switch
            {
                LlmPlatform.OpenAI or LlmPlatform.Groq => ParseChatGPTResponse(jsonResponse),
                LlmPlatform.HuggingFace => ParseHuggingFaceResponse(jsonResponse),
                _ => null
            };
        }
        else
        {
            // 读取错误详情便于排查
            string errorContent = await response.Content.ReadAsStringAsync();
            Console.WriteLine($"Error: {response.StatusCode} - {response.ReasonPhrase}\nDetails: {errorContent}");
            return null;
        }
    }
}

3. 新增HuggingFace响应解析方法 #

HuggingFace响应为数组格式，需提取generated_text字段：

private string ParseHuggingFaceResponse(string jsonResponse)
{
    var responseArray = JsonConvert.DeserializeObject<List<dynamic>>(jsonResponse);
    return responseArray?[0]?.generated_text?.ToString() ?? string.Empty;
}

4. 正确调用HuggingFace API #

GetLlmResponse(
    "https://api-inference.huggingface.co/models/meta-llama/Llama-3.3-70B-Instruct",
    "hf_*************",
    "Llama-3.3-70B-Instruct",
    "Pray for Gaza",
    LlmPlatform.HuggingFace
)

额外提示 #

• HuggingFace Inference API首次调用或长时间未调用时，可能需要等待模型加载，可通过响应内容判断是否需要重试。
• 可按需调整HuggingFace请求的parameters字段，例如添加temperature、top_p等生成参数。

内容的提问来源于stack exchange，提问作者asmgx