C语言调用Baichuan-M2-32B模型API实战-开发者社区

C语言调用Baichuan-M2-32B模型API实战

1. 引言

在嵌入式系统和资源受限环境中，直接使用C语言调用大模型API是一个极具挑战性但又非常实用的需求。本文将带你从零开始，使用纯C语言实现与Baichuan-M2-32B医疗增强推理模型的交互。

Baichuan-M2-32B作为一款专注于医疗领域的强大模型，其API调用通常使用Python等高级语言，但在嵌入式医疗设备、边缘计算等场景中，C语言的轻量级和高效性使其成为更优选择。我们将通过构建一个轻量级HTTP客户端，实现完整的模型调用流程。

2. 环境准备与基础概念

2.1 前置条件

在开始之前，请确保你已准备好：

一台运行Linux的开发机（x86或ARM架构均可）
已部署好的Baichuan-M2-32B API服务（可使用vLLM或SGLang部署）
基本的C语言开发环境（gcc、make等）
libcurl开发库（用于HTTP通信）

安装libcurl开发包：

sudo apt-get install libcurl4-openssl-dev

2.2 Baichuan-M2-32B API基础

Baichuan-M2-32B提供兼容OpenAI格式的API接口，主要端点包括：

/v1/completions：文本补全
/v1/chat/completions：对话补全（推荐）
/v1/models：模型信息查询

我们的C语言实现将主要针对/v1/chat/completions端点进行交互。

3. C语言HTTP客户端实现

3.1 基础HTTP请求封装

首先实现一个通用的HTTP POST请求函数：

#include <stdio.h> #include <stdlib.h> #include <string.h> #include <curl/curl.h> struct string { char *ptr; size_t len; }; void init_string(struct string *s) { s->len = 0; s->ptr = malloc(s->len+1); if (s->ptr == NULL) { fprintf(stderr, "malloc() failed\n"); exit(EXIT_FAILURE); } s->ptr[0] = '\0'; } size_t writefunc(void *ptr, size_t size, size_t nmemb, struct string *s) { size_t new_len = s->len + size*nmemb; s->ptr = realloc(s->ptr, new_len+1); if (s->ptr == NULL) { fprintf(stderr, "realloc() failed\n"); exit(EXIT_FAILURE); } memcpy(s->ptr+s->len, ptr, size*nmemb); s->ptr[new_len] = '\0'; s->len = new_len; return size*nmemb; } char* http_post(const char* url, const char* post_data, const char* api_key) { CURL *curl; CURLcode res; struct string s; init_string(&s); curl = curl_easy_init(); if(curl) { struct curl_slist *headers = NULL; headers = curl_slist_append(headers, "Content-Type: application/json"); if(api_key != NULL) { char auth_header[256]; snprintf(auth_header, sizeof(auth_header), "Authorization: Bearer %s", api_key); headers = curl_slist_append(headers, auth_header); } curl_easy_setopt(curl, CURLOPT_URL, url); curl_easy_setopt(curl, CURLOPT_POSTFIELDS, post_data); curl_easy_setopt(curl, CURLOPT_HTTPHEADER, headers); curl_easy_setopt(curl, CURLOPT_WRITEFUNCTION, writefunc); curl_easy_setopt(curl, CURLOPT_WRITEDATA, &s); res = curl_easy_perform(curl); if(res != CURLE_OK) { fprintf(stderr, "curl_easy_perform() failed: %s\n", curl_easy_strerror(res)); free(s.ptr); return NULL; } curl_easy_cleanup(curl); curl_slist_free_all(headers); } return s.ptr; }

3.2 构建请求JSON

我们需要构建符合Baichuan-M2-32B API要求的JSON请求体：

#include <jansson.h> char* build_chat_request(const char* prompt) { json_t *root = json_object(); json_t *messages = json_array(); json_t *message = json_object(); json_object_set_new(message, "role", json_string("user")); json_object_set_new(message, "content", json_string(prompt)); json_array_append_new(messages, message); json_object_set_new(root, "messages", messages); json_object_set_new(root, "model", json_string("Baichuan-M2-32B")); json_object_set_new(root, "max_tokens", json_integer(512)); json_object_set_new(root, "temperature", json_real(0.7)); char *request = json_dumps(root, JSON_COMPACT); json_decref(root); return request; }

3.3 解析响应JSON

处理API返回的JSON响应：

char* parse_chat_response(const char* json_response) { json_error_t error; json_t *root = json_loads(json_response, 0, &error); if(!root) { fprintf(stderr, "error: on line %d: %s\n", error.line, error.text); return NULL; } json_t *choices = json_object_get(root, "choices"); if(!json_is_array(choices)) { fprintf(stderr, "error: choices is not an array\n"); json_decref(root); return NULL; } json_t *first_choice = json_array_get(choices, 0); json_t *message = json_object_get(first_choice, "message"); json_t *content = json_object_get(message, "content"); const char *text = json_string_value(content); char *result = strdup(text); // 复制字符串以便后续使用 json_decref(root); return result; }

4. 完整调用示例

将上述组件组合成完整的调用流程：

#include <stdio.h> #include <stdlib.h> #define API_URL "http://localhost:8000/v1/chat/completions" #define API_KEY NULL // 替换为你的API密钥 int main() { // 初始化libcurl curl_global_init(CURL_GLOBAL_ALL); // 构建请求 char *request = build_chat_request("患者主诉头痛3天，伴恶心无呕吐，无发热。可能的诊断是什么？"); printf("Request JSON: %s\n", request); // 发送请求 char *response = http_post(API_URL, request, API_KEY); if(response == NULL) { fprintf(stderr, "API request failed\n"); free(request); curl_global_cleanup(); return 1; } printf("Raw response: %s\n", response); // 解析响应 char *content = parse_chat_response(response); if(content != NULL) { printf("\n模型回复:\n%s\n", content); free(content); } // 清理资源 free(request); free(response); curl_global_cleanup(); return 0; }

5. 高级主题与优化

5.1 内存管理优化

在嵌入式环境中，内存管理至关重要。我们可以优化之前的实现：

// 内存池实现 typedef struct { void **ptrs; size_t count; size_t capacity; } MemoryPool; void pool_init(MemoryPool *pool, size_t initial_capacity) { pool->ptrs = malloc(initial_capacity * sizeof(void*)); pool->count = 0; pool->capacity = initial_capacity; } void* pool_alloc(MemoryPool *pool, size_t size) { if(pool->count >= pool->capacity) { pool->capacity *= 2; pool->ptrs = realloc(pool->ptrs, pool->capacity * sizeof(void*)); } void *ptr = malloc(size); pool->ptrs[pool->count++] = ptr; return ptr; } void pool_free(MemoryPool *pool) { for(size_t i = 0; i < pool->count; i++) { free(pool->ptrs[i]); } free(pool->ptrs); pool->count = 0; pool->capacity = 0; }

5.2 流式响应处理

对于长文本生成，可以使用流式API：

size_t stream_callback(char *ptr, size_t size, size_t nmemb, void *userdata) { size_t real_size = size * nmemb; json_error_t error; json_t *root = json_loadb(ptr, real_size, 0, &error); if(root) { json_t *choices = json_object_get(root, "choices"); if(json_is_array(choices) && json_array_size(choices) > 0) { json_t *choice = json_array_get(choices, 0); json_t *delta = json_object_get(choice, "delta"); json_t *content = json_object_get(delta, "content"); if(content && json_is_string(content)) { printf("%s", json_string_value(content)); fflush(stdout); } } json_decref(root); } return real_size; } void stream_chat(const char* prompt) { CURL *curl = curl_easy_init(); if(curl) { char *request = build_chat_request(prompt); struct curl_slist *headers = NULL; headers = curl_slist_append(headers, "Content-Type: application/json"); headers = curl_slist_append(headers, "Accept: text/event-stream"); curl_easy_setopt(curl, CURLOPT_URL, API_URL); curl_easy_setopt(curl, CURLOPT_POSTFIELDS, request); curl_easy_setopt(curl, CURLOPT_HTTPHEADER, headers); curl_easy_setopt(curl, CURLOPT_WRITEFUNCTION, stream_callback); CURLcode res = curl_easy_perform(curl); if(res != CURLE_OK) { fprintf(stderr, "curl_easy_perform() failed: %s\n", curl_easy_strerror(res)); } free(request); curl_slist_free_all(headers); curl_easy_cleanup(curl); } }

5.3 错误处理与重试机制

#define MAX_RETRIES 3 #define RETRY_DELAY 1000 // 毫秒 char* robust_http_post(const char* url, const char* post_data, const char* api_key) { int retry_count = 0; CURLcode res; struct string s; while(retry_count < MAX_RETRIES) { init_string(&s); CURL *curl = curl_easy_init(); if(curl) { // ... 设置headers和其他选项 ... res = curl_easy_perform(curl); if(res == CURLE_OK) { curl_easy_cleanup(curl); return s.ptr; } fprintf(stderr, "Attempt %d failed: %s\n", retry_count+1, curl_easy_strerror(res)); curl_easy_cleanup(curl); free(s.ptr); } retry_count++; if(retry_count < MAX_RETRIES) { usleep(RETRY_DELAY * 1000); } } return NULL; }

6. 实际应用示例

6.1 嵌入式医疗设备集成

// 医疗问答系统集成示例 void medical_qa_loop() { printf("医疗问答系统已启动(输入quit退出)\n"); char input[1024]; while(1) { printf("\n患者症状描述: "); if(fgets(input, sizeof(input), stdin) == NULL) break; // 移除换行符 input[strcspn(input, "\n")] = '\0'; if(strcmp(input, "quit") == 0) break; char *request = build_chat_request(input); char *response = robust_http_post(API_URL, request, API_KEY); if(response) { char *content = parse_chat_response(response); if(content) { printf("\n诊断建议:\n%s\n", content); free(content); } free(response); } free(request); } }

6.2 性能基准测试

#include <time.h> void benchmark_api_call(int iterations) { struct timespec start, end; double total_time = 0; int success_count = 0; const char *test_prompt = "简要说明感冒和流感的区别"; for(int i = 0; i < iterations; i++) { clock_gettime(CLOCK_MONOTONIC, &start); char *request = build_chat_request(test_prompt); char *response = http_post(API_URL, request, API_KEY); clock_gettime(CLOCK_MONOTONIC, &end); if(response) { double elapsed = (end.tv_sec - start.tv_sec) + (end.tv_nsec - start.tv_nsec) / 1e9; total_time += elapsed; success_count++; free(response); } free(request); } printf("完成%d次调用，成功%d次\n", iterations, success_count); printf("平均响应时间: %.3f秒\n", total_time / success_count); printf("QPS: %.2f\n", success_count / total_time); }