DeepSeekMath API参考：完整接口文档与使用示例

DeepSeekMath是一个专注于数学推理的大语言模型，在MATH基准测试中达到了51.7%的准确率。本文档提供完整的API接口参考和使用示例，帮助开发者快速集成和使用DeepSeekMath模型。## 模型版本DeepSeekMath提供三个版本的模型：| 模型名称 | 描述 | 适用场景 ||---------|------|---------|| `deepseek-ai/...

gitblog_00020

1118人浏览 · 2025-09-02 04:50:29

gitblog_00020 · 2025-09-02 04:50:29 发布

DeepSeekMath API参考：完整接口文档与使用示例

【免费下载链接】DeepSeek-Math 项目地址: https://gitcode.com/GitHub_Trending/de/DeepSeek-Math

概述

DeepSeekMath是一个专注于数学推理的大语言模型，在MATH基准测试中达到了51.7%的准确率。本文档提供完整的API接口参考和使用示例，帮助开发者快速集成和使用DeepSeekMath模型。

模型版本

DeepSeekMath提供三个版本的模型：

模型名称	描述	适用场景
`deepseek-ai/deepseek-math-7b-base`	基础版本	文本补全、数学推理
`deepseek-ai/deepseek-math-7b-instruct`	指令调优版本	对话式数学问题解答
`deepseek-ai/deepseek-math-7b-rl`	强化学习版本	高级数学问题求解

核心API接口

1. Hugging Face Transformers接口

文本补全接口（Base模型）

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, GenerationConfig

def deepseek_math_base_completion(prompt, max_new_tokens=100, temperature=1.0, top_p=0.9, top_k=50):
    """
    DeepSeekMath Base模型文本补全接口
    
    参数:
    prompt (str): 输入文本提示
    max_new_tokens (int): 最大生成token数
    temperature (float): 温度参数，控制生成随机性
    top_p (float): 核采样参数
    top_k (int): Top-K采样参数
    
    返回:
    str: 生成的文本结果
    """
    model_name = "deepseek-ai/deepseek-math-7b-base"
    tokenizer = AutoTokenizer.from_pretrained(model_name)
    model = AutoModelForCausalLM.from_pretrained(
        model_name, 
        torch_dtype=torch.bfloat16, 
        device_map="auto"
    )
    model.generation_config = GenerationConfig.from_pretrained(model_name)
    model.generation_config.pad_token_id = model.generation_config.eos_token_id

    inputs = tokenizer(prompt, return_tensors="pt")
    outputs = model.generate(
        **inputs.to(model.device),
        max_new_tokens=max_new_tokens,
        temperature=temperature,
        top_p=top_p,
        top_k=top_k,
        do_sample=True
    )

    result = tokenizer.decode(outputs[0], skip_special_tokens=True)
    return result

# 使用示例
prompt = "The integral of x^2 from 0 to 2 is"
result = deepseek_math_base_completion(prompt)
print(result)

对话补全接口（Instruct模型）

def deepseek_math_instruct_chat(question, max_new_tokens=200, temperature=0.7, top_p=0.9, top_k=50):
    """
    DeepSeekMath Instruct模型对话接口
    
    参数:
    question (str): 数学问题，建议使用chain-of-thought提示
    max_new_tokens (int): 最大生成token数
    temperature (float): 温度参数
    top_p (float): 核采样参数
    top_k (int): Top-K采样参数
    
    返回:
    str: 生成的推理过程和答案
    """
    model_name = "deepseek-ai/deepseek-math-7b-instruct"
    tokenizer = AutoTokenizer.from_pretrained(model_name)
    model = AutoModelForCausalLM.from_pretrained(
        model_name,
        torch_dtype=torch.bfloat16,
        device_map="auto"
    )
    model.generation_config = GenerationConfig.from_pretrained(model_name)
    model.generation_config.pad_token_id = model.generation_config.eos_token_id

    # 构建对话消息（推荐使用chain-of-thought提示）
    messages = [
        {
            "role": "user", 
            "content": f"{question}\nPlease reason step by step, and put your final answer within \\boxed{{}}."
        }
    ]
    
    input_tensor = tokenizer.apply_chat_template(
        messages, 
        add_generation_prompt=True, 
        return_tensors="pt"
    )
    
    outputs = model.generate(
        input_tensor.to(model.device),
        max_new_tokens=max_new_tokens,
        temperature=temperature,
        top_p=top_p,
        top_k=top_k,
        do_sample=True
    )

    result = tokenizer.decode(outputs[0][input_tensor.shape[1]:], skip_special_tokens=True)
    return result

# 使用示例
question = "what is the integral of x^2 from 0 to 2?"
result = deepseek_math_instruct_chat(question)
print(result)

2. Replicate云端API接口

Base模型API端点

import replicate
import os

# 设置Replicate API token
os.environ["REPLICATE_API_TOKEN"] = "your_replicate_api_token"

def replicate_math_base(prompt, max_new_tokens=100, temperature=1.0, top_p=0.9, top_k=50):
    """
    Replicate平台DeepSeekMath Base模型API
    
    参数:
    prompt (str): 输入文本提示
    max_new_tokens (int): 最大生成token数
    temperature (float): 温度参数
    top_p (float): 核采样参数
    top_k (int): Top-K采样参数
    
    返回:
    str: 流式生成的文本结果
    """
    output = replicate.run(
        "cjwbw/deepseek-math-7b-base",
        input={
            "text": prompt,
            "max_new_tokens": max_new_tokens,
            "temperature": temperature,
            "top_p": top_p,
            "top_k": top_k
        }
    )
    
    # 流式输出处理
    full_output = ""
    for item in output:
        full_output += item
        print(item, end="", flush=True)
    
    return full_output

# 使用示例
result = replicate_math_base("The derivative of sin(x) is")

Instruct模型API端点

def replicate_math_instruct(question, max_new_tokens=200, temperature=0.7, top_p=0.9, top_k=50):
    """
    Replicate平台DeepSeekMath Instruct模型API
    
    参数:
    question (str): 数学问题
    max_new_tokens (int): 最大生成token数
    temperature (float): 温度参数
    top_p (float): 核采样参数
    top_k (int): Top-K采样参数
    
    返回:
    str: 流式生成的推理结果
    """
    output = replicate.run(
        "cjwbw/deepseek-math-7b-instruct",
        input={
            "text": f"{question}\nPlease reason step by step, and put your final answer within \\boxed{{}}.",
            "max_new_tokens": max_new_tokens,
            "temperature": temperature,
            "top_p": top_p,
            "top_k": top_k
        }
    )
    
    full_output = ""
    for item in output:
        full_output += item
        print(item, end="", flush=True)
    
    return full_output

# 使用示例
result = replicate_math_instruct("Solve the equation: 2x + 5 = 13")

参数详解

生成参数

参数	类型	默认值	描述	推荐范围
`max_new_tokens`	int	100-200	最大生成token数量	50-500
`temperature`	float	0.7-1.0	控制生成随机性	0.1-1.5
`top_p`	float	0.9	核采样参数	0.5-1.0
`top_k`	int	50	Top-K采样参数	10-100

提示工程最佳实践

英文数学问题提示模板

# 推荐使用chain-of-thought提示
prompt = f"{question}\nPlease reason step by step, and put your final answer within \\boxed{{}}."

中文数学问题提示模板

prompt = f"{question}\n请通过逐步推理来解答问题，并把最终答案放置于\\boxed{{}}中。"

使用示例集合

示例1：微积分问题

# 计算定积分
question = "Calculate the definite integral of x^3 from 0 to 1"
result = deepseek_math_instruct_chat(question)
print(result)

示例2：代数方程求解

# 解二次方程
question = "Solve the quadratic equation: x^2 - 5x + 6 = 0"
result = deepseek_math_instruct_chat(question)
print(result)

示例3：几何问题

# 计算圆的面积
question = "What is the area of a circle with radius 5?"
result = deepseek_math_instruct_chat(question)
print(result)

示例4：概率统计

# 概率计算
question = "If you flip a fair coin 3 times, what is the probability of getting exactly 2 heads?"
result = deepseek_math_instruct_chat(question)
print(result)

错误处理与最佳实践

错误处理

import requests
from transformers import AutoTokenizer, AutoModelForCausalLM
from huggingface_hub import HfApi, ModelFilter

def safe_model_loading(model_name, max_retries=3):
    """
    安全的模型加载函数，包含重试机制
    """
    for attempt in range(max_retries):
        try:
            tokenizer = AutoTokenizer.from_pretrained(model_name)
            model = AutoModelForCausalLM.from_pretrained(
                model_name,
                torch_dtype=torch.bfloat16,
                device_map="auto"
            )
            return tokenizer, model
        except Exception as e:
            print(f"Attempt {attempt + 1} failed: {e}")
            if attempt == max_retries - 1:
                raise
            time.sleep(2 ** attempt)  # 指数退避

# 使用安全加载
try:
    tokenizer, model = safe_model_loading("deepseek-ai/deepseek-math-7b-base")
except Exception as e:
    print(f"Failed to load model: {e}")

性能优化建议

# 使用批处理提高效率
def batch_predict(questions, model, tokenizer, batch_size=4):
    """
    批量预测函数
    """
    results = []
    for i in range(0, len(questions), batch_size):
        batch = questions[i:i+batch_size]
        inputs = tokenizer(batch, return_tensors="pt", padding=True, truncation=True)
        
        with torch.no_grad():
            outputs = model.generate(
                **inputs.to(model.device),
                max_new_tokens=150,
                temperature=0.7,
                do_sample=True
            )
        
        batch_results = tokenizer.batch_decode(outputs, skip_special_tokens=True)
        results.extend(batch_results)
    
    return results

API响应格式

成功响应示例

{
    "status": "success",
    "result": "The integral of x^2 from 0 to 2 is ∫x²dx from 0 to 2 = (1/3)x³|0 to 2 = (1/3)*8 - 0 = 8/3 ≈ 2.6667",
    "model": "deepseek-math-7b-instruct",
    "inference_time": 2.45
}

错误响应示例

{
    "status": "error",
    "error_code": "MODEL_LOAD_FAILED",
    "message": "Failed to load model: Connection timeout",
    "suggestion": "Please check your internet connection and try again"
}