Le Rong Blog

Free sharing lab

Deepseek-Coder HuggingFace

DeepSeek-Coder is a family of state-of-the-art code-focused language models developed by DeepSeek AI, available on Hugging Face. These models are optimized for code generation, understanding, and editing tasks and support a wide range of programming languages. Below is a guide to using DeepSeek-Coder via Hugging Face:

1. Model Access on Hugging Face

The models are hosted on Hugging Face Hub under the deepseek-ai namespace. Example models include:

  • deepseek-ai/deepseek-coder-1.3b
  • deepseek-ai/deepseek-coder-6.7b
  • deepseek-ai/deepseek-coder-33b

Repository:

2. Key Features

  • Multi-Language Support: Python, Java, C++, JavaScript, Go, Rust, and more.
  • Long Context Handling: Up to 16K token context for large codebases.
  • Code Generation: Generate code snippets, scripts, or entire functions from natural language prompts.
  • Code Completion: Autocomplete code in real time.
  • Code Understanding: Explain code logic, debug errors, or refactor existing code.

3. How to Use

Step 1: Install Dependencies

pip install transformers torch

Step 2: Load the Model

from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "deepseek-ai/deepseek-coder-6.7b-instruct"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    trust_remote_code=True,
    torch_dtype="auto"
).cuda()  # Use GPU if available

Step 3: Generate Code

prompt = "Write a Python function to calculate the Fibonacci sequence."

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256)
generated_code = tokenizer.decode(outputs[0], skip_special_tokens=True)

print(generated_code)

4. Advanced Usage

Code Completion

Use the model to autocomplete partial code:

partial_code = "def factorial(n):\n    if n == 0:\n        return 1\n    else:"
inputs = tokenizer(partial_code, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=100)

Chat Format (for Instruct Models)

For instruction-tuned models (e.g., deepseek-coder-6.7b-instruct):

messages = [
    {"role": "user", "content": "Write a quicksort algorithm in Python."}
]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to(model.device)
outputs = model.generate(inputs, max_new_tokens=512)

5. Quantization (Reduce GPU Memory)

Use bitsandbytes for 4-bit/8-bit quantization:

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    device_map="auto",
    load_in_4bit=True,  # or load_in_8bit=True
    trust_remote_code=True
)

6. Applications

  • IDE Plugins: Power autocomplete features in tools like VS Code.
  • Code Review Automation: Identify bugs or security issues.
  • Documentation Generation: Create comments or docs from code.
  • Educational Tools: Teach programming interactively.

7. Performance

  • Achieves top-tier results on benchmarks like HumanEval and MBPP.
  • Outperforms similar-sized models (e.g., CodeLlama, StarCoder).

8. License

  • Most DeepSeek-Coder models are licensed under Apache 2.0, allowing commercial and research use. Check the specific model card for details.

For full documentation and examples, visit the official Hugging Face repository:

Tags：
Categories：