English

Local Deployment Guide for DeepSeek V3: From Basics to Advanced

Overview

This guide provides detailed instructions on deploying and running the DeepSeek V3 model in your local environment. We'll cover the complete process from basic setup to advanced deployment options, helping you choose the most suitable deployment strategy.

Environment Setup

Basic Requirements

  • NVIDIA GPU (A100 or H100 recommended) or AMD GPU
  • Sufficient system memory (32GB+ recommended)
  • Linux operating system (Ubuntu 20.04 or higher recommended)
  • Python 3.8 or higher

Code and Model Preparation

  1. Clone the official repository:
git clone https://github.com/deepseek-ai/DeepSeek-V3.git cd DeepSeek-V3/inference pip install -r requirements.txt
  1. Download model weights:
  • Download official model weights from HuggingFace
  • Place weight files in the designated directory

Deployment Options

1. DeepSeek-Infer Demo Deployment

This is the basic deployment method, suitable for quick testing and experimentation:

python convert.py --hf-ckpt-path /path/to/DeepSeek-V3 \ --save-path /path/to/DeepSeek-V3-Demo \ --n-experts 256 \ --model-parallel 16 torchrun --nnodes 2 --nproc-per-node 8 generate.py \ --node-rank $RANK \ --master-addr $ADDR \ --ckpt-path /path/to/DeepSeek-V3-Demo \ --config configs/config_671B.json \ --interactive \ --temperature 0.7 \ --max-new-tokens 200

2. SGLang Deployment (Recommended)

SGLang v0.4.1 offers optimal performance:

  • MLA optimization support
  • FP8 (W8A8) support
  • FP8 KV cache support
  • Torch Compile support
  • NVIDIA and AMD GPU support

3. LMDeploy Deployment (Recommended)

LMDeploy provides enterprise-grade deployment solutions:

  • Offline pipeline processing
  • Online service deployment
  • PyTorch workflow integration
  • Optimized inference performance

4. TRT-LLM Deployment (Recommended)

TensorRT-LLM features:

  • BF16 and INT4/INT8 weight support
  • Upcoming FP8 support
  • Optimized inference speed

5. vLLM Deployment (Recommended)

vLLM v0.6.6 features:

  • FP8 and BF16 mode support
  • NVIDIA and AMD GPU support
  • Pipeline parallelism capability
  • Multi-machine distributed deployment

Performance Optimization Tips

  1. Memory Optimization:

    • Use FP8 or INT8 quantization to reduce memory usage
    • Enable KV cache optimization
    • Set appropriate batch sizes

  2. Speed Optimization:

    • Enable Torch Compile
    • Use pipeline parallelism
    • Optimize input/output processing

  3. Stability Optimization:

    • Implement error handling mechanisms
    • Add monitoring and logging
    • Regular system resource checks

Common Issues and Solutions

  1. Memory Issues:

    • Reduce batch size
    • Use lower precision
    • Enable memory optimization options

  2. Performance Issues:

    • Check GPU utilization
    • Optimize model configuration
    • Adjust parallel strategies

  3. Deployment Errors:

    • Check environment dependencies
    • Verify model weights
    • Review detailed logs

Next Steps

After basic deployment, you can:

  • Conduct performance benchmarking
  • Optimize configuration parameters
  • Integrate with existing systems
  • Develop custom features

Now you have mastered the main methods for locally deploying DeepSeek V3. Choose the deployment option that best suits your needs and start building your AI applications!