Deepseek Coder

See marcoscannabrava/local-llm for a more recent and robust multi-model implementation leveraging the excellent llm package.

This package uses Deepseek Coder 1.3B Instruct-tuned Language Model for code generation in the terminal.

A GPU is not required but makes the responses faster.

It's based on:

• Aider
• vLLM

Installation

pip install deepseek-coder

Usage

deepseek-coder

The first time you run deepseek-coder the model weights will be downloaded which can take a while.

tail -f /tmp/deepseek_coder.log # this will show the download progress

Options

deepseek-coder --logpath /path/to/logs

Backlog

• [] improve Aider default prompts and prevent multiple completions due to "improperly formatted response"
• [] env var for tokens
• [] multi-platform tests