GitHub Copilot, advertised by GitHub as “Your AI pair programmer”, is intended to support programmers in their work. The AI software was trained to generate code from natural language using tons of code hosted on GitHub. The system is based on OpenAI Codex. Just like the competition Tabnine, the Copilot is available as a paid subscription for integration into common IDEs.
With FauxPilot there is now an alternative that can be installed locally and used in Visual Studio Code. It uses the codegen models developed by SalesForce in a research project for code generation. These models, which come in four different sizes with different storage requirements, generate program code from natural language. Nvidia’s Triton Inference Server with the FasterTransformer backend efficiently transfers the models to Nvidia GPUs that are required to run FauxPilot.
In addition to an Nvidia GPU with sufficient VRAM (depending on the Codegen model, 2 to 32 GB), you need Docker, since FauxPilot runs in a Docker container. After the start, FauxPilot can be addressed from your own programs via the URL; a subset of the OpenAI API is implemented here. More interesting, however, is the option to enter your own FauxPilot server in the configuration of the Visual Studio Code Copilot plugin and thus use Copilot from the development environment instead of the GitHub.