fauxpilot

Why we included this project

FauxPilot runs a Copilot-style completion server on your own hardware, so your code never has to leave the machine. It serves Salesforce CodeGen models through NVIDIA's Triton Inference Server with the FasterTransformer backend, doing the inference on your GPU instead of in a cloud data center. A setup script downloads the model you choose, converts it for FasterTransformer, and the server then exposes the OpenAI API, a Copilot plugin endpoint, and a REST API, which means most existing editor clients can connect with minimal reconfiguration. You need Docker, an NVIDIA GPU with compute capability 6.0 or newer, and enough VRAM for the model size, though you can split a model across multiple GPUs. The documentation covers both server setup and client wiring, making this a practical option for teams that want private completions or want to try local code generation without a hosted subscription.