A Free Alternative to GitHub Copilot

For Visual Studio Code, the usual option to consider is the FauxPilot plugin paired with a FauxPilot server, but in practice the FauxPilot plugin is not especially usable once installed.

A better solution is to install the FauxPilot server yourself, then install the GitLab plugin and modify the address of its code assistant, so that you can do AI-assisted coding.

To avoid a needlessly painful environment setup, you can install Baidu Paddle's PaddleNLP and start the code-assist service directly with a single line of code. It is still FauxPilot under the hood, but installation is convenient: you can simply run it in the background, and it supports models with 350 million, 6 billion, and 16 billion parameters. The 350M model needs about 3.5 GB of VRAM, the 6B model needs 13 GB, and the VRAM requirement for the 16B model has not been computed.

If you have the means, you can consider using DeepSpeed with ZeRO: generally an 8 GB GPU should still manage the 6B model, and if you have a 3090, loading the 16B model is no problem either.