gpyou

Command-line client for GPYou: prepaid GPUs for fine-tuning and serving open models.

uv tool install gpyou     # or: pipx install gpyou
gpyou signup              # or: gpyou login
gpyou                     # interactive menu

The sign-in session is kept in the OS keychain. On systems without one (headless Linux, WSL) it is kept in ~/.config/gpyou/sessions.json, readable only by you; gpyou logout ends it.

Any public Hugging Face model that fits one GPU can be fine-tuned and served (--model Qwen/Qwen3-8B). Qwen/Qwen2.5-1.5B-Instruct and Qwen/Qwen2.5-7B-Instruct are pinned and train on every assistant turn; any other model trains with its own chat template on the final assistant turn. Fine-tuning uses one fixed recipe: 1 epoch, learning rate 2e-4, LoRA r 16 / alpha 32 on all linear layers, 8 conversations per step per GPU, conversations up to 2,048 tokens (longer ones are rejected, not truncated), and loss on every assistant turn only. gpyou finetune validate prints it with your dataset.

The CLI talks to https://app.gpyou.ai. Set GPYOU_API_URL to use another GPYou service, such as a development server.