JupyterHub with PyTorch on the BC-250, running on k3s
A tutorial repo sets up JupyterHub notebooks on the BC-250, with PyTorch using the GPU through OpenCL.
The BC-250 is an ex-mining board with a cut-down PS5 APU, and a repository by mxreyer details how to run JupyterHub on a single-node Kubernetes cluster on the board. PyTorch connects to the GPU via PyTorch through OpenCL.
Each user gets a pod with dedicated storage and GPU access routed through a device plugin exposing /dev/dri/renderD128. LAN access uses Traefik, while tailnet access runs through a Tailscale sidecar. The software stack consists of Fedora 44, amdgpu, and Mesa rusticl on radeonsi.
Work in a separate repository, pytorch-dlprim-gfx1013, yielded 2.54× on training and 1.59× on inference. The documentation notes that this project is intended for learning rather than production use.
