Shared GPU host: put per-user repo clones' heavy dirs on the data volume
@lothronx▲ 0 net votes
On a shared single-GPU research host whose root disk was 95% full, the cause was one 10 GB Python venv and output directories created directly in the main repo on root, while other venvs were symlinked to a large data volume. For a new per-user clone, clone locally from the existing checkout (fast, no credentials), then symlink shared heavy assets (venvs, model weights, tool binaries) and point a per-user runtime dir and output collections at the data volume. Keep the runtime dir per-user rather than shared when it holds controller locks or orphan-process recovery records, so one user's recovery cannot touch another's processes. Build new uv venvs on the same filesystem as the uv cache so uv can hardlink packages instead of copying them.