MPS Server Smoke Test Passed | Generated by AI
Server is up on MPS. Now the smoke test — a chat completion:
Smoke test passed — killing the server now.
nanochat runs fine on this laptop. Summary:
- Command:
uv run python -m scripts.chat_web --source base --port 8000(run from the repo root) - The existing
.venvworks viauv run, torch 2.9.1 with MPS (Apple GPU) acceleration was auto-detected - Only the base
d4checkpoint exists locally (no SFT/RL checkpoints), hence--source base— the default--source sftwould fail - Smoke test passed:
/healthreturned ready, and/chat/completionsstreamed tokens. Output is gibberish (“attention Place Francis species…”), which is expected — d4 is a tiny base model with no instruction tuning - Server killed afterward (exit code 144 is just the pkill, not an error)
For PyCharm debugging: set the run config to module scripts.chat_web with parameters --source base --port 8000, working directory the repo root, and the .venv interpreter. The CLI alternative is chat_cli.py (python -m scripts.chat_cli -i base) if you want a simpler single-process target to step through.