vast-experiment-opslisted

# vast-experiment-ops Error-informed operational playbook for renting, running, and reporting ML experiments on Vast.ai. It supplements the platform docs and the instance's own guide; it replaces neither. ## Quick reference | Situation | Action | |---|---| | Strict FP32 equivalence passes on CPU, fails on CUDA | Re-run the unchanged check in a fresh process with `NVIDIA_TF32_OVERRIDE=0` set before Python starts | | `JSONDecodeError` parsing remote metrics | Capture stdout/stderr separately, parse stdout only; collector emits exactly one JSON document on stdout | | Asked whether the test split was used | Answer from this run's queue config and result metadata plus prior artifacts — never open test data to answer | | Reporting progress | Include UTC time, snapshot age, active trial/epoch, and complete/pruned/failed/running counts | | Sizing a job | Check cgroup CPU/RAM limits, free GPU memory, free disk, and storage persistence — host totals mislead | ## Ground rules - Get the instance address and experiment paths from the user or project config. Never reuse an address remembered from this skill. - Read the instance's own guide (e.g. `/etc/vast-agents-guide.md`) before operating on it; use its Python environment and service manager. - Treat a status request as read-only: do not restart services, change numerical settings, or read held-out data just to report progress. ## Case 1 — CPU passes, CUDA fails strict recurrent equivalence **Observed:** step-versus-full-rep