owned-gpu-smoke-5da02d3-7c03a644
completed —
Hypothesis
Explicitly bypassing the NGC entrypoint will preserve clean JSON stdout while the already-proven GPU and durable artifact smoke remain successful.
Why run it
Confirms the complete owned-worker adapter contract, including machine-readable result recovery, before any larger training.
Written result
Status: completed.
This retry changed only entrypoint handling after the parent run completed GPU and artifact work but failed clean-JSON result parsing. The adapter staged the immutable Git snapshot, ran the typed-codec/render and one-step GPU forward/backward smoke, wrote and reloaded the checkpoint, and copied selected artifacts into the persistent artifact volume.
The RTX 4080 step used PyTorch 2.8.0a0+5228986c39.nv25.06 with CUDA 12.9. Loss was 0.20076081156730652 and gradient norm was 0.3963698446750641. Tensor fields were stable through the codec round trip. The 16,179-byte checkpoint has SHA-256 77e93414e04707efb0469718426bb2364d5a7b22e575e519124d5c628442348f.
A second complete invocation used create-or-identical semantics and reproduced both the workspace and durable artifacts exactly. No detached job remains active.
State transitions
- planned2026-09-20T09:19:27Z
- staged2026-09-20T09:20:43Z
- completed2026-09-20T09:21:22Z
Run record
Verbatim from runs/owned-gpu-smoke-5da02d3-7c03a644/run.yaml, the record committed before launch.
e3b0c44298fc1c14…7c03a6440fe64d2a…