openmoji-g1-gpu-215bcb8-0bafd5c-9b9b1699

Written result

Status: failed on the adapter result contract after the GPU step completed.

This run is bounded to the exact one-step local pilot and stages only its six selected raw SVGs plus the pinned palette with the committed source snapshot.

It retries openmoji-g1-gpu-e7dc920-0bafd5c-9b9b1699, which loaded the staged config, data, normalizers, model, and optimizer correctly but was rejected inside the first CUDA step because torch.use_deterministic_algorithms(True) requires CUBLAS_WORKSPACE_CONFIG on CUDA >= 10.2. The only changed factor is that launcher environment variable. The config, fixture, seed, step budget, image, and cap are unchanged.

The deterministic-cuBLAS correction worked. The pipeline ran to completion on the RTX 4080: device is cuda, deterministic_algorithms is true, one optimizer step produced train loss 15.216644 and token accuracy 0.005068, validation loss is 15.344566 with 0.009404 aggregate / 0.008850 changed / 0.009709 retained accuracy over 226 changed and 412 retained fields, checkpoint_round_trip is true, and locked_path_exact is true. Near-random first-step accuracy is expected and is not learning evidence.

The run is still recorded as failed because the adapter's result contract failed: the staged wrapper computed its summary digest from the output root rather than the pilot's report root, so it raised before printing the JSON result. The GPU work and its artifacts are intact and durable on the worker and are listed in artifacts.json; nothing was deleted. The correction is confined to the wrapper's summary path and requires a new code and run identity.

Recovered evidence on the worker:

• checkpoint.zip, 7,075,309 bytes, sha256 4b265e5575e3aa455a0d427e340ec407eaaaf39222709305d060281ae7e453f9;
• summary.json sha256 15ede088423678eb308548481e1c348850c04bf588101d4f74fecaaa224d07b7;
• 577,552 model parameters over the 3,359-icon bucket-p32-t128;
• torch 2.8.0a0+5228986c39.nv25.06, CUDA 12.9.

The local CPU pilot checkpoint is the same 7,075,309 bytes but hashes differently (d11efa00...), which is the expected CPU/GPU floating-point difference. This run was never intended to establish cross-device artifact identity.