openmoji-g1-gpu-7ba1aa4-0bafd5c-9b9b1699

Written result

Status: 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-215bcb8-0bafd5c-9b9b1699, whose GPU step, checkpoint round trip, and locked-path check all succeeded and whose artifacts remain durable on the worker. That run failed only because the staged wrapper digested summary.json from the output root rather than the pilot's report root. The only changed factor is that path, now derived from the resolved pilot configuration, plus a new CPU regression test that pins the wrapper to the layout the pilot actually writes.

Because nothing that affects computation changed, this run is also a same-worker reproduction check against the retry parent's recorded checkpoint and summary digests.

Result

The complete Gate G pipeline smoke passed on the owned RTX 4080. device is cuda, deterministic_algorithms is true under CUBLAS_WORKSPACE_CONFIG=:4096:8, the staged identity verified, one bounded optimizer step ran over the real dominant-bucket data, the canonical checkpoint round-tripped, and locked paths stayed byte-exact.

Recorded values:

• train loss 15.216644, train token accuracy 0.005068 at step 1;
• validation loss 15.344566; aggregate 0.009404, changed 0.008850 over 226 changed fields, retained 0.009709 over 412 retained fields;
• 577,552 model parameters; bucket-p32-t128 with 3,359 icons;
• checkpoint 7,075,309 bytes, sha256 4b265e5575e3aa455a0d427e340ec407eaaaf39222709305d060281ae7e453f9;
• summary sha256 15ede088423678eb308548481e1c348850c04bf588101d4f74fecaaa224d07b7;
• torch 2.8.0a0+5228986c39.nv25.06, CUDA 12.9, driver 595.71.05.

These first-step accuracies are near-random by construction. They are a pipeline liveness check, not learning evidence.

Reproduction

A second identical invocation returned the same JSON result, so the adapter's create-or-identical behavior holds. The checkpoint and summary digests also match the retry parent openmoji-g1-gpu-215bcb8-0bafd5c-9b9b1699 exactly, which makes this an independent same-worker reproduction across separate containers rather than a single observation.

The local CPU pilot checkpoint has the same 7,075,309 bytes but a different digest (d11efa00...). That is the expected CPU/GPU floating-point difference; cross-device artifact identity was never part of this run's contract.

Scope

Fixed-topology, geometry-only, one optimizer step. This does not establish learning, topology generation, or unconditional generation. It establishes that the selected representation, data join, conditioning, corruption, optimizer, checkpoint, and locked-edit path all execute correctly in the target GPU environment.