openmoji-g1-gpu-e7dc920-0bafd5c-9b9b1699

Written result

Status: failed during the first CUDA step; no checkpoint written.

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-cd3250e-0bafd5c-9b9b1699, which verified its stage but failed before model construction because the staged pilot resolved relative input paths against the image working directory. The only changed factor is the wrapper's working directory; the config, fixture, seed, step budget, image, and cap are unchanged.

The working-directory correction worked: the staged config, pinned palette, six raw SVGs, selected normalizers, conditioned model, and optimizer all loaded from the immutable source root. The run then failed inside the first CUDA step because the pipeline declares torch.use_deterministic_algorithms(True) and CUDA >= 10.2 cuBLAS needs CUBLAS_WORKSPACE_CONFIG to honor that declaration. The container left an empty artifact directory, no checkpoint, and no running container.

This is a launcher environment gap, not a weakening of the determinism contract. The correction sets CUBLAS_WORKSPACE_CONFIG=:4096:8 explicitly in the owned Docker smoke launcher, which makes the declared determinism achievable rather than relaxing it.