A captured StableHLO training step compiles on all eight ranks, then stalls during execution with the original XLA flags in NVIDIA JAX Toolbox's November 16, 2026 nightly. Adding --xla_gpu_autotune_level=0 lets every rank complete.
The graph comes from a four-layer, 8,192-token synthetic Snowball training step with reference attention, scatter dispatch, and XLA ragged dot. It contains zero StableHLO custom calls. The runner imports Python standard-library modules, JAX, jaxlib, and NumPy and calls PJRT directly. It creates 203 input buffers filled with zeros. The test measures execution completion. Replay requires no model checkpoint, dataset, Marin installation, or storage credentials.
Use one Linux x86-64 host with eight NVIDIA H100 80 GB GPUs and NVIDIA Container Toolkit. The observed runs used driver 595.71.05, 32 CPUs, and a 384 GiB host-memory limit. This fixture is an eight-process replay, with one GPU per process.
The pinned image is `ghcr.io/nvidia/jax@sha256