The ROCm jobs in wheel_tests_nightly_release.yml run against two ROCm variants: the pinned TheRock 10.0.0 and the moving TheRock Next (pre-release) (therock-latest). Tracking the moving one is the point on nightlies. On a release run it is noise: those legs must test the ROCm that actually ships, and a therock-latest failure there muddies a release it has no bearing on.

Technical details

• Adds a conditional exclude entry to each of the three ROCm matrices (build-rocm-artifacts, run-pytest-rocm, run-bazel-test-rocm) dropping the therock-latest leg on release runs. The rocm: lists themselves are unchanged.


• A run counts as a release run when github.ref_name starts with release/ (upstream release branches) or rocm-jaxlib-v (downstream ROCm release branches), or when the artifact-type input is release.


• The prefixes are anchored deliberately to avoid matching other branches.


• Follows the existing idiom in this file — the &&-only expression that renders the matched value on a release branch and short-circuits to false (a no-op exclude) otherwise, as in the TPU libtpu-version-type excludes, combined with partial-object exclude against a dict-valued dimension as in run-pytest-cuda.

Test plan

Release-mode workflow_dispatch on ROCm/jax.

Test result

| artifact-type | build legs |
| --- | --- |
| release | 10 — all TheRock 10.0.0, zero pre-release |
| default | 20 — 10 pinned + 10 pre-release |