Repository navigation
Conversation
CPU tests (e.g., rocgdb-cpu) are architecture-independent but were running redundantly for every GPU family. Only assign the CPU test runner to the first family so they execute exactly once. Also switch CPU test runner from aws-linux-scale-rocm-prod to aws-linux-scale-rocm-medium since these tests don't need large machines. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Test that only the first family gets a CPU runner assignment and that subsequent families correctly have an empty test-runs-on-cpu, ensuring CPU tests execute exactly once across GPU families. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
❌ PR Check — Action Required
📖 Need help? See the Policy FAQ for details on every check and how to fix failures. |
|
🚫 Please fix the failed policies before requesting reviews. The following policy checks failed:
The |
| # CPU tests are architecture-independent, so only schedule them for the | ||
| # first family to avoid redundant runs across GPU targets. | ||
| if not per_family_info: | ||
| # First family: assign CPU runner | ||
| test_runs_on_cpu = get_cpu_test_runner(platform) |
There was a problem hiding this comment.
I'm not sure this is generally true. We should check across all ROCm subprojects.
Don't the tests only install artifacts for the requested GPU families? Consider a test that checks for kpack files for the current "gpu targets to test" - if we only run that on gfx942 we could miss that gfx1100 expects files to be produced but they aren't.
There was a problem hiding this comment.
We could create a separate test matrix that is just CPU, so we'd have:
- test CPU (fetch all gfx targets)
- test GPU (gfx942)
- test GPU (gfx950)
- test GPU (gfx1151)
(I'm convinced this de-duplication is worth doing... but we need to think a bit carefully about which components need what testing and how to plug that in)
|
Scott brings up a good point, so this can be closed as we want to keep testing for those archs |
since cpu tests are agnostic of arch, we run cpu tests only once
we also only run on medium runners too per comment here: #8768 (comment)