[AMDGPU] Skip GJK collision edge-case tests under fast-math - #3067
[AMDGPU] Skip GJK collision edge-case tests under fast-math#3067gayathiri-venkataraman wants to merge 1 commit into
Conversation
Some collision_edge_cases modes fail on AMDGPU with fast-math enabled due to tight FP32 tolerances; they pass with fast-math off. Skip the known cases so the rest of the rigid test suite can run on AMD GPUs. Ported from amd-integration #33 (2b519a8) after test suite restructure (Genesis-Embodied-AI#3037). Co-authored-by: Cursor <cursoragent@cursor.com>
|
Could you share more context? What is happening exactly? Any video? We cannot just disable tests like this. Otherwise it means AMDGPU is not supported basically. |
|
Converted to draft — GPU verification was not run before opening; running now on MI300X (job 67545433). Target: Will mark ready for review once we have a green run (expect 5 skips on AMDGPU for known fast-math tolerance cases). |
|
MI300X verification update (job 67545433): did not complete — Slurm 1h TIMEOUT at 24/36 cases. All cases so far ERROR at fixture/setup (no skips observed yet); likely env mismatch (Genesis 1.2.2 venv vs upstream-main test layout on Log: |
|
Closing for inactivity. Let's reopen if some progress is made. |
2 participants
Summary
Skip selected
collision_edge_casesmodes on AMDGPU when fast-math is enabled. These cases fail due to tight FP32 tolerances but pass with fast-math off; skipping them unblocks running the rest of the rigid collision test suite on AMD GPUs.Ported to the restructured test layout (
tests/rigid/test_collision.py::test_edge_cases) after upstream #3037.Correctness
Test-only change — no physics or solver behavior change.
Verification
Matches amd-integration #33 (
2b519a8c). To verify on AMDGPU:Skipped modes:
(gjk_collision, mode)in{(True,5), (True,7), (True,8), (False,5), (False,7)}whengs.backend == gs.amdgpu.Made with Cursor