From 7245460494cfedfdd84dcd7480a353c3cf4c9f57 Mon Sep 17 00:00:00 2001 From: Alex Magro Date: Tue, 28 Apr 2026 22:24:26 +0000 Subject: [PATCH] Disable all UB layer tests for gfx942 --- tests/pytorch/distributed/test_comm_gemm_overlap.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/tests/pytorch/distributed/test_comm_gemm_overlap.py b/tests/pytorch/distributed/test_comm_gemm_overlap.py index 8d98c6263..68671447b 100644 --- a/tests/pytorch/distributed/test_comm_gemm_overlap.py +++ b/tests/pytorch/distributed/test_comm_gemm_overlap.py @@ -105,9 +105,9 @@ def _run_layer_with_overlap( pytest.skip("Bulk overlap is not yet supported on HIP/ROCm.") # On gfx942, non-determinism across the 8 XCDs causes small jitter that compounds # This should not affect training convergence, but creates larger numerical differences. + # TODO: Fix gfx942 issues arising from deterministic bwd attention and other jitter if (IS_HIP_EXTENSION - and get_device_compute_capability() < (9, 5) - and layer_type == te.TransformerLayer.__name__): + and get_device_compute_capability() < (9, 5)): pytest.skip("TransformerLayer overlap can exceed numerical tolerance on pre-MI350 due to jitter.") test_path = TEST_ROOT / "run_layer_with_overlap.py" test_cmd = LAUNCH_CMD + [