[ROCm][CI] Keep rocprofiler-sdk out of DeepEP HT MoE test workers (#51173)
Signed-off-by: Stefan Koncarevic <Stefan.Koncarevic@amd.com> Co-authored-by: Andreas Karatzas <akaratza@amd.com>
stefankoncarevic committed
Aug 5, 2026 at 22:28 UTC
65addac701c0f019152cc00ffedb6edec442a3e2
1 file changed
+7
tests/kernels/moe/test_deepep_moe.py
+7
@@ -488,10 +488,17 @@ def test_deep_ep_moe(
488
world_dp_size: tuple[int, int],
489
per_act_token_quant: bool,
490
workspace_init,
491
+ monkeypatch,
492
):
493
low_latency_mode = False
494
use_fp8_dispatch = False
495
496
+ if current_platform.is_rocm():
497
+ # The cooperative kernel launches on this path segfault inside the ROCm
498
+ # HSA runtime at process exit while rocprofiler-sdk is attached, and
499
+ # torch attaches it implicitly. Workers are spawned, so they inherit it.
500
+ monkeypatch.setenv("ROCPROFILER_REGISTER_ENABLED", "0")
501
+
502
set_random_seed(7)
503
world_size, dp_size = world_dp_size
504
config = TestConfig(dtype=dtype, topk=topk, m=m, k=k, n=n, num_experts=num_experts)