test_mla_helix: Skip cp_size > 1 tests that require helix alltoall

The alltoall_helix C++ operation requires NCCL process groups to be
fully initialized via init_pg, which depends on Ray infrastructure
(TorchDist). The MPI-based test environment cannot properly set up
these NCCL groups.

For now, skip tests with cp_size > 1 which require helix communication.
Only pure TP configurations (cp_size=1) will run in the MPI environment.

Author: brb-nv

tests/unittest/_torch/modules/test_mla_helix.py

@@ -1143,6 +1143,13 @@ def test_mla_helix_distributed_mixed_tp_cp(
     # Validate that num_heads is divisible by (tp_size * cp_size)
     if scenario.num_heads % (tp_size * cp_size) != 0:
         pytest.skip(f"num_heads {scenario.num_heads} not divisible by tp_size*cp_size {tp_size * cp_size}")
+
+    # Skip helix tests (cp_size > 1) in MPI-based test environment
+    # The alltoall_helix operation requires NCCL process groups to be fully
+    # initialized via init_pg, which needs Ray infrastructure (TorchDist).
+    # For now, only test pure TP configurations (cp_size=1) in the MPI environment.
+    if cp_size > 1:
+        pytest.skip(f"cp_size={cp_size} requires helix alltoall which needs Ray/TorchDist for NCCL init")
 
     gen_steps = scenario.ref_steps if gen_steps is None else gen_steps