[GG] fix(moe): serialize hybrid resident-grid launches

tests/quantization/test_nvfp4_nf3_hybrid.py

@@ -8,6 +8,7 @@
 from vllm.model_executor.layers.quantization import get_quantization_config
 from vllm.model_executor.layers.quantization.nvfp4_nf3_hybrid import (
     NvFp4Nf3HybridConfig,
+    NvFp4Nf3HybridMoEMethod,
     _combined_tier_local_descriptors,
     _read_hybrid_keys,
     _unpack_nf3_codes,
@@ -109,3 +110,10 @@ def test_grid188_tier_descriptors_encode_exact_partition():
 def test_grid188_tier_descriptors_reject_incomplete_partition():
     with pytest.raises(ValueError, match="does not cover all 256"):
         _combined_tier_local_descriptors({0: (0, 0)})
+
+
+@pytest.mark.parametrize("num_tokens", [1, 8, 256, 3072])
+def test_hybrid_moe_rejects_shared_expert_aux_stream(num_tokens):
+    method = object.__new__(NvFp4Nf3HybridMoEMethod)
+
+    assert not method.supports_shared_experts_aux_stream(num_tokens)

vllm/model_executor/layers/quantization/nvfp4_nf3_hybrid.py

@@ -333,6 +333,16 @@ def __init__(
         super().__init__(moe_config)
         self.quant_config = quant_config
 
+    def supports_shared_experts_aux_stream(self, num_tokens: int) -> bool:
+        """The hybrid B12X launches must own the device while they run.
+
+        Both the unified Grid188 decode and the per-tier fused fallback use
+        resident-grid software barriers. Concurrent shared-expert work can
+        occupy an SM needed by a barrier participant and deadlock the grid;
+        the resulting asynchronous fault may surface in a later CUDA op.
+        """
+        return False
+
     def maybe_make_prepare_finalize(
         self,
         routing_tables: tuple[torch.Tensor, torch.Tensor, torch.Tensor] | None = None,