Profile MoE TP-local shards without collectives

runs/frontier-qwen30-vllm020-profile-v1/profile_vllm020_moe.py

@@ -199,8 +199,12 @@ def main() -> None:
             in_dtype=torch.bfloat16,
             max_num_tokens=max_num_tokens,
         )
+        # This process profiles one TP-local weight shard.  Keep the global
+        # runtime context single-rank so vLLM does not initialize a collective;
+        # the action-conditioned shard size remains explicit in moe_config and
+        # the real TP2/TP4 all-reduce is profiled in a separate multi-GPU run.
         vllm_config = VllmConfig(
-            parallel_config=ParallelConfig(tensor_parallel_size=tp)
+            parallel_config=ParallelConfig(tensor_parallel_size=1)
         )
         with set_current_vllm_config(vllm_config):
             backend, experts_cls = select_unquantized_moe_backend(moe_config)
@@ -348,8 +352,8 @@ def main() -> None:
             "norm_topk_prob": True,
         },
         "measurement_scope": (
-            "vLLM modular MoE prepare+FlashInfer CUTLASS experts+finalize; "
-            "router linear/top-k and TP all-reduce excluded"
+            "one TP-local weight shard: vLLM modular MoE prepare+FlashInfer "
+            "CUTLASS experts+finalize; router linear/top-k and TP all-reduce excluded"
         ),
         "rows": rows,
     }