Profile collective dispatch across tensor sizes

runs/frontier-qwen30-vllm020-profile-v1/profile_vllm020_allreduce.py

@@ -113,10 +113,7 @@ def main() -> None:
                     f"FlashInfer all-reduce was not selected at TP={world_size}, "
                     f"tokens={num_tokens}"
                 )
-            if not fi_comm.should_use_fi_ar(input_tensor):
-                raise SystemExit(
-                    f"FlashInfer rejected serving tensor shape {tuple(input_tensor.shape)}"
-                )
+            uses_flashinfer = fi_comm.should_use_fi_ar(input_tensor)
 
             samples: list[float] = []
             for _ in range(args.repeats):
@@ -142,7 +139,9 @@ def main() -> None:
                     * torch.tensor([], dtype=torch.bfloat16).element_size(),
                     "dtype": "bfloat16",
                     "communicator": "vllm.tensor_model_parallel_all_reduce",
-                    "selected_backend": "flashinfer_trtllm",
+                    "selected_backend": (
+                        "flashinfer_trtllm" if uses_flashinfer else "nccl_fallback"
+                    ),
                     "per_rank_time_ms": per_rank,
                     "critical_path_median_ms": max(
                         rank_stats["median"] for rank_stats in per_rank