[Perf][Bugfix] Prefer MNNVL for single-node FlashInfer allreduce again (#61989)

Signed-off-by: khluu <khluu000@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Wei Zhao <51183510+wzhao18@users.noreply.github.com>

tests/distributed/test_comm_ops.py

@@ -462,8 +462,8 @@ def test_flashinfer_standalone_workspace_size(
 @pytest.mark.parametrize(
     ("node_count", "is_blackwell", "expected"),
     [
-        (1, False, ("trtllm", False)),
-        (1, True, ("trtllm", False)),
+        (1, False, ("mnnvl", True)),
+        (1, True, ("mnnvl", True)),
         (2, False, (None, False)),
         (2, True, ("mnnvl", False)),
     ],

vllm/distributed/device_communicators/flashinfer_all_reduce.py

@@ -156,10 +156,12 @@ def _resolve_fi_ar_backend(group: ProcessGroup) -> tuple[str | None, bool]:
 
     node_count = _node_count(group)
     if node_count == 1:
-        # TRTLLM supports single-node allreduce on Hopper and Blackwell. Do not
-        # create an MNNVL workspace when the process group is node-local.
-        backend = "trtllm"
-        allow_trtllm_fallback = False
+        # MNNVL decodes faster than TRTLLM on single-node NVSwitch systems
+        # (about 3-5% throughput at TP4/TP8 on H200 and B200). Fall back to
+        # TRTLLM where the node lacks NVSwitch multicast and MNNVL cannot
+        # create its workspace.
+        backend = "mnnvl"
+        allow_trtllm_fallback = True
     elif current_platform.has_device_capability(100):
         # MNNVL requires Blackwell-class GPUs and is the only supported
         # FlashInfer backend for a multi-node allreduce process group.