fix(deepep): tuning and enhance deepep on single node

feat(deepep): enhance single node deepep

python/sglang/srt/layers/moe/token_dispatcher/deepep.py

@@ -218,14 +218,19 @@ class DeepEPBuffer:
                     f"Consider using --deepep-config to change the behavior."
                 )
 
+        is_single_node = group.size() <= torch.cuda.device_count()
+        allow_mnnvl = True
+        if is_single_node and not deepep_mode.enable_low_latency():
+            num_rdma_bytes = 0
+            allow_mnnvl = False
+
         cls._buffer = Buffer(
             group,
             num_nvl_bytes,
             num_rdma_bytes,
             low_latency_mode=deepep_mode.enable_low_latency(),
             num_qps_per_rank=num_qps_per_rank,
-            # TODO can be false when unneeded
-            allow_mnnvl=True,
+            allow_mnnvl=allow_mnnvl,
         )
         return cls._buffer