support 1 shot allreduce in 1-node and 2-node using mscclpp (#6277)

python/sglang/srt/server_args.py

@@ -165,6 +165,7 @@ class ServerArgs:
     enable_tokenizer_batch_encode: bool = False
     disable_outlines_disk_cache: bool = False
     disable_custom_all_reduce: bool = False
+    enable_mscclpp: bool = False
     disable_overlap_schedule: bool = False
     enable_mixed_chunk: bool = False
     enable_dp_attention: bool = False
@@ -1168,6 +1169,11 @@ class ServerArgs:
             action="store_true",
             help="Disable the custom all-reduce kernel and fall back to NCCL.",
         )
+        parser.add_argument(
+            "--enable-mscclpp",
+            action="store_true",
+            help="Enable using mscclpp for small messages for all-reduce kernel and fall back to NCCL.",
+        )
         parser.add_argument(
             "--disable-overlap-schedule",
             action="store_true",