Add support for NCCL symmetric memory for TP allreduces (#8238)

python/sglang/srt/server_args.py

@@ -218,6 +218,7 @@ class ServerArgs:
     enable_profile_cuda_graph: bool = False
     enable_cudagraph_gc: bool = False
     enable_nccl_nvls: bool = False
+    enable_symm_mem: bool = False
     enable_tokenizer_batch_encode: bool = False
     disable_outlines_disk_cache: bool = False
     disable_custom_all_reduce: bool = False
@@ -1599,6 +1600,11 @@ class ServerArgs:
             action="store_true",
             help="Enable NCCL NVLS for prefill heavy requests when available.",
         )
+        parser.add_argument(
+            "--enable-symm-mem",
+            action="store_true",
+            help="Enable NCCL symmetric memory for fast collectives.",
+        )
         parser.add_argument(
             "--enable-tokenizer-batch-encode",
             action="store_true",