Compatible with lower versions of torch

Author: shihaobai

lightllm/server/router/model_infer/mode_backend/base_backend.py

@@ -84,8 +84,10 @@ def init_model(self, kvargs):
             init_method=f'tcp://127.0.0.1:{kvargs["nccl_port"]}',
             rank=self.tp_rank,
             world_size=self.world_size,
-            device_id=torch.device(f"cuda:{self.tp_rank}"),
         )
+        # warmup nccl communicator
+        a = torch.zeros([1]).to(f"cuda:{self.tp_rank}")
+        dist.all_reduce(a)
 
         from lightllm.distributed import set_custom_reduce