解决custom allreduce在dp情况下的其服务错误问题

vllm/distributed/device_communicators/custom_all_reduce.py

@@ -142,7 +142,11 @@ class CustomAllreduce:
         else:
             device_ids = list(range(cuda_device_count_stateless()))
 
+        if (world_size == len(device_ids)):
+            physical_device_id = device_ids[device.index % world_size]
+        else:
             physical_device_id = device_ids[device.index]
+
         tensor = torch.tensor([physical_device_id], dtype=torch.int, device="cpu")
         gather_list = [
             torch.tensor([0], dtype=torch.int, device="cpu") for _ in range(world_size)