diff --git a/vllm/distributed/parallel_state.py b/vllm/distributed/parallel_state.py
index fa493fef..e0eeeffb 100644
--- a/vllm/distributed/parallel_state.py
+++ b/vllm/distributed/parallel_state.py
@@ -194,9 +194,11 @@ class GroupCoordinator:
 
         from vllm.platforms import current_platform
 
-        # TODO: fix it for other platforms
         if current_platform.is_cuda_alike():
             self.device = torch.device(f"cuda:{local_rank}")
+        elif current_platform.is_out_of_tree():
+            self.device = torch.device(
+                f"{current_platform.device_name}:{local_rank}")
         else:
             self.device = torch.device("cpu")