diff --git a/torch/_inductor/runtime/coordinate_descent_tuner.py b/torch/_inductor/runtime/coordinate_descent_tuner.py
index 6626c88a1e0d..f58f4da06113 100644
--- a/torch/_inductor/runtime/coordinate_descent_tuner.py
+++ b/torch/_inductor/runtime/coordinate_descent_tuner.py
@@ -63,9 +63,14 @@ def get_config_max(self, prefix: str) -> int:
 
     @lru_cache(maxsize=1)
     def get_warpsmax(self):
-        # Currently, CUDA has a maximum of 1024 threads, so 32 is the max
-        # number of warps.
-        return 1024 // 32
+        # CUDA/ROCm has a maximum of 1024 threads per block
+        from torch.cuda import current_device, get_device_properties, is_available
+        
+        warp_size = (
+            get_device_properties(current_device()).warp_size if is_available() else 32
+        )
+
+        return 1024 // warp_size
 
     def cache_benchmark_result(self, config, timing):
         self.cached_benchmark_results[triton_config_to_hashable(config)] = timing