[AUTOGENERATED] [release/2.5] [release/2.7] [SWDEV-543214] Reland #2416 Fix warps runtime part 2 (#2452)

Cherry-pick of #2442

---------

Co-authored-by: Jack Taylor <[email protected]>

Author: okakarpa

torch/_inductor/runtime/coordinate_descent_tuner.py

@@ -3,6 +3,7 @@
 import itertools
 import logging
 from typing import Callable, Optional
+from functools import lru_cache
 
 from .hints import TRITON_MAX_BLOCK
 from .runtime_utils import red_text, triton_config_to_hashable
@@ -69,10 +70,16 @@ def get_config_max(self, prefix: str) -> int:
         size_hint = self.prefix_to_size_hint(prefix)
         return min(max_block, size_hint) if size_hint is not None else max_block
 
+    @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