misc: cache is_hopper_arch (#6799)

sgl-kernel/python/sgl_kernel/utils.py

@@ -13,6 +13,7 @@
 # limitations under the License.
 # ==============================================================================
 
+import functools
 from typing import Dict, Tuple
 
 import torch
@@ -41,6 +42,7 @@ def _to_tensor_scalar_tuple(x):
         return (None, x)
 
 
+@functools.lru_cache(maxsize=1)
 def is_hopper_arch() -> bool:
     # Hopper arch's compute capability == 9.0
     device = torch.cuda.current_device()