enable csgmv automatically on cuda (#13600)

python/sglang/srt/utils/common.py

@@ -3618,6 +3618,12 @@ def cached_triton_kernel(key_fn=None):
     """
 
     def decorator(fn):
+        # Auto-enable the custom kernel cache for CUDA, where it is
+        # known to be compatible.
+        if is_cuda() and not envs.SGLANG_USE_CUSTOM_TRITON_KERNEL_CACHE.is_set():
+            logger.debug("Detected platform CUDA, using custom triton kernel cache.")
+            return CachedKernel(fn, key_fn)
+
         if envs.SGLANG_USE_CUSTOM_TRITON_KERNEL_CACHE.get():
             logger.debug(
                 f"{envs.SGLANG_USE_CUSTOM_TRITON_KERNEL_CACHE.name} = True. Using custom triton kernel cache."