Make HELION_FORCE_AUTOTUNE or kernel.autotune() skip the cache

helion/autotuner/base_cache.py

@@ -157,8 +157,13 @@ def _get_cache_info_message(self) -> str:
         """Return a message describing where the cache is and how to clear it."""
         return ""
 
-    def autotune(self) -> Config:
-        if os.environ.get("HELION_SKIP_CACHE", "") not in {"", "0", "false", "False"}:
+    def autotune(self, *, skip_cache: bool = False) -> Config:
+        if skip_cache or os.environ.get("HELION_SKIP_CACHE", "") not in {
+            "",
+            "0",
+            "false",
+            "False",
+        }:
             return self.autotuner.autotune()
 
         if (config := self.get()) is not None:

helion/autotuner/base_search.py

@@ -65,7 +65,7 @@ class BaseAutotuner(abc.ABC):
     """
 
     @abc.abstractmethod
-    def autotune(self) -> Config:
+    def autotune(self, *, skip_cache: bool = False) -> Config:
         raise NotImplementedError
 
 
@@ -420,7 +420,7 @@ def parallel_benchmark(
                 results.append((config, fn, inf))
         return results
 
-    def autotune(self) -> Config:
+    def autotune(self, *, skip_cache: bool = False) -> Config:
         """
         Perform autotuning to find the best configuration.
 

helion/runtime/kernel.py

@@ -253,7 +253,7 @@ def autotune(
         self,
         args: Sequence[object],
         *,
-        force: bool = False,
+        force: bool = True,
         **options: object,
     ) -> Config:
         """
@@ -475,7 +475,7 @@ def autotune(
         self,
         args: Sequence[object],
         *,
-        force: bool = False,
+        force: bool = True,
         **kwargs: object,
     ) -> Config:
         """
@@ -508,7 +508,9 @@ def autotune(
                 config = FiniteSearch(self, args, self.configs).autotune()
         else:
             self.settings.check_autotuning_disabled()
-            config = self.settings.autotuner_fn(self, args, **kwargs).autotune()
+            config = self.settings.autotuner_fn(self, args, **kwargs).autotune(
+                skip_cache=force
+            )
 
         self.set_config(config)
         return config
@@ -623,7 +625,7 @@ def __call__(self, *args: object) -> _R:
             if (config := self._implicit_config()) is not None:
                 self.set_config(config)
             else:
-                self.autotune(args)
+                self.autotune(args, force=False)
             assert self._run is not None
 
         assert self._config is not None