[Inductor] Fix arguments passed to triton kernel launch hooks (pytorc…

…h#128732)

`binary.launch_enter_hook` is treated as an instance method and will add a `self` argument to the hooks.
`CompiledKernel.launch_enter_hook` is a static method, which matches the hook calling convention of profilers (i.e., a single `LazyDict` argument only).

Pull Request resolved: pytorch#128732
Approved by: https://github.com/shunting314, https://github.com/bertmaher

test/inductor/test_profiler.py

@@ -158,10 +158,10 @@ def test_inductor_profiling_triton_hooks(self):
 
         hooks_called = {"enter": False, "exit": False}
 
-        def launch_enter_hook(*args):
+        def launch_enter_hook(lazy_dict):
             hooks_called["enter"] = True
 
-        def launch_exit_hook(*args):
+        def launch_exit_hook(lazy_dict):
             hooks_called["exit"] = True
 
         CompiledKernel.launch_enter_hook = launch_enter_hook

torch/_inductor/runtime/triton_heuristics.py

@@ -50,6 +50,7 @@
 
 if triton is not None:
     from triton import Config
+    from triton.compiler import CompiledKernel
     from triton.runtime.autotuner import OutOfResources
     from triton.runtime.jit import KernelInterface
 
@@ -453,8 +454,8 @@ def _precompile_config(self, cfg: Config, warm_cache_only: bool):
         scope = {
             "grid_meta": cfg.kwargs,
             "bin": binary,
-            "launch_enter_hook": binary.launch_enter_hook,
-            "launch_exit_hook": binary.launch_exit_hook,
+            "launch_enter_hook": CompiledKernel.launch_enter_hook,
+            "launch_exit_hook": CompiledKernel.launch_exit_hook,
             "metadata": binary.packed_metadata
             if hasattr(binary, "packed_metadata")
             else binary.metadata,