Adding the best autotuner config (#146121)

Mingming-Ding · pytorchmergebot · commit bcd0ba0f6968 · 2025-02-01T03:43:33.000Z
Summary: Adding logs to log the best config for autotune configs Test Plan: Testing in Mast : aps-omnifmv1-5_32_test_with_best_config-c5e9ceccf8 {F1974838864} Reviewed By: oulgen Differential Revision: D68931164 Pull Request resolved: #146121 Approved by: https://github.com/oulgen
diff --git a/torch/_inductor/runtime/triton_heuristics.py b/torch/_inductor/runtime/triton_heuristics.py
@@ -740,8 +740,21 @@ def autotune_to_one_config(self, *args, **kwargs):
         self.autotune_time_taken_ns = (
             self.precompile_time_taken_ns + benchmark_time_taken_ns
         )
+
+        # log the best config
+        launcher = self.launchers[0]
+        log.debug(
+            "Best config for %s: %s: %f, nreg %d, nspill %d, #shared-mem %s",
+            self.fn.__name__,
+            launcher.config,
+            timings[launcher],
+            launcher.n_regs,
+            launcher.n_spills,
+            launcher.shared,
+        )
+
         if self.save_cache_hook:
-            self.save_cache_hook(self.launchers[0].config, self.autotune_time_taken_ns)
+            self.save_cache_hook(launcher.config, self.autotune_time_taken_ns)
 
     def save_gpu_kernel(self, grid, stream, launcher):
         if callable(grid):
