commit 9f96d4b61b1e0e8ef36a7906f41b4720baacda61
author Michael Lazos <mlazos@meta.com> Tue Jul 23 16:14:29 2024 -0700
committer PyTorch MergeBot <pytorchmergebot@users.noreply.github.com> Wed Jul 24 04:00:02 2024 +0000
tree 4baf85116497344d163bdfe5fa202f1edf22b4b4
parent 9575b1afad8c3b96c96e8644238bcb898fd5ba82

Disable inlining on cudagraph fallback tests (#131557)

The cudagraph fallback tests should only run without nn module inlining. The [rerecord limit](https://github.com/pytorch/pytorch/blob/fc3d2b26cd55e865655b099e47d5a5ecdaf2bd85/torch/_inductor/cudagraph_trees.py#L1922) is ignored if nn module inlining is disabled. Arguably it should just be higher, but this PR addresses the failures and allows inlining to be on by default on main.

Pull Request resolved: https://github.com/pytorch/pytorch/pull/131557
Approved by: https://github.com/anijain2305
ghstack dependencies: #131556

test/inductor/test_cudagraph_trees.py

diff --git a/test/inductor/test_cudagraph_trees.py b/test/inductor/test_cudagraph_trees.py
index 417e120..db74e48 100644
--- a/test/inductor/test_cudagraph_trees.py
+++ b/test/inductor/test_cudagraph_trees.py
@@ -1973,6 +1973,7 @@
             self.run_static_input_param_test(fn, 6)
 
         @torch._dynamo.config.patch("error_on_recompile", True)
+        @torch._dynamo.config.patch("inline_inbuilt_nn_modules", False)
         @torch._inductor.config.patch("triton.cudagraph_support_input_mutation", True)
         @torch._inductor.config.patch("triton.cudagraph_unexpected_rerecord_limit", 0)
         def test_fallback_to_eager_if_recompiling_too_many_times(self):
@@ -2008,6 +2009,7 @@
             self.assertEqual(counters["inductor"]["cudagraph_skips"], 1)
 
         @torch._dynamo.config.patch("error_on_recompile", True)
+        @torch._dynamo.config.patch("inline_inbuilt_nn_modules", False)
         @torch._inductor.config.patch("triton.cudagraph_support_input_mutation", True)
         @torch._inductor.config.patch("triton.cudagraph_unexpected_rerecord_limit", 0)
         def test_fallback_to_eager_if_recompiling_too_many_times_warn_only_once(self):
@@ -2052,6 +2054,7 @@
             )
             self.assertEqual(counters["inductor"]["cudagraph_skips"], 2)
 
+        @torch._dynamo.config.patch("inline_inbuilt_nn_modules", False)
         @torch._inductor.config.patch("triton.cudagraph_support_input_mutation", True)
         @torch._inductor.config.patch("triton.cudagraph_unexpected_rerecord_limit", 0)
         def test_fallback_to_eager_if_recompiling_too_many_times_due_to_cudagraph_managed_tensor(
@@ -2096,6 +2099,7 @@
             ).run(captured_output[0])
             self.assertEqual(counters["inductor"]["cudagraph_skips"], 1)
 
+        @torch._dynamo.config.patch("inline_inbuilt_nn_modules", False)
         @torch._dynamo.config.patch("error_on_recompile", True)
         @torch._dynamo.config.patch("inline_inbuilt_nn_modules", True)
         @torch._inductor.config.patch("triton.cudagraph_unexpected_rerecord_limit", 1)