Update on "[Inductor] Construct subgraph with benchmarking args not example_inputs"

PaulZhang12 · PaulZhang12 · commit c81b6e8da02b · 2025-05-16T15:45:04.000-07:00
If the inputs to a subgraph has FlexibleLayout, the subgraph does not currently freeze the layouts here. Therefore, the `example_inputs` generated might not be consistent in layout with the `args` based in for benchmarking Differential Revision: [D74900879](https://our.internmc.facebook.com/intern/diff/D74900879/) cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov [ghstack-poisoned]
diff --git a/test/inductor/test_subgraph_choice.py b/test/inductor/test_subgraph_choice.py
@@ -106,7 +106,7 @@ def func(mat1, mat2):
         res = compiled_func(a_in, b_in)
 
         # Check same results of compiled result and regular torch.mm
-        torch.testing.assert_close(res, a_in @ b_in, atol=1e-2, rtol=1e-2)
+        torch.testing.assert_close(res, a_in @ b_in, atol=1e-1, rtol=1e-1)
 
     @skipIfXpu
     @unittest.skipIf(TEST_WITH_ROCM, "decompose_k not supported on ROCm")
@@ -186,8 +186,7 @@ def func(mat1, mat2):
             res = compiled_func(a_in, b_in)
 
             # Check same results of compiled result and regular torch.mm
-            # Relax precision as decomposeK does first accumulation in fp16
-            torch.testing.assert_close(res, (a_in + 1.0) @ b_in, atol=1e-2, rtol=1e-2)
+            torch.testing.assert_close(res, (a_in + 1.0) @ b_in, atol=1e-1, rtol=1e-1)
 
 
 if __name__ == "__main__":
