[inductor] switch assume_aligned_inputs to False

ghstack-source-id: 8c5d6c9c965c3310bf61c76afdc7e34c6baa35d9
Pull Request resolved: #124336

benchmarks/dynamo/ci_expected_accuracy/cpu_inductor_torchbench_inference.csv

@@ -58,23 +58,23 @@ detectron2_fasterrcnn_r_101_c4,pass,42
 
 
 
-detectron2_fasterrcnn_r_101_dc5,pass,42
+detectron2_fasterrcnn_r_101_dc5,pass,43
 
 
 
 detectron2_fasterrcnn_r_101_fpn,pass,46
 
 
 
-detectron2_fasterrcnn_r_50_c4,pass,42
+detectron2_fasterrcnn_r_50_c4,pass,43
 
 
 
 detectron2_fasterrcnn_r_50_dc5,pass,42
 
 
 
-detectron2_fasterrcnn_r_50_fpn,pass,46
+detectron2_fasterrcnn_r_50_fpn,pass,47
 
 
 
@@ -86,15 +86,15 @@ detectron2_maskrcnn_r_101_c4,fail_accuracy,57
 
 
 
-detectron2_maskrcnn_r_101_fpn,pass,64
+detectron2_maskrcnn_r_101_fpn,pass,65
 
 
 
-detectron2_maskrcnn_r_50_c4,pass,57
+detectron2_maskrcnn_r_50_c4,pass,58
 
 
 
-detectron2_maskrcnn_r_50_fpn,pass,64
+detectron2_maskrcnn_r_50_fpn,pass,65
 
 
 

test/inductor/test_cudagraph_trees.py

@@ -481,7 +481,7 @@ def foo(x, y):
             new_id = self.get_manager().new_graph_id().id
             self.assertEqual(new_id, 3)
 
-        def _test_unaligned_static_input_impl(self):
+        def _test_unaligned_static_input_impl(self, expected_clones):
             def fn(x, y):
                 return (x + y,)
 
@@ -512,21 +512,21 @@ def __torch_dispatch__(self, func, types, args=(), kwargs=None):
             for _ in range(3):
                 with CloneCounterMode() as m:
                     compiled_f(get_unaligned_inputs())
-                    self.assertEqual(m.count, 2)
+                    self.assertEqual(m.count, expected_clones)
 
                     compiled_f(get_aligned_inputs())
-                    self.assertEqual(m.count, 2)
+                    self.assertEqual(m.count, expected_clones)
 
         def test_unaligned_static_input_trees(self):
-            self._test_unaligned_static_input_impl()
+            self._test_unaligned_static_input_impl(expected_clones=0)
 
         @torch._inductor.config.patch("triton.cudagraph_trees", False)
         def test_unaligned_static_input_non_trees(self):
-            self._test_unaligned_static_input_impl()
+            self._test_unaligned_static_input_impl(expected_clones=0)
 
         @torch._inductor.config.patch("triton.cudagraphs", False)
         def test_unaligned_static_input_no_cudagraphs(self):
-            self._test_unaligned_static_input_impl()
+            self._test_unaligned_static_input_impl(expected_clones=0)
 
         def test_sparsity(self):
             def foo(view_6, buf31):

torch/_inductor/config.py

@@ -491,7 +491,7 @@ def decide_compile_threads():
 # assume_aligned_inputs means that we assume that inputs will be aligned; we generate
 # code using this assumption, and clone tensors before use if they aren't aligned.
 # In the common case, most inputs will be aligned.
-assume_aligned_inputs: bool = True
+assume_aligned_inputs: bool = False
 
 
 # config specific to codegen/cpp.py