training ir torchao migration (#1006)

Summary:

Migrate capture_pre_autograd_graph to export_for_training.

We still need to keep capture_pre_autograd_graph call because torch/ao's CI tests uses earlier version of pytorch that does not have export_for_training.

See https://github.com/pytorch/ao/blob/main/.github/workflows/regression_test.yml

Differential Revision: D63859678

Author: yushangdi

test/dtypes/test_uint4.py

@@ -7,7 +7,6 @@
 from torch.ao.quantization.quantize_pt2e import prepare_pt2e, convert_pt2e
 from torch.ao.quantization.quantizer import QuantizationSpec, Quantizer
 
-from torch._export import capture_pre_autograd_graph
 from torch.testing._internal.common_quantization import (
     NodeSpec as ns,
     QuantizationTestCase,
@@ -26,6 +25,14 @@
 )
 import copy
 
+has_export_for_training = False
+try:
+    from torch.export import export_for_training
+    has_export_for_training = True
+except ImportError:
+    # capture_pre_autograd_graph is deprecated, it's
+    # left here to work with previous versions of pytorch
+    from torch._export import capture_pre_autograd_graph
 
 def _apply_weight_only_uint4_quant(model):
     def fn(mod):
@@ -203,10 +210,16 @@ def forward(self, x):
 
         # program capture
         m = copy.deepcopy(m_eager)
-        m = capture_pre_autograd_graph(
-            m,
-            example_inputs,
-        )
+        if has_export_for_training:
+            m = export_for_training(
+                m,
+                example_inputs,
+            ).module()
+        else:
+            m = capture_pre_autograd_graph(
+                    m,
+                    example_inputs,
+                ).module()
 
         m = prepare_pt2e(m, quantizer)
         # Calibrate

test/integration/test_integration.py

@@ -1484,11 +1484,21 @@ def forward(self, x):
 
         # make sure it compiles
         example_inputs = (x,)
-        from torch._export import capture_pre_autograd_graph
         # TODO: export changes numerics right now, this is because of functionalization according to Zhengxu
         # we can re-enable this after non-functional IR is enabled in export
         # model = torch.export.export(model, example_inputs).module()
-        model = capture_pre_autograd_graph(model, example_inputs)
+        has_export_for_training = False
+        try:
+            from torch.export import export_for_training
+            has_export_for_training = True
+        except ImportError:
+            # capture_pre_autograd_graph is deprecated, it's
+            # left here to work with previous versions of pytorch
+            from torch._export import capture_pre_autograd_graph
+        if has_export_for_training:
+            model = export_for_training(model, example_inputs).module()
+        else:
+            model = capture_pre_autograd_graph(model, example_inputs)
         after_export = model(x)
         self.assertTrue(torch.equal(after_export, ref))
         if api is _int8da_int8w_api:

torchao/utils.py

@@ -180,7 +180,7 @@ def _the_op_that_needs_to_be_preserved(...)
 
         # after this, `_the_op_that_needs_to_be_preserved` will be preserved as
         # torch.ops.my_namespace.the_op_that_needs_to_be_preserved operator after
-        # torch.export.export / torch._export.capture_pre_autograd_graph
+        # torch.export.export / torch._export.export_for_training
 
     """
     from torch._inductor.decomposition import register_decomposition