Fix export config in examples (#34)

Summary:
Pull Request resolved: #34

## Context
It doesn't look correct that want to export models with one config but test the export with a different one.

## This DIff
 - Ensure canonical config is used for both export and tests
 - Ensure the test is loading the lowered model through `_load_for_executorch_from_buffer` for output consistency checking
 - Remove confusion about certain export flags by adding more inline comments to those flags, e.g. why  is `enable_dynamic_shape` there not enabled by default, what does `enable_aot` do with dynamic shapes, etc.

Reviewed By: JacobSzwejbka

Differential Revision: D48018569

fbshipit-source-id: 5e0696c19dea539dec5c23ac3566199bd2ce981f

Author: guangy10

examples/export/test/TARGETS

@@ -10,5 +10,6 @@ python_unittest(
         "//executorch/examples/export:utils",
         "//executorch/examples/models:models",
         "//executorch/exir:lib",
+        "//executorch/extension/pybindings:portable",  # @manual
     ],
 )

examples/export/test/test_export.py

@@ -8,26 +8,33 @@
 
 import torch
 
-from executorch.examples.export.utils import _EDGE_COMPILE_CONFIG
+from executorch.examples.export.utils import _CAPTURE_CONFIG, _EDGE_COMPILE_CONFIG
 from executorch.examples.models import MODEL_NAME_TO_MODEL
 
+# pyre-ignore[21]: Could not find module `executorch.extension.pybindings.portable`.
+from executorch.extension.pybindings.portable import (  # @manual
+    _load_for_executorch_from_buffer,
+)
+
 
 class ExportTest(unittest.TestCase):
     def _assert_eager_lowered_same_result(
         self, eager_model: torch.nn.Module, example_inputs
     ):
         import executorch.exir as exir
 
-        capture_config = exir.CaptureConfig(enable_dynamic_shape=False)
-        edge_model = exir.capture(eager_model, example_inputs, capture_config).to_edge(
+        edge_model = exir.capture(eager_model, example_inputs, _CAPTURE_CONFIG).to_edge(
             _EDGE_COMPILE_CONFIG
         )
 
         executorch_model = edge_model.to_executorch()
+        # pyre-ignore
+        pte_model = _load_for_executorch_from_buffer(executorch_model.buffer)
+
         with torch.no_grad():
             eager_output = eager_model(*example_inputs)
         with torch.no_grad():
-            executorch_output = executorch_model.graph_module(*example_inputs)
+            executorch_output = pte_model.forward(example_inputs)
         self.assertTrue(
             torch.allclose(eager_output, executorch_output[0], rtol=1e-5, atol=1e-5)
         )

examples/export/utils.py

@@ -11,7 +11,7 @@
 # Reason is that there memory allocation ops with symbolic shape nodes.
 # and when evaulating shape, it doesnt seem that we presenting them with shape env
 # that contain those variables.
-_CAPTURE_CONFIG = exir.CaptureConfig(enable_aot=True, _unlift=False)
+_CAPTURE_CONFIG = exir.CaptureConfig(enable_aot=True)
 _EDGE_COMPILE_CONFIG = exir.EdgeCompileConfig(
     _check_ir_validity=False,
 )

exir/capture/_config.py

@@ -19,8 +19,8 @@
 class CaptureConfig:
     pt2_mode: bool = True
     enable_functionalization: bool = True
-    enable_dynamic_shape: bool = False
-    enable_aot: bool = False
+    enable_dynamic_shape: bool = False  # This flag does nothing if enable_aot is True
+    enable_aot: bool = False  # When it's true it implies automatic dynamic shapes via default dynamo config
     _dynamo_config: "ExirDynamoConfig" = field(default_factory=ExirDynamoConfig)
     _unlift: bool = False
     _use_old_decomp_table: bool = False