Add back inliner to version converter (#2813)

Basically, version converter only supports inlined model. In terms of
torch onnx exporter, we should just remind/warn users that if their
models includes local-function. We just don't support it.

series of version converter PRs:
#2791 (Removed assertion of
no local function included)
#2799 (Missing opset import)
#2806 (Raise for RefAttr)

Maybe we should just try revert
#2791?

cc @gramalingam @justinchuby

Author: titaiwangms

onnxscript/version_converter/__init__.py

@@ -37,7 +37,9 @@ def __init__(self, target_version: int, fallback: bool = False) -> None:
         super().__init__()
         self.target_version = target_version
         self.fallback = fallback
-        self._convert_pass = _ConvertVersionPass(
+        # NOTE: The current version converter only supports inlined models.
+        self._inline_pass = common_passes.InlinePass()
+        self._convert_pass = _ConvertVersionPassRequiresInline(
             target_version=target_version,
             fallback=fallback,
         )
@@ -51,15 +53,16 @@ def call(self, model: ir.Model) -> ir.passes.PassResult:
         # Run the conversion pass outside of Sequential so that errors
         # (e.g. VersionConverterError) propagate directly without being
         # wrapped in PassError.
-        result = self._convert_pass(model)
-        cleanup_result = self._cleanup_passes(result)
+        inline_result = self._inline_pass(model)
+        result = self._convert_pass(inline_result.model)
+        cleanup_result = self._cleanup_passes(result.model)
         return ir.passes.PassResult(
             cleanup_result.model,
-            result.modified or cleanup_result.modified,
+            result.modified or cleanup_result.modified or inline_result.modified,
         )
 
 
-class _ConvertVersionPass(ir.passes.InPlacePass):
+class _ConvertVersionPassRequiresInline(ir.passes.InPlacePass):
     """Convert the model to the specified ONNX opset version.
 
     This pass leverages the onnxscript version converter to convert the model. If

onnxscript/version_converter/_version_converter_test.py

@@ -208,6 +208,7 @@ def test_version_convert_gridsample_cubic(self):
         self.assertEqual(model.graph.node(4).version, 20)
         self.assertEqual(model.graph.node(4).attributes["mode"].value, "cubic")
 
+    @pytest.mark.xfail(reason="Version converter does not currently support local-function.")
     def test_version_convert_function_nodes(self):
         """Test that version converter processes nodes inside model functions."""
         model = ir.from_onnx_text(
@@ -253,6 +254,7 @@ def test_version_convert_function_nodes(self):
         self.assertEqual(func[3].version, 20)
         self.assertEqual(len(func[3].inputs), 3)  # DFT 19->20 adds dft_length input
 
+    @pytest.mark.xfail(reason="Version converter does not currently support local-function.")
     def test_version_convert_function_with_control_flow_subgraph(self):
         """Test that version converter processes subgraphs inside control flow nodes in functions."""
         model = ir.from_onnx_text(
@@ -455,6 +457,7 @@ def test_metadata_is_copied_to_multiple_replacement_nodes(self):
                     f"Node {i} ({node.op_type}) should have metadata copied",
                 )
 
+    @pytest.mark.xfail(reason="Version converter does not currently support local-function.")
     def test_version_convert_raises_on_function_node_with_ref_attribute(self):
         """Test that version conversion raises when a function contains a node with a ref attribute."""
         # Build a function with a LeakyRelu node that uses a RefAttr for 'alpha'