Fix ONNX operator_export_type on the new registry (#87735) (#90044)

Fixes #87313

Our ONNX pipelines do not run with BUILD_CAFFE2=0, so tests for operator_export_type ONNX_ATEN and ONNX_ATEN_FALLBACK will not be fully tested, allowing regressions to happen again.

We need to run the same set of tests for both BUILD_CAFFE2=0 and 1
Pull Request resolved: #87735
Approved by: https://github.com/AllenTiTaiWang, https://github.com/BowenBao

(cherry picked from commit 2aed670)

Co-authored-by: Thiago Crepaldi <thiago.crepaldi@microsoft.com>

test/onnx/test_pytorch_onnx_no_runtime.py

@@ -32,6 +32,7 @@ def export_to_onnx(
     mocks: Optional[Iterable] = None,
     operator_export_type: torch.onnx.OperatorExportTypes = torch.onnx.OperatorExportTypes.ONNX,
     opset_version: int = GLOBALS.export_onnx_opset_version,
+    **torch_onnx_export_kwargs,
 ) -> onnx.ModelProto:
     """Exports `model(input)` to ONNX and returns it.
 
@@ -44,6 +45,7 @@ def export_to_onnx(
         mocks: list of mocks to use during export
         operator_export_type: export type as described by `torch.onnx.export(...operator_export_type,...)`
         opset_version: ONNX opset version as described by `torch.onnx.export(...opset_version,...)`
+        torch_onnx_export_kwargs: extra torch.onnx.export kwargs arguments
     Returns:
         A valid ONNX model (`onnx.ModelProto`)
     """
@@ -60,6 +62,7 @@ def export_to_onnx(
             f,
             operator_export_type=operator_export_type,
             opset_version=opset_version,
+            **torch_onnx_export_kwargs,
         )
 
     # Validate ONNX graph before returning it
@@ -777,6 +780,91 @@ def forward(self, x):
             model, inputs, f, dynamic_axes={"x": [0, 1]}, input_names=["x"]
         )
 
+    @common_utils.skipIfNoCaffe2
+    def test_caffe2_aten_fallback(self):
+        class ModelWithAtenNotONNXOp(torch.nn.Module):
+            def forward(self, x, y):
+                abcd = x + y
+                defg = torch.linalg.qr(abcd)
+                return defg
+
+        x = torch.rand(3, 4)
+        y = torch.rand(3, 4)
+        f = io.BytesIO()
+        torch.onnx.export(
+            ModelWithAtenNotONNXOp(),
+            (x, y),
+            f,
+            do_constant_folding=False,
+            operator_export_type=torch.onnx.OperatorExportTypes.ONNX_ATEN_FALLBACK,
+            # support for linalg.qr was added in later op set versions.
+            opset_version=9,
+        )
+        onnx_model = onnx.load(io.BytesIO(f.getvalue()))
+        self.assertAtenOp(onnx_model, "linalg_qr")
+
+    @common_utils.skipIfNoCaffe2
+    def test_caffe2_onnx_aten(self):
+        class ModelWithAtenFmod(torch.nn.Module):
+            def forward(self, x, y):
+                return torch.fmod(x, y)
+
+        x = torch.randn(3, 4, dtype=torch.float32)
+        y = torch.randn(3, 4, dtype=torch.float32)
+        f = io.BytesIO()
+        torch.onnx.export(
+            ModelWithAtenFmod(),
+            (x, y),
+            f,
+            do_constant_folding=False,
+            operator_export_type=torch.onnx.OperatorExportTypes.ONNX_ATEN,
+            opset_version=10,  # or higher
+        )
+        onnx_model = onnx.load(io.BytesIO(f.getvalue()))
+        assert onnx_model.graph.node[0].op_type == "Mod"
+
+    @common_utils.skipIfCaffe2
+    def test_aten_fallback(self):
+        class ModelWithAtenNotONNXOp(torch.nn.Module):
+            def forward(self, x, y):
+                abcd = x + y
+                defg = torch.linalg.qr(abcd)
+                return defg
+
+        x = torch.rand(3, 4)
+        y = torch.rand(3, 4)
+        f = io.BytesIO()
+        torch.onnx.export(
+            ModelWithAtenNotONNXOp(),
+            (x, y),
+            f,
+            do_constant_folding=False,
+            operator_export_type=torch.onnx.OperatorExportTypes.ONNX_ATEN_FALLBACK,
+            # support for linalg.qr was added in later op set versions.
+            opset_version=9,
+        )
+        onnx_model = onnx.load(io.BytesIO(f.getvalue()))
+        self.assertAtenOp(onnx_model, "linalg_qr")
+
+    @common_utils.skipIfCaffe2
+    def test_onnx_aten(self):
+        class ModelWithAtenFmod(torch.nn.Module):
+            def forward(self, x, y):
+                return torch.fmod(x, y)
+
+        x = torch.randn(3, 4, dtype=torch.float32)
+        y = torch.randn(3, 4, dtype=torch.float32)
+        f = io.BytesIO()
+        torch.onnx.export(
+            ModelWithAtenFmod(),
+            (x, y),
+            f,
+            do_constant_folding=False,
+            operator_export_type=torch.onnx.OperatorExportTypes.ONNX_ATEN,
+        )
+        onnx_model = onnx.load(io.BytesIO(f.getvalue()))
+        self.assertAtenOp(onnx_model, "fmod", "Tensor")
+
 
 if __name__ == "__main__":
     common_utils.run_tests()

torch/onnx/utils.py

@@ -1739,17 +1739,22 @@ def _add_output_to_block(block: _C.Block, value: _C.Value) -> int:
 
 @_beartype.beartype
 def _should_aten_fallback(
-    name: str,
-    opset_version: int,
-    operator_export_type: _C_onnx.OperatorExportTypes,
+    name: str, opset_version: int, operator_export_type: _C_onnx.OperatorExportTypes
 ):
+    # For BUILD_CAFFE2=0 builds, if domain=="aten" and operator_export_type==ONNX_ATEN,
+    #   an aten::ATen operator is created regardless of symbolics existence
+    # For BUILD_CAFFE2=1, the same applies only if there is no symbolic available
+
     is_exportable_aten_op = registration.registry.is_registered_op(name, opset_version)
     is_onnx_aten_export = operator_export_type == _C_onnx.OperatorExportTypes.ONNX_ATEN
     is_aten_fallback_export = (
         operator_export_type == _C_onnx.OperatorExportTypes.ONNX_ATEN_FALLBACK
     )
-    return is_onnx_aten_export or (
-        not is_exportable_aten_op and is_aten_fallback_export
+    is_caffe2_build = _C_onnx._CAFFE2_ATEN_FALLBACK
+
+    return name.startswith("aten::") and (
+        ((is_onnx_aten_export or is_aten_fallback_export) and not is_caffe2_build)
+        or (not is_exportable_aten_op and is_aten_fallback_export)
     )
 
 
@@ -1844,6 +1849,21 @@ def _run_symbolic_function(
         env=env,
     )
 
+    # Direct ATen export requested
+    if _should_aten_fallback(ns_op_name, opset_version, operator_export_type):
+        attrs = {
+            k + "_" + node.kindOf(k)[0]: symbolic_helper._node_get(node, k)
+            for k in node.attributeNames()
+        }
+        outputs = node.outputsSize()
+        attrs["outputs"] = outputs
+        return graph_context.at(
+            op_name,
+            *inputs,
+            overload_name=_get_aten_op_overload_name(node),
+            **attrs,
+        )
+
     try:
         # Caffe2-specific: Quantized op symbolics are registered for opset 9 only.
         if symbolic_helper.is_caffe2_aten_fallback() and opset_version == 9:
@@ -1861,6 +1881,7 @@ def _run_symbolic_function(
         if symbolic_function_group is not None:
             symbolic_fn = symbolic_function_group.get(opset_version)
             if symbolic_fn is not None:
+                # TODO Wrap almost identical attrs assignment or comment the difference.
                 attrs = {
                     k: symbolic_helper._node_get(node, k) for k in node.attributeNames()
                 }
@@ -1874,18 +1895,6 @@ def _run_symbolic_function(
             # Clone node to trigger ONNX shape inference
             return graph_context.op(op_name, *inputs, **attrs, outputs=node.outputsSize())  # type: ignore[attr-defined]
 
-        if _should_aten_fallback(ns_op_name, opset_version, operator_export_type):
-            # Direct ATen export requested
-            outputs = node.outputsSize()
-            attrs["outputs"] = outputs
-            # `overload_name` is set for non-Caffe2 builds only
-            return graph_context.at(
-                op_name,
-                *inputs,
-                overload_name=_get_aten_op_overload_name(node),
-                **attrs,
-            )
-
         raise errors.UnsupportedOperatorError(
             domain,
             op_name,