Add preserve_ops to EdgeCompileConfig (#12546)

1. Add `preserve_ops` to `EdgeCompileConfig`
2. Remove preserved ops from the decomposition table in `to_edge`.
3. Add checks to the verifier ensuring that preserved ops do not have mutations or views.
4. Update 'core_aten_exception_list' to be 'preserved_ops' in `to_edge_transform_and_lower`.

Context/Usage
**core_aten_ops_exception_list**
- Contains operators that are missing a decomposition to core aten.
- Exclude these so that verification can still be run on the rest of the graph.
- Ideally, this list should be empty.

**preserve_ops**
- Contains operators that the user specifically does not want decomposed.
- Must be aten; custom ops are ignored by verifier.

Edge case:
- If an aten operator does not have a decomp, and the user specifically wants it to be preserved, put it in preserve_ops rather than core_aten_ops_exception_list.

Differential Revision: [D78298749](https://our.internmc.facebook.com/intern/diff/D78298749/)

Author: pytorchbot

backends/nxp/nxp_backend.py

@@ -174,7 +174,7 @@ def preprocess(
             # Otherwise, we get violation that this op is not part of ATen Core ops.
             edge_program._verifiers = [
                 EXIREdgeDialectVerifier(
-                    class_only=True, exception_list=[torch.ops.aten.max_pool2d.default]
+                    class_only=True, core_aten_ops_exception_list=[torch.ops.aten.max_pool2d.default]
                 )
             ]
 

exir/capture/_config.py

@@ -40,9 +40,14 @@ class EdgeCompileConfig:
     # TODO(larryliu): remove this
     _use_edge_ops: bool = True
     # Allow core ATen ops check to be skipped for certain ops, but continue with the rest of the checks.
+    # Note: only use this for core ATen ops that are missing decompositions. This is temporary,
+    # enabling verification on the rest of the program until decomposition coverage is improved.
     _core_aten_ops_exception_list: List[torch._ops.OpOverload] = field(
         default_factory=list
     )
+    # Allow ops to be preserved in the graph, i.e., prevent them from being decomposed.
+    # These may be core or non-core ATen ops; custom ops should not be here.
+    _preserve_ops: List[torch.torch._ops.OpOverload] = field(default_factory=list)
     # TODO(gasoonjia): remove this
     _skip_dim_order: bool = False
 

exir/program/_program.py

@@ -795,9 +795,19 @@ def _generate_edge_program(
     name: str,
     config: EdgeCompileConfig,
     program: ExportedProgram,
-    ops_set_to_not_decompose: Optional[List[torch._ops.OpOverload]] = None,
+    core_aten_ops_exception_list: Optional[List[torch._ops.OpOverload]] = None,
+    preserve_ops: Optional[List[torch._ops.OpOverload]] = None,
 ) -> ExportedProgram:
-
+    """
+    Args:
+        name: The name of the program.
+        config: The configuration for the edge program.
+        program: The exported program to be converted to an edge program.
+        core_aten_ops_exception_list: A list of aten ops that are missing decompositions to core aten.
+        preserve_ops: A list of aten ops that should not be decomposed.
+    Returns:
+        An ExportedProgram in edge dialect.
+    """
     # Remove invalid assert ops, such as _assert_tensor_metadata
     gm = program.graph_module
     gm_res = RemoveNonCoreAtenOpGraphAssertsPass()(gm)
@@ -812,7 +822,8 @@ def _generate_edge_program(
             EXIRATenDialectVerifier(
                 edge_compile_config=config,
                 class_only=False,
-                exception_list=ops_set_to_not_decompose,
+                core_aten_ops_exception_list=core_aten_ops_exception_list,
+                preserve_ops=preserve_ops,
             )(gm)
         except ExportError as e:
             logging.info(f"Input program {name} is not in ATen dialect.")
@@ -848,7 +859,8 @@ def _generate_edge_program(
             EXIREdgeDialectVerifier(
                 edge_compile_config=config,
                 class_only=True,
-                exception_list=ops_set_to_not_decompose,
+                core_aten_ops_exception_list=core_aten_ops_exception_list,
+                preserve_ops=preserve_ops,
             )
         ],
     )
@@ -864,7 +876,7 @@ def _replace_aten_ops_with_transformed_ops(
     program: ExportedProgram,
     partitioner,
 ):
-    ops_to_not_decompose = set()
+    preserve_ops = set()
     partitioners = partitioner.get(name)
     if partitioners is None:
         return
@@ -889,7 +901,7 @@ def _replace_aten_ops_with_transformed_ops(
                 and node.target in ops_set_to_not_decompose
                 and is_op_supported
             ):
-                ops_to_not_decompose.add(node.target)
+                preserve_ops.add(node.target)
                 node.target = aten_op_to_transform_op[node.target]
 
         for _, submod, _ in get_control_flow_submodules(program.graph_module):
@@ -900,10 +912,10 @@ def _replace_aten_ops_with_transformed_ops(
                     and node.target in ops_set_to_not_decompose
                     and is_op_supported
                 ):
-                    ops_to_not_decompose.add(node.target)
+                    preserve_ops.add(node.target)
                     node.target = aten_op_to_transform_op[node.target]
 
-    return ops_to_not_decompose
+    return preserve_ops
 
 
 def _restore_transformed_ops_to_aten_ops(program: ExportedProgram):
@@ -1014,7 +1026,7 @@ def _sanity_check_graph_for_non_decomp_ops(
 
 
 def _remove_invalid_ops_for_not_decompose(
-    ops_to_not_decompose: List[torch._ops.OpOverload],
+    preserve_ops: List[torch._ops.OpOverload],
 ) -> List[torch._ops.OpOverload]:
     _logged_warnings = set()
 
@@ -1079,7 +1091,7 @@ def keep(op):
             return False
         return True
 
-    return list(filter(keep, ops_to_not_decompose))
+    return list(filter(keep, preserve_ops))
 
 
 def _gen_edge_manager_for_partitioners(
@@ -1136,7 +1148,7 @@ def _gen_edge_manager_for_partitioners(
             name,
             config,
             program,
-            list(ops_set_to_not_decompose_by_program.get(name, [])),
+            preserve_ops=list(ops_set_to_not_decompose_by_program.get(name, [])),
         )
 
     edge_manager = EdgeProgramManager(
@@ -1281,7 +1293,7 @@ def to_edge_transform_and_lower(
             EXIREdgeDialectVerifier(
                 edge_compile_config=config,
                 class_only=True,
-                exception_list=list(ops_set_to_not_decompose),
+                preserve_ops=list(ops_set_to_not_decompose),
             )()(program.graph_module)
 
     return edge_manager
@@ -1328,7 +1340,7 @@ def to_edge_with_preserved_ops(
             table.pop(op, None)
         program = program.run_decompositions(table)
         edge_programs[name] = _generate_edge_program(
-            name, config, program, list(preserve_ops)
+            name, config, program, preserve_ops=list(preserve_ops)
         )
 
     return EdgeProgramManager(
@@ -1367,8 +1379,16 @@ def to_edge(
 
     for name, program in aten_programs.items():
         # Decompose to Core ATen
-        program = program.run_decompositions(_default_decomposition_table())
-        edge_programs[name] = _generate_edge_program(name, config, program)
+        table = _default_decomposition_table()
+        preserve_ops = []
+        if compile_config:
+            preserve_ops = compile_config._preserve_ops
+            for op in compile_config._preserve_ops:
+                table.pop(op, None)
+        program = program.run_decompositions(table)
+        edge_programs[name] = _generate_edge_program(
+            name, config, program, preserve_ops=preserve_ops
+        )
 
     return EdgeProgramManager(edge_programs, constant_methods, config)
 
@@ -1389,7 +1409,8 @@ def __init__(
         edge_programs: Union[ExportedProgram, Dict[str, ExportedProgram]],
         constant_methods: Optional[Dict[str, Any]] = None,
         compile_config: Optional[EdgeCompileConfig] = None,
-        ops_set_to_not_decompose: Optional[List[torch._ops.OpOverload]] = None,
+        core_aten_ops_exception_list: Optional[List[torch._ops.OpOverload]] = None,
+        preserve_ops: Optional[List[torch._ops.OpOverload]] = None,
     ):
         """
         Should not be called directly by users. User should use :func:'to_edge' instead.
@@ -1404,7 +1425,8 @@ def __init__(
             try:
                 EXIREdgeDialectVerifier(
                     edge_compile_config=self.compile_config,
-                    exception_list=ops_set_to_not_decompose,
+                    core_aten_ops_exception_list=core_aten_ops_exception_list,
+                    preserve_ops=preserve_ops,
                 )(program.graph_module)
             except ExportError as e:
                 logging.info(f"Input program {name} is not in aten dialect.")

exir/program/test/test_program.py

@@ -27,7 +27,6 @@
     ExecutorchProgramManager,
     to_edge,
     to_edge_transform_and_lower,
-    to_edge_with_preserved_ops,
 )
 from executorch.exir.tracer import _default_decomposition_table
 from executorch.exir.verification.verifier import EXIREdgeDialectVerifier
@@ -784,7 +783,9 @@ def forward(self, x: torch.Tensor) -> torch.Tensor:
     def _test_to_edge_with_preserved_ops(
         self, program, preserved_ops, expected_preserved_ops
     ):
-        edge = to_edge_with_preserved_ops(program, preserve_ops=preserved_ops)
+        edge = to_edge(
+            program, compile_config=EdgeCompileConfig(_preserve_ops=preserved_ops)
+        )
 
         def count_nodes(graph_module, target):
             count = 0

exir/verification/test/test_verifier.py

@@ -161,3 +161,17 @@ def forward(self, input, label):
         edge_verifier = EXIREdgeDialectVerifier()
 
         edge_verifier(edge.exported_program())
+
+    def test_verifier_preserve_ops_view(self) -> None:
+        class TestExpand(nn.Module):
+            def __init__(self):
+                super().__init__()
+
+            def forward(self, x):
+                return x.expand(2, 2, 2, 2)
+
+        model = TestExpand()
+        config = EdgeCompileConfig(_preserve_ops=[torch.ops.aten.expand.default])
+        export_model = export(model, (torch.randn(2, 2, 2, 2),), strict=True)
+        with self.assertRaises(RuntimeError):
+            to_edge(export_model, compile_config=config)

exir/verification/verifier.py

@@ -3,8 +3,11 @@
 #
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
+#
+# pyre-unsafe
 
 import itertools
+import logging
 import operator
 import types
 from contextlib import nullcontext
@@ -81,26 +84,33 @@ def __call__(self, *args, **kwargs):
 def EXIRATenDialectVerifier(  # noqa: C901
     edge_compile_config: Optional[EdgeCompileConfig] = None,
     class_only: bool = False,
-    exception_list: Optional[List[torch._ops.OpOverload]] = None,
+    core_aten_ops_exception_list: Optional[List[torch._ops.OpOverload]] = None,
+    preserve_ops: Optional[List[torch._ops.OpOverload]] = None,
 ):
     """
     Returns a verifier class that runs ATen dialect specific checks on the graph module.
     """
+    _core_aten_ops_exception_list = core_aten_ops_exception_list or []
+    _preserve_ops = preserve_ops or []
     # merge the exception list from edge_compile_config and exception_list
-    if edge_compile_config and edge_compile_config._core_aten_ops_exception_list:
-        exception_list = edge_compile_config._core_aten_ops_exception_list + (
-            exception_list or []
-        )
+    if edge_compile_config:
+        if edge_compile_config._core_aten_ops_exception_list:
+            _core_aten_ops_exception_list.extend(
+                edge_compile_config._core_aten_ops_exception_list
+            )
+        if edge_compile_config._preserve_ops:
+            _preserve_ops.extend(edge_compile_config._preserve_ops)
 
     class _EXIRATenDialectVerifier(EXIRATenDialectVerifierBase):
         dialect = "OLD_EXIR_ATEN"
 
         def __init__(self) -> None:
             super().__init__()
             # Note: here we are using the exception list passed from EXIRATenDialectVerifier function!
-            self._exception_list = exception_list if exception_list else []
+            self._core_aten_ops_exception_list = _core_aten_ops_exception_list
+            self._preserve_ops = _preserve_ops
 
-        def _get_exception_list(self) -> List[torch._ops.OpOverload]:
+        def _get_core_aten_ops_exception_list(self) -> List[torch._ops.OpOverload]:
             exception_list = (
                 [
                     torch.ops.aten.mkldnn_rnn_layer.default,
@@ -113,15 +123,35 @@ def _get_exception_list(self) -> List[torch._ops.OpOverload]:
                 ]
                 + list(_EXECUTORCH_SYM_OPS)
                 + DISALLOW_LIST
-                + self._exception_list
+                + self._core_aten_ops_exception_list
             )
 
             return exception_list
 
         def check_valid_op(self, op):
             if isinstance(op, OpOverload):
                 # TODO These special ops should be removable easily.
-                if op.namespace != "aten" or op in self._get_exception_list():
+                if (
+                    op.namespace != "aten"
+                    or op in self._get_core_aten_ops_exception_list()
+                ):
+                    return
+                if op in self._preserve_ops:
+                    if op.namespace != "aten":
+                        raise RuntimeError(
+                            f"Only preserve aten ops. Received op {op} with namespace {op.namespace}."
+                        )
+                    # Preserved ops should not include mutation or view,
+                    # which may affect memory planning.
+                    if op.is_view:
+                        raise RuntimeError(
+                            f"Cannot preserve operator {op} because it is a view or mutation."
+                        )
+                    if op._schema.is_mutable:
+                        logging.warning(
+                            f"Preserving mutation ops like {op} is a no-op because run_decomposition functionalizes it and prevents it from showing up."
+                        )
+
                     return
                 if torch.Tag.core not in op.tags and torch.Tag.view_copy not in op.tags:
                     # NOTE(qihan): whether view_copy operators are marked as canonical is still under
@@ -149,7 +179,9 @@ def check_valid_op(self, op):
 def get_aten_verifier(config: EdgeCompileConfig):
     return (
         EXIRATenDialectVerifier(
-            class_only=True, exception_list=config._core_aten_ops_exception_list
+            class_only=True,
+            core_aten_ops_exception_list=config._core_aten_ops_exception_list,
+            preserve_ops=config._preserve_ops,
         )
         if config._check_ir_validity
         else EXIRATenDialectVerifierBase
@@ -210,13 +242,19 @@ def _check_tensor_args_matching_op_allowed_dtype(gm: GraphModule) -> None:
 def EXIREdgeDialectVerifier(  # noqa: C901
     edge_compile_config: Optional[EdgeCompileConfig] = None,
     class_only: bool = False,
-    exception_list: Optional[List[torch._ops.OpOverload]] = None,
+    core_aten_ops_exception_list: Optional[List[torch._ops.OpOverload]] = None,
+    preserve_ops: Optional[List[torch._ops.OpOverload]] = None,
 ):
+    _core_aten_ops_exception_list = core_aten_ops_exception_list or []
+    _preserve_ops = preserve_ops or []
     # merge the exception list from edge_compile_config and exception_list
-    if edge_compile_config and edge_compile_config._core_aten_ops_exception_list:
-        exception_list = edge_compile_config._core_aten_ops_exception_list + (
-            exception_list or []
-        )
+    if edge_compile_config:
+        if edge_compile_config._core_aten_ops_exception_list:
+            _core_aten_ops_exception_list.extend(
+                edge_compile_config._core_aten_ops_exception_list
+            )
+        if edge_compile_config._preserve_ops:
+            _preserve_ops.extend(edge_compile_config._preserve_ops)
 
     class _EXIREdgeDialectVerifier(Verifier):
         dialect = "EDGE"
@@ -228,16 +266,19 @@ def __init__(self) -> None:
             self.check_edge_ops = _edge_compile_config._use_edge_ops
             self.use_dim_order = not _edge_compile_config._skip_dim_order
 
+            self._core_aten_ops_exception_list = _core_aten_ops_exception_list
+            self._preserve_ops = _preserve_ops
+
             self.aten_op_verifier = EXIRATenDialectVerifier(
-                exception_list=exception_list
+                core_aten_ops_exception_list=_core_aten_ops_exception_list,
+                preserve_ops=_preserve_ops,
             )
             self.check_valid_aten_op = self.aten_op_verifier.check_valid_op
 
             if self.check_edge_ops:
                 self.check_valid_op = self.check_valid_edge_op
             else:
                 self.check_valid_op = self.check_valid_aten_op
-            self._exception_list = exception_list if exception_list else []
 
         def allowed_getattr_types(self) -> Tuple[Type[Any], ...]:
             return (
@@ -258,7 +299,7 @@ def check_valid_edge_op(self, op):
                 in [operator.getitem]
                 + DISALLOW_LIST
                 + list(_EXECUTORCH_SYM_OPS)
-                + self._exception_list
+                + self._core_aten_ops_exception_list
             ):
                 return