use _preserve_ops for to_edge_transform_and_lower (#4273)

Summary:
Pull Request resolved: #4273

## Motivation
`run_decompositions()` has a new preserve_ops functionality which allows us to specify which ops we want to refrain from decomposing. This is super helpful for the to_edge_transform_and_lower api because it allows us to preserve decomposition that occur beyond the first level.

For example consider LSTM. when exported using torch.export, we see a torch.ops.aten.LSTM() operator in the graph. When running decompositions this is decomposed into linear, and then further decomposed into addmm. Since the linear op is produced from decomposing LSTM and does not exist until after we run_decompositions(), we can not perform our trick of changing the name space to prevent its decomposition. However, now using `_preserve_ops=(torch.ops.aten.linear.default,)` we are able to prevent this second layer decomposition.

## API Implementation Change
So in the implementation we do two passes. The first pass is we run_decompositions preserving all aten ops specified by our partitioners using `_preserve_ops`. On our second pass, we further filter which aten ops should be preserved by using the check_op_fn given to us by partitioners. We then use our namespace trick to prevent the decomposition of all aten ops which pass our check_op_fn.

## Testing Changes
To strengthen our tests, I first change the functionality of the NonDecompPartitioner. We partition only pre-decomp aten ops. And each of these ops live within their own delegate (this allows us to have a 1:1 mapping for call_delegate and pre_decomp aten nodes). In testing, this will allow us to ensure that the number of ops which are to preserved is correct by counting the number of delegates calls.

In testing we then count the number of aten ops which should correctly be preserved. And then check after the fact that all these ops are
1. No longer in the graph after to_edge_transform_and_lower
2. Each of these preserved ops are transformed into a call_delegate node

Reviewed By: tarun292

Differential Revision: D59786323

fbshipit-source-id: 7ea946e0d5afc8ebddd26913f6e843305116ad3b

exir/backend/test/op_partitioner_demo.py

@@ -4,6 +4,7 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
+import itertools
 from typing import Callable, Dict, final, List, Optional, Tuple
 
 import torch
@@ -24,6 +25,7 @@
 from executorch.exir.graph_module import get_control_flow_submodules
 from torch._export.utils import is_buffer, is_lifted_tensor_constant, is_param
 from torch.export import ExportedProgram
+from torch.fx.passes.infra.partitioner import Partition
 from torch.fx.passes.operator_support import any_chain, OperatorSupportBase
 
 
@@ -145,10 +147,12 @@ def is_node_supported(self, submodules, node: torch.fx.Node) -> bool:
 @final
 class NonDecompTestPartitioner(Partitioner):
     """
-    Partitions all add/mul nodes regardless of order
+    Non Decomp Test Partitioner, preserves aten ops from decomposition for delegate
+    consumption. Ensures that non_decomposed_edge_ops are all within their own delegate
     """
 
     def __init__(self) -> None:
+        self.supported_non_decomposed_edge_ops = edge_ops_non_decomposed
         self.op_support = any_chain(OpsToNotDecomposeOperatorSupport())
         self.delegation_spec = DelegationSpec(
             BackendWithCompilerDemo.__name__,
@@ -171,14 +175,29 @@ def filter_ops(node: torch.fx.Node) -> bool:
 
         return (ops_not_to_decompose, filter_ops)
 
+    def _generate_single_node_partition(
+        self, gm: torch.fx.GraphModule
+    ) -> List[Partition]:
+        partitions = []
+        partition_id = itertools.count()
+        nodes_seen = set()
+        for node in gm.graph.nodes:
+            if (
+                node.op == "call_function"
+                and node.target in self.supported_non_decomposed_edge_ops
+                and node not in nodes_seen
+            ):
+                partitions.append(Partition(nodes=[node], id=next(partition_id)))
+                nodes_seen.add(node)
+
+        return partitions
+
     def _partition_graph_module(
         self,
         graph_module: torch.fx.GraphModule,
     ) -> Dict[str, DelegationSpec]:
         partition_tags: Dict[str, DelegationSpec] = {}
-        partition_list = generate_pattern_op_partitions(
-            graph_module, op_support=self.op_support
-        )
+        partition_list = self._generate_single_node_partition(graph_module)
         for partition in partition_list:
             for node in partition.nodes:
                 delegation_tag = f"tag{partition.id}"

exir/program/_program.py

@@ -859,6 +859,66 @@ def _sanity_check_graph_for_non_decomp_ops(
                     logging.warning(warning_str)
 
 
+def _gen_edge_manager_for_partitioners(
+    partitioner: Dict[str, List[Partitioner]],
+    aten_programs: Dict[str, ExportedProgram],
+    config: EdgeCompileConfig,
+    constant_methods: Optional[Dict[str, Any]],
+) -> "EdgeProgramManager":
+    """
+    Generates EdgeProgramManager for subsequent lowering to the
+    partitioners specified by partitioner. The EdgeProgramManager is generated from
+    aten_programs.
+
+    Partitioners specify what nodes should not be decomposed from the original aten programs.
+    This is done through two passes of run_decompositions.
+        - First pass preserves all aten_targets specified by partitioners to preserve
+          them from nested decompositions
+        - Second pass uses check_op fn provided by partitioners to perform additional checks
+          on nodes with preserved aten targets. They are then replaces with transformed ops to
+          keep them through the second pass of decompositions
+    """
+    ops_set_to_not_decompose_by_program = {}
+    edge_programs: Dict[str, ExportedProgram] = {}
+    for name, program in aten_programs.items():
+        if partitioner is not None:
+            # preserve all ops listed by all partitioners first
+            all_ops_no_decomp = set()
+            for curr_partitioner in partitioner.get(name, []):
+                curr_ops_no_decomp, _ = curr_partitioner.ops_to_not_decompose(program)
+                all_ops_no_decomp |= set(curr_ops_no_decomp)
+
+            program = program.run_decompositions(
+                _default_decomposition_table(), _preserve_ops=tuple(all_ops_no_decomp)
+            )
+            # Among all the preserved aten ops, use the check_op_fn to do an additional
+            # check on which ops need to be preserved and which ops need to be decomposed
+            # Those which are truly preserved will be replaced with transformed ops
+            ops_set_to_not_decompose_by_program[name] = (
+                _replace_aten_ops_with_transformed_ops(name, program, partitioner)
+            )
+        program = program.run_decompositions(_default_decomposition_table())
+
+        _restore_transformed_ops_to_aten_ops(program)
+
+        edge_programs[name] = program
+
+        edge_programs[name] = _generate_edge_program(
+            name,
+            config,
+            program,
+            list(ops_set_to_not_decompose_by_program.get(name, [])),
+        )
+
+    edge_manager = EdgeProgramManager(
+        edge_programs,
+        constant_methods,
+        config,
+        list(set().union(*ops_set_to_not_decompose_by_program.values())),
+    )
+    return edge_manager
+
+
 def _to_edge_transform_and_lower(
     programs: Union[ExportedProgram, Dict[str, ExportedProgram]],
     transform_passes: Optional[
@@ -909,8 +969,6 @@ def _to_edge_transform_and_lower(
     Returns:
         EdgeProgramManager
     """
-    ops_set_to_not_decompose = set()
-
     assert not isinstance(constant_methods, EdgeCompileConfig)
     config = compile_config or EdgeCompileConfig()
     if not isinstance(programs, dict):
@@ -923,31 +981,8 @@ def _to_edge_transform_and_lower(
     else:
         partitioner = {}
 
-    ops_set_to_not_decompose_by_program = {}
-    edge_programs: Dict[str, ExportedProgram] = {}
-    for name, program in aten_programs.items():
-        if partitioner is not None:
-            ops_set_to_not_decompose_by_program[name] = (
-                _replace_aten_ops_with_transformed_ops(name, program, partitioner)
-            )
-        program = program.run_decompositions(_default_decomposition_table())
-
-        _restore_transformed_ops_to_aten_ops(program)
-
-        edge_programs[name] = program
-
-        edge_programs[name] = _generate_edge_program(
-            name,
-            config,
-            program,
-            list(ops_set_to_not_decompose_by_program.get(name, [])),
-        )
-
-    edge_manager = EdgeProgramManager(
-        edge_programs,
-        constant_methods,
-        config,
-        list(set().union(*ops_set_to_not_decompose_by_program.values())),
+    edge_manager = _gen_edge_manager_for_partitioners(
+        partitioner, aten_programs, config, constant_methods
     )
 
     if transform_passes is not None:

exir/program/test/test_program.py

@@ -6,7 +6,7 @@
 
 # pye-strict
 
-import operator
+import copy
 import unittest
 from typing import Any, Dict
 
@@ -27,6 +27,7 @@
     ExecutorchProgramManager,
     to_edge,
 )
+from executorch.exir.tracer import _default_decomposition_table
 from executorch.exir.verification.verifier import EXIREdgeDialectVerifier
 
 from executorch.extension.pybindings.portable_lib import (
@@ -102,6 +103,19 @@ def _get_random_inputs(cls):
         return (x,)
 
 
+class TestLSTM(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.lstm = torch.nn.LSTM(input_size=8, hidden_size=16, batch_first=True)
+
+    def forward(self, x):
+        return self.lstm(x)
+
+    @classmethod
+    def _get_random_inputs(cls):
+        return (torch.rand(1, 10, 8),)
+
+
 class WrapperModule(torch.nn.Module):
     def __init__(self, fn):
         super().__init__()
@@ -550,23 +564,65 @@ def _use_foo_add(a: torch.Tensor, b: torch.Tensor):
         except SpecViolationError:
             self.fail("Should not error out on custom op")
 
+    def get_num_nondecomposed_ops(self, ep, partitioner):
+        # count the number of aten ops that the partitioner can delegate
+        # we do this by running run_decompositions() with the preserved ops given
+        # to us by the partitioner. Then we count the number of preserved aten ops
+        # which pass the filter_ops fn given by the partitioner
+        reference_ep = copy.deepcopy(ep)
+        aten_ops_not_decomposed, filter_ops = partitioner.ops_to_not_decompose(ep)
+        reference_decomp_ep = reference_ep.run_decompositions(
+            decomp_table=_default_decomposition_table(),
+            _preserve_ops=tuple(aten_ops_not_decomposed),
+        )
+        num_non_decomposed_aten_ops = 0
+        for node in reference_decomp_ep.graph.nodes:
+            if (
+                node.op == "call_function"
+                and node.target in aten_ops_not_decomposed
+                and (filter_ops(node) if filter_ops else True)
+            ):
+                num_non_decomposed_aten_ops += 1
+        return num_non_decomposed_aten_ops
+
     def _test_model_with_non_decomp_partitioner(self, model: torch.nn.Module):
         # This is the pre-dispatch export that we will be switching to primarily
         # in the near future. The input to _to_edge_transform_and_lower needs to
         # be a graph generated by this pre dispatch export.
         ep = _export(model, model._get_random_inputs(), pre_dispatch=True)
+        non_decomp_partitioner = NonDecompTestPartitioner()
+
+        num_non_decomposed_aten_ops = self.get_num_nondecomposed_ops(
+            ep, non_decomp_partitioner
+        )
+
+        # run to_edge_trasnform_and_lower
         edge = _to_edge_transform_and_lower(
             ep,
             compile_config=EdgeCompileConfig(),
             partitioner=[NonDecompTestPartitioner()],
         )
+        # Check that non_decomposed_edge_ops are all consumed by the delegate
+        non_decomposed_edge_ops = (
+            non_decomp_partitioner.supported_non_decomposed_edge_ops
+        )
+        for node in edge.exported_program().graph.nodes:
+            if node.op == "call_function":
+                self.assertTrue(node.target not in non_decomposed_edge_ops)
+
+        # check that the number of call_delegate_nodes is equal to the number of
+        # non_decomposed_aten_ops we found above
+        num_call_delegates = 0
         for node in edge.exported_program().graph_module.graph.nodes:
             # There should only be a single call_function node in the graph
             # and that should be a call_delegate node.
-            if node.op == "call_function" and node.target != operator.getitem:
-                self.assertEqual(
-                    node.target, torch.ops.higher_order.executorch_call_delegate
-                )
+            if (
+                node.op == "call_function"
+                and node.target == torch.ops.higher_order.executorch_call_delegate
+            ):
+                num_call_delegates += 1
+
+        self.assertEqual(num_call_delegates, num_non_decomposed_aten_ops)
 
     def test_to_edge_transform_and_lower(self):
         self._test_model_with_non_decomp_partitioner(TestLinear())
@@ -577,6 +633,8 @@ def test_to_edge_transform_and_lower(self):
 
         self._test_model_with_non_decomp_partitioner(TestUpsample())
 
+        self._test_model_with_non_decomp_partitioner(TestLSTM())
+
     def test_to_edge_transform_and_lower_with_exception(self):
         class TestLinear(torch.nn.Module):
             def __init__(self):