[ExecuTorch][to_backend] Enable passing Delegation Spec to to_backend

Pull Request resolved: #8165

This will be used for the backend weight sharing so backends which do entire graph delegation can still share data across methods.

Support Entire Graph Delegation Flow through EdgeProgramManager's to_backend.

### Motivation
A current usecase for backend lowering is through the `to_backend(backend_id, exported_program, compile_spec)` API which lowers the entire exported program to the specified backend_id. However, lowering via the EdgeProgramManager only allows for partitioner based lowering. the EdgeProgramManager is the main component which enables support for multiple methods, as a result backends which leverage the old `to_backend(backend_id, ...)` api can not export executorch models with multiple methods.

### Design
We override EdgeProgramManager to also allow Partitioner to be replaceable by DelegationSpec. DelegationSpec is essentially a wrapper around the backend_id and the compile_spec, so any where a partitioenr is specified to lower a graph, the delegation spec can also be used to do entier graph lowering.

### Intended Flow
```
del_spec = DelegationSpec("BackendWithCompilerDemo", [CompileSpec(...)])
encode_graph = torch.export.export(Encoder(), sample_inputs)
decode_graph = torch.export.export(Decoder(), sample_inputs)
edge_manager = to_edge({
    "encode": encode_graph,
    "decode": decode_graph,
})
lowered_edge_manager = edge_manager.to_backend(del_spec)
# or if you want to specify which methods to lower to with del_spec
lowered_edge_manager= edge_manager.to_backend({
    "encode": del_spec,
})
```
ghstack-source-id: 266326224
@exported-using-ghexport

Differential Revision: [D69086565](https://our.internmc.facebook.com/intern/diff/D69086565/)

Author: mcr229

exir/backend/canonical_partitioners/TARGETS

@@ -7,6 +7,7 @@ runtime.python_library(
     srcs = [
         "duplicate_dequant_node_pass.py",
         "pattern_op_partitioner.py",
+        "all_node_partitioner.py",
     ],
     visibility = [
         "//executorch/...",

exir/backend/canonical_partitioners/all_node_partitioner.py

@@ -0,0 +1,54 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+from typing import Dict, List
+
+import torch
+from executorch.exir.backend.backend_details import ExportedProgram
+from executorch.exir.backend.compile_spec_schema import CompileSpec
+from executorch.exir.backend.partitioner import (
+    DelegationSpec,
+    Partitioner,
+    PartitionResult,
+)
+from torch._export.utils import is_buffer, is_lifted_tensor_constant, is_param
+
+def is_non_tensor_placeholder(node: torch.fx.Node, ep: ExportedProgram) -> bool:
+    """
+    Returns true if the node is a placeholder node and it is not a tensor
+    """
+    return node.op == "placeholder" and not (
+        is_param(ep, node) or is_buffer(ep, node) or is_lifted_tensor_constant(ep, node)
+    )
+
+
+class AllNodePartitioner(Partitioner):
+    def __init__(
+        self,
+        backend_id: str,
+        compile_specs: List[CompileSpec],
+    ):
+        """
+        Partitioner that lowers every single node in the graph module to the
+        specified backend_id
+        """
+        super().__init__()
+        self.delegation_spec = DelegationSpec(backend_id, compile_specs)
+
+    def partition(self, exported_program: ExportedProgram) -> PartitionResult:
+        # tag all nodes
+        partition_tags: Dict[str, DelegationSpec] = {}
+        for node in exported_program.graph_module.graph.nodes:
+            if is_non_tensor_placeholder(node, exported_program) or node.op == "output":
+                continue
+
+            delegation_tag = self.delegation_spec.backend_id
+            node.meta["delegation_tag"] = delegation_tag
+            partition_tags[delegation_tag] = self.delegation_spec
+
+        return PartitionResult(
+            tagged_exported_program=exported_program, partition_tags=partition_tags
+        )

exir/backend/test/test_backends.py

@@ -10,7 +10,12 @@
 
 import executorch.exir as exir
 import torch
+from executorch.exir import to_edge
 from executorch.exir.backend.backend_api import LoweredBackendModule, to_backend
+from executorch.exir.backend.canonical_partitioners.all_node_partitioner import 
+(
+    AllNodePartitioner,
+)
 from executorch.exir.backend.compile_spec_schema import CompileSpec
 from executorch.exir.backend.partitioner import (
     DelegationSpec,
@@ -1266,3 +1271,179 @@ def forward(self, x: List[torch.Tensor]):
 
         gm = exir.capture(ComposedM(), inputs, exir.CaptureConfig()).to_edge()
         gm(*inputs)
+
+    def test_to_backend_delegation_spec(self):
+        class SinModule(torch.nn.Module):
+            def __init__(self):
+                super().__init__()
+
+            def forward(self, x):
+                return [torch.sin(x)]
+            
+
+        sin_module = SinModule()
+        model_inputs = (torch.ones(1),)
+        max_value = model_inputs[0].shape[0]
+
+        partitioner = AllNodePartitioner(
+            "BackendWithCompilerDemo", [CompileSpec("max_value", bytes([max_value]))]
+        )
+
+        edgeir_m = to_edge(torch.export.export(sin_module, model_inputs))
+        edgeir_m = edgeir_m.to_backend(partitioner)
+        exec_prog = edgeir_m.to_executorch()
+        graph_module = exec_prog.exported_program().graph_module
+        # Check that there is not an aten.sin node.
+        self.assertTrue(
+            exir_ops.edge.aten.sin
+            not in {node.target for node in graph_module.graph.nodes}
+        )
+
+        # Check that there exists a call_delegate, representing the call to the
+        # delegated function
+        FileCheck().check("torch.ops.higher_order.executorch_call_delegate").run(
+            graph_module.code
+        )
+        lowered_submodules = get_lowered_submodules(graph_module)
+        self.assertEqual(len(lowered_submodules), 1)
+
+        for node in graph_module.graph.nodes:
+            if node.op == "call_function" and node.target == executorch_call_delegate:
+                # Check that first arg is lowered_module_{unique_id}
+                self.assertEqual(node.args[0].target, "lowered_module_0")
+
+        program = exec_prog.executorch_program
+
+        # Check the program can be printed
+        print_program(program)
+
+        # Check the backend delegate
+        self.check_backend_delegate(
+            program=program,
+            delegate=program.execution_plan[0].delegates[0],
+            expected_id=BackendWithCompilerDemo.__name__,
+            expected_processed=b"1version:0#op:demo::aten.sin.default, numel:1, dtype:torch.float32<debug_handle>2#",
+        )
+
+        # Check the delegate instruction
+        self.assertTrue(
+            isinstance(
+                program.execution_plan[0].chains[0].instructions[0].instr_args,
+                DelegateCall,
+            )
+        )
+        buff = exec_prog.buffer
+
+        executorch_module = _load_for_executorch_from_buffer(buff)
+        model_inputs = torch.ones(1)
+        model_outputs = executorch_module.forward([model_inputs])
+        self.assertEqual(
+            model_inputs,
+            torch.ones(1),
+        )
+        expected_output = 0.8333 * torch.ones(1)
+
+        self.assertTrue(
+            torch.allclose(model_outputs[0], expected_output, atol=1e-03, rtol=1e-03)
+        )
+
+    def test_to_backend_multimethod_delegation_spec(self):
+        class SinModule(torch.nn.Module):
+            def __init__(self):
+                super().__init__()
+
+            def forward(self, x):
+                return torch.sin(x)
+ 
+            def inputs(self):
+                return (torch.ones(1),)
+
+        class AddMulModule(torch.nn.Module):
+            def __init__(self):
+                super().__init__()
+
+            def forward(self, a, x, b):
+                y = torch.mm(a, x)
+                z = torch.add(y, b)
+                return z
+ 
+            def inputs(self):
+                return (torch.ones(2, 2), 2 * torch.ones(2, 2), 3 * torch.ones(2, 2))
+
+        sin_module = SinModule()
+        max_value_sin = sin_module.inputs()[0].shape[0]
+        sin_partitioner = AllNodePartitioner(
+            "BackendWithCompilerDemo", 
+            [CompileSpec("max_value", bytes([max_value_sin]))],
+        )
+
+        add_mul_module = AddMulModule()
+        max_value_add_mul = add_mul_module.inputs()[0].shape[0]
+        add_mul_partitioner = AllNodePartitioner(
+            "BackendWithCompilerDemo", 
+            [CompileSpec("max_value", bytes([max_value_add_mul]))],
+        )
+
+        edgeir_m = to_edge(
+            {
+                "sin": torch.export.export(sin_module, sin_module.inputs()),
+                "add_mul": torch.export.export(add_mul_module, add_mul_module.inputs()),
+            }
+        )
+        edgeir_m = edgeir_m.to_backend(
+            {
+                "sin": sin_partitioner,
+                "add_mul": add_mul_partitioner,
+            }
+        )
+        exec_prog = edgeir_m.to_executorch()
+
+        for method_name in ["sin", "add_mul"]:
+            graph_module = exec_prog.exported_program(method_name).graph_module
+            # Check delegated nodes are gone
+            self.assertTrue(
+                exir_ops.edge.aten.sin
+                not in {node.target for node in graph_module.graph.nodes}
+            )
+            self.assertTrue(
+                exir_ops.edge.aten.add
+                not in {node.target for node in graph_module.graph.nodes}
+            )
+            self.assertTrue(
+                exir_ops.edge.aten.mm
+                not in {node.target for node in graph_module.graph.nodes}
+            )
+            # Check that there exists a call_delegate, representing the call to the
+            # delegated function
+            FileCheck().check("torch.ops.higher_order.executorch_call_delegate").run(
+                graph_module.code
+            )
+            lowered_submodules = get_lowered_submodules(graph_module)
+            self.assertEqual(len(lowered_submodules), 1)
+
+        program = exec_prog.executorch_program
+
+        # Check the program can be printed
+        print_program(program)
+
+        buff = exec_prog.buffer
+
+        executorch_module = _load_for_executorch_from_buffer(buff)
+
+        for method_name, module in {
+            "sin": sin_module, 
+            "add_mul": add_mul_module
+        }.items():
+            inputs_flattened, _ = tree_flatten(module.inputs())
+            model_outputs = executorch_module.run_method(
+                method_name, tuple(inputs_flattened)
+            )
+
+            if method_name == "sin":
+                # backend with compiler demo does a taylor approximation of sin
+                ref_output = 0.8333 * torch.ones(1)
+            else:
+                ref_output = module(*module.inputs())
+            self.assertTrue(
+                torch.allclose(model_outputs[0], ref_output, atol=1e-03, rtol=1e-03)
+            )

exir/backend/test/test_backends_lifted.py

@@ -11,6 +11,10 @@
 import torch
 from executorch.exir import to_edge
 from executorch.exir.backend.backend_api import LoweredBackendModule, to_backend
+from executorch.exir.backend.canonical_partitioners.all_node_partitioner import 
+(
+    AllNodePartitioner,
+)
 from executorch.exir.backend.compile_spec_schema import CompileSpec
 from executorch.exir.backend.partitioner import (
     DelegationSpec,
@@ -138,6 +142,18 @@ def forward(self, x):
 
         self.assertTrue(torch.allclose(new_res, expected_res))
 
+        # Test same flow but through edge_program_manager
+        edgeir_m = to_edge(export(sin_module, model_inputs, strict=True))
+        loweredir_m = edgeir_m.to_backend(
+            AllNodePartitioner(BackendWithCompilerDemo.__name__, [])
+        )
+        lowered_sin_module = get_lowered_submodules(
+            loweredir_m.exported_program().graph_module
+        )[0][1]
+
+        new_res = lowered_sin_module(*model_inputs)[0]
+
+        self.assertTrue(torch.allclose(new_res, expected_res))
         # TODO(tkaruturi): emitting single LoweredBackendModule
         # program = to_edge(export(graph_module)).to_exectorch()._emitter_output.program
 

exir/backend/test/test_compatibility.py

@@ -10,6 +10,10 @@
 from executorch.exir import to_edge
 from executorch.exir._serialize import _serialize_pte_binary
 from executorch.exir.backend.backend_api import to_backend
+from executorch.exir.backend.canonical_partitioners.all_node_partitioner import 
+(
+    AllNodePartitioner,
+)
 from executorch.exir.backend.compile_spec_schema import CompileSpec
 from executorch.exir.backend.test.backend_with_compiler_demo import (
     BackendWithCompilerDemo,
@@ -65,3 +69,49 @@ def forward(self, x):
             "loading method forward failed with error 0x30",
         ):
             executorch_module = _load_for_executorch_from_buffer(buff)
+
+    def test_compatibility_in_runtime_edge_program_manager(self):
+        class SinModule(torch.nn.Module):
+            def __init__(self):
+                super().__init__()
+
+            def forward(self, x):
+                return torch.sin(x)
+
+        sin_module = SinModule()
+        model_inputs = (torch.ones(1),)
+        edgeir_m = to_edge(export(sin_module, model_inputs, strict=True))
+        max_value = model_inputs[0].shape[0]
+        compile_specs = [CompileSpec("max_value", bytes([max_value]))]
+        lowered_edge_irm = edgeir_m.to_backend(
+            AllNodePartitioner("BackendWithCompilerDemo", compile_specs)
+        )
+        exec_prog = lowered_edge_irm.to_executorch()
+
+        buff = exec_prog.buffer
+
+        # The demo backend works well
+        executorch_module = _load_for_executorch_from_buffer(buff)
+        model_inputs = torch.ones(1)
+        _ = executorch_module.forward([model_inputs])
+
+        prog = exec_prog.executorch_program
+        # Rewrite the delegate version number from 0 to 1.
+        prog.backend_delegate_data[0].data = bytes(
+            "1version:1#op:demo::aten.sin.default, numel:1, dtype:torch.float32<debug_handle>1#",
+            encoding="utf8",
+        )
+
+        # Generate the .pte file with the wrong version.
+        buff = bytes(
+            _serialize_pte_binary(
+                program=prog,
+            )
+        )
+
+        # Throw runtime error with error code 0x30, meaning delegate is incompatible.
+        with self.assertRaisesRegex(
+            RuntimeError,
+            "loading method forward failed with error 0x30",
+        ):
+            executorch_module = _load_for_executorch_from_buffer(buff)

exir/program/TARGETS

@@ -31,6 +31,7 @@ python_library(
         "//executorch/exir/_serialize:lib",
         "//executorch/exir/backend:backend_api",
         "//executorch/exir/backend:partitioner",
+        "//executorch/exir/backend/canonical_partitioners:canonical_partitioner_lib",
         "//executorch/exir/capture:config",
         "//executorch/exir/emit:emit",
         "//executorch/exir/emit:lib",