NXP backend: Update aot_neutron_compile.py example pipeline (#14142)

### Summary
Updates example in `aot_neutron_compile.py` to reflect current state of
NXP backend. RemoveIOQuantOpsPass call is moved to NeutronEdgePassManager. `export_to_edge()` is replaced by `to_edge_transform_and_lower()` call.

### Test plan
AOT examples should run automatically in CI. You can manually test it
using `backends/nxp/run_aot_example.sh`

cc @digantdesai @JakeStevens @robert-kalmar @MartinPavella
@roman-janik-nxp

---------

Co-authored-by: Roman Janik <[email protected]>

Author: StrycekSimon

backends/nxp/edge_passes/neutron_edge_pass_manager.py

@@ -10,6 +10,10 @@
     MoveTrailingAuxiliaryOperatorIntoSeparateQDQClusterPass,
 )
 from executorch.backends.nxp.edge_passes.neutron_edge_pass import NeutronEdgePass
+
+from executorch.backends.nxp.edge_passes.remove_io_quant_ops_pass import (
+    RemoveIOQuantOpsPass,
+)
 from executorch.exir import EdgeProgramManager
 from executorch.exir.program._program import (
     _get_updated_graph_signature,
@@ -24,7 +28,9 @@
 
 class NeutronEdgePassManager(PassManager):
 
-    def __init__(self, passes: list[NeutronEdgePass] = None):
+    def __init__(
+        self, passes: list[NeutronEdgePass] = None, remove_io_quant_ops: bool = False
+    ):
         passes: list[NeutronEdgePass] = passes or [
             MoveLeadingAuxiliaryOperatorIntoSeparateQDQClusterPass(),
             MoveTrailingAuxiliaryOperatorIntoSeparateQDQClusterPass(),
@@ -35,6 +41,8 @@ def __init__(self, passes: list[NeutronEdgePass] = None):
             steps=10,  # Empirical value. At most 10 cycles of passes will be run.
         )
 
+        self.remove_io_quant_ops = remove_io_quant_ops
+
     def _transform_graph_module(self, module: nn.Module) -> PassResult:
         """Apply the passes to a single graph module."""
         pass_result: PassResult = super().__call__(module)
@@ -78,12 +86,17 @@ def __call__(self, epm: EdgeProgramManager) -> EdgeProgramManager:
 
             new_programs[name] = new_program
 
-        if len(new_programs) == 0:
-            # No passes were run, return the old EdgeProgramManager.
-            return epm
+        result = epm
 
-        else:
-            # Return a new EdgeProgramManager with the updated programs.
-            return EdgeProgramManager(
+        if len(new_programs) > 0:
+            # Use a new EdgeProgramManager with the updated programs if any update was performed.
+            result = EdgeProgramManager(
                 new_programs, copy.deepcopy(epm._config_methods), epm.compile_config
             )
+
+        if self.remove_io_quant_ops:
+            result = result.transform(
+                [RemoveIOQuantOpsPass(edge_program_manager=result)]
+            )
+
+        return result

backends/nxp/tests/executorch_pipeline.py

@@ -15,9 +15,6 @@
 from executorch.backends.nxp.edge_passes.neutron_edge_pass_manager import (
     NeutronEdgePassManager,
 )
-from executorch.backends.nxp.edge_passes.remove_io_quant_ops_pass import (
-    RemoveIOQuantOpsPass,
-)
 from executorch.backends.nxp.neutron_partitioner import NeutronPartitioner
 from executorch.backends.nxp.nxp_backend import generate_neutron_compile_spec
 from executorch.backends.nxp.quantizer.neutron_quantizer import NeutronQuantizer
@@ -115,7 +112,9 @@ def to_quantized_edge_program(
         edge_compile_config=edge_compile_config,
     )
 
-    edge_program_manager = NeutronEdgePassManager()(edge_program_manager)
+    edge_program_manager = NeutronEdgePassManager(
+        remove_io_quant_ops=remove_quant_io_ops
+    )(edge_program_manager)
 
     compile_spec = generate_neutron_compile_spec(
         target,
@@ -125,11 +124,6 @@ def to_quantized_edge_program(
     partitioner = NeutronPartitioner(compile_spec, custom_delegation_options)
     edge_program_manager = edge_program_manager.to_backend(partitioner)
 
-    if remove_quant_io_ops:
-        edge_program_manager = edge_program_manager.transform(
-            [RemoveIOQuantOpsPass(edge_program_manager=edge_program_manager)]
-        )
-
     return edge_program_manager
 
 

examples/nxp/aot_neutron_compile.py

@@ -15,8 +15,8 @@
 import executorch.kernels.quantized  # noqa F401
 
 import torch
-from executorch.backends.nxp.edge_passes.remove_io_quant_ops_pass import (
-    RemoveIOQuantOpsPass,
+from executorch.backends.nxp.edge_passes.neutron_edge_pass_manager import (
+    NeutronEdgePassManager,
 )
 from executorch.backends.nxp.neutron_partitioner import NeutronPartitioner
 from executorch.backends.nxp.nxp_backend import generate_neutron_compile_spec
@@ -33,7 +33,6 @@
 from torchao.quantization.pt2e.quantize_pt2e import convert_pt2e, prepare_pt2e
 
 from .experimental.cifar_net.cifar_net import CifarNet, test_cifarnet_model
-
 from .models.mobilenet_v2 import MobilenetV2
 
 FORMAT = "[%(levelname)s %(asctime)s %(filename)s:%(lineno)s] %(message)s"
@@ -228,7 +227,7 @@ def _get_batch_size(data):
 
     module = exported_program.module()
 
-    # 4. Quantize if required
+    # 3. Quantize if required
     if args.quantize:
         if calibration_inputs is None:
             logging.warning(
@@ -254,39 +253,30 @@ def _get_batch_size(data):
         quantized_str = "quantized " if args.quantize else ""
         print(f"\nAccuracy of the {quantized_str}`{args.model_name}`: {accuracy}\n")
 
-    # 5. Export to edge program
-    partitioner_list = []
-    if args.delegate is True:
-        partitioner_list = [
-            NeutronPartitioner(
-                generate_neutron_compile_spec(
-                    args.target,
-                    args.neutron_converter_flavor,
-                    operators_not_to_delegate=args.operators_not_to_delegate,
-                )
-            )
-        ]
+    # 4. Transform and lower
+
+    compile_spec = generate_neutron_compile_spec(
+        args.target,
+        operators_not_to_delegate=args.operators_not_to_delegate,
+        neutron_converter_flavor=args.neutron_converter_flavor,
+    )
+    partitioners = [NeutronPartitioner(compile_spec)] if args.delegate else []
 
-    edge_program = to_edge_transform_and_lower(
+    edge_program_manager = to_edge_transform_and_lower(
         export(module, example_inputs, strict=True),
-        partitioner=partitioner_list,
-        compile_config=EdgeCompileConfig(
-            _check_ir_validity=False,
-        ),
+        partitioner=partitioners,
+        compile_config=EdgeCompileConfig(),
     )
-    logging.debug(f"Exported graph:\n{edge_program.exported_program().graph}")
 
-    if args.remove_quant_io_ops:
-        edge_program = edge_program.transform(
-            [RemoveIOQuantOpsPass(edge_program_manager=edge_program)]
-        )
-        logging.debug(
-            f"Exported graph (RemoveIOQuantOpsPass):\n{edge_program.exported_program().graph}"
-        )
+    edge_program_manager = NeutronEdgePassManager(
+        remove_io_quant_ops=args.remove_quant_io_ops
+    )(edge_program_manager)
+
+    logging.debug(f"Lowered graph:\n{edge_program_manager.exported_program().graph}")
 
-    # 6. Export to ExecuTorch program
+    # 5. Export to ExecuTorch program
     try:
-        exec_prog = edge_program.to_executorch(
+        exec_prog = edge_program_manager.to_executorch(
             config=ExecutorchBackendConfig(extract_delegate_segments=False)
         )
     except RuntimeError as e:
@@ -306,7 +296,7 @@ def executorch_program_to_str(ep, verbose=False):
 
     logging.debug(f"Executorch program:\n{executorch_program_to_str(exec_prog)}")
 
-    # 7. Serialize to *.pte
+    # 6. Serialize to *.pte
     model_name = f"{args.model_name}" + (
         "_nxp_delegate" if args.delegate is True else ""
     )