(WIP) Refactor XNNPACK tester to extract base class (#11596)

Summary:
Refactor the XNNPACK tester to split out reusable base components from XNNPACK-specific parts. I've relocated the base classes to backends/test/harness.

I've kept the tester structure pretty much unchanged, except for replacing stage names with an enum.

It looks like Arm tests are directly importing for XNNPACK's tester currently. Ideally, we'll want to refactor to have their own stage implementations, but I've left that as a follow-up to minimize changes for the initial refactor.

Pull Request resolved: #11596

Test Plan:
CI

Rollback Plan:

Differential Revision: D76547310

Pulled By: GregoryComer

Author: GregoryComer

backends/arm/test/misc/test_lifted_tensor.py

@@ -12,6 +12,7 @@
     TosaPipelineBI,
     TosaPipelineMI,
 )
+from executorch.backends.test.harness.stages import StageType
 from executorch.backends.xnnpack.test.tester import ToEdge
 
 
@@ -72,9 +73,8 @@ def test_partition_lifted_tensor_tosa_MI(test_data: input_t1):
         use_to_edge_transform_and_lower=False,
     )
     pipeline.run()
-    to_edge_stage_name = pipeline.tester.stage_name(ToEdge)
     signature = (
-        pipeline.tester.stages[to_edge_stage_name]
+        pipeline.tester.stages[StageType.TO_EDGE]
         .artifact.exported_program()
         .graph_signature
     )
@@ -94,9 +94,8 @@ def test_partition_lifted_tensor_tosa_BI(test_data: input_t1):
         use_to_edge_transform_and_lower=False,
     )
     pipeline.run()
-    to_edge_stage_name = pipeline.tester.stage_name(ToEdge)
     signature = (
-        pipeline.tester.stages[to_edge_stage_name]
+        pipeline.tester.stages[StageType.TO_EDGE]
         .artifact.exported_program()
         .graph_signature
     )

backends/arm/test/tester/analyze_output_utils.py

@@ -14,6 +14,7 @@
     get_output_quantization_params,
 )
 
+from executorch.backends.test.harness.stages import StageType
 from executorch.backends.xnnpack.test.tester.tester import Export, Quantize
 
 logger = logging.getLogger(__name__)
@@ -236,8 +237,8 @@ def dump_error_output(
     if path_to_tosa_files is None:
         path_to_tosa_files = tempfile.mkdtemp(prefix="executorch_result_dump_")
 
-    export_stage = tester.stages.get(tester.stage_name(Export), None)
-    quantize_stage = tester.stages.get(tester.stage_name(Quantize), None)
+    export_stage = tester.stages.get(StageType.EXPORT, None)
+    quantize_stage = tester.stages.get(StageType.QUANTIZE, None)
     if export_stage is not None and quantize_stage is not None:
         output_nodes = get_output_nodes(export_stage.artifact)
         qp_input = get_input_quantization_params(export_stage.artifact)

backends/arm/test/tester/arm_tester.py

@@ -50,6 +50,7 @@
 from executorch.backends.arm.tosa_partitioner import TOSAPartitioner
 from executorch.backends.arm.tosa_specification import TosaSpecification
 
+from executorch.backends.test.harness.stages import Stage, StageType
 from executorch.backends.xnnpack.test.tester import Tester
 from executorch.devtools.backend_debug import get_delegation_info
 
@@ -242,9 +243,12 @@ def wrapped_ep_pass(ep: ExportedProgram) -> ExportedProgram:
         super().run(artifact, inputs)
 
 
-class InitialModel(tester.Stage):
+class InitialModel(Stage):
     def __init__(self, model: torch.nn.Module):
         self.model = model
+    
+    def stage_type(self) -> StageType:
+        return StageType.INITIAL_MODEL
 
     def run(self, artifact, inputs=None) -> None:
         pass
@@ -284,13 +288,13 @@ def __init__(
         self.constant_methods = constant_methods
         self.compile_spec = compile_spec
         super().__init__(model, example_inputs, dynamic_shapes)
-        self.pipeline[self.stage_name(InitialModel)] = [
-            self.stage_name(tester.Quantize),
-            self.stage_name(tester.Export),
+        self.pipeline[StageType.INITIAL_MODEL] = [
+            StageType.QUANTIZE,
+            StageType.EXPORT,
         ]
 
         # Initial model needs to be set as a *possible* but not yet added Stage, therefore add None entry.
-        self.stages[self.stage_name(InitialModel)] = None
+        self.stages[StageType.INITIAL_MODEL] = None
         self._run_stage(InitialModel(self.original_module))
 
     def quantize(self, quantize_stage: Optional[tester.Quantize] = None):
@@ -385,7 +389,7 @@ def serialize(
         return super().serialize(serialize_stage)
 
     def is_quantized(self) -> bool:
-        return self.stages[self.stage_name(tester.Quantize)] is not None
+        return self.stages[StageType.QUANTIZE] is not None
 
     def run_method_and_compare_outputs(
         self,
@@ -414,18 +418,16 @@ def run_method_and_compare_outputs(
         """
 
         if not run_eager_mode:
-            edge_stage = self.stages[self.stage_name(tester.ToEdge)]
+            edge_stage = self.stages[StageType.TO_EDGE]
             if edge_stage is None:
-                edge_stage = self.stages[
-                    self.stage_name(tester.ToEdgeTransformAndLower)
-                ]
+                edge_stage = self.stages[StageType.TO_EDGE_TRANSFORM_AND_LOWER]
             assert (
                 edge_stage is not None
             ), "To compare outputs, at least the ToEdge or ToEdgeTransformAndLower stage needs to be run."
         else:
             # Run models in eager mode. We do this when we want to check that the passes
             # are numerically accurate and the exported graph is correct.
-            export_stage = self.stages[self.stage_name(tester.Export)]
+            export_stage = self.stages[StageType.EXPORT]
             assert (
                 export_stage is not None
             ), "To compare outputs in eager mode, the model must be at Export stage"
@@ -435,11 +437,11 @@ def run_method_and_compare_outputs(
         is_quantized = self.is_quantized()
 
         if is_quantized:
-            reference_stage = self.stages[self.stage_name(tester.Quantize)]
+            reference_stage = self.stages[StageType.QUANTIZE]
         else:
-            reference_stage = self.stages[self.stage_name(InitialModel)]
+            reference_stage = self.stages[StageType.INITIAL_MODEL]
 
-        exported_program = self.stages[self.stage_name(tester.Export)].artifact
+        exported_program = self.stages[StageType.EXPORT].artifact
         output_nodes = get_output_nodes(exported_program)
 
         output_qparams = get_output_quantization_params(output_nodes)
@@ -449,7 +451,7 @@ def run_method_and_compare_outputs(
             quantization_scales.append(getattr(output_qparams[node], "scale", None))
 
         logger.info(
-            f"Comparing Stage '{self.stage_name(test_stage)}' with Stage '{self.stage_name(reference_stage)}'"
+            f"Comparing Stage '{test_stage.stage_type()}' with Stage '{reference_stage.stage_type()}'"
         )
 
         # Loop inputs and compare reference stage with the compared stage.
@@ -500,14 +502,12 @@ def get_graph(self, stage: str | None = None) -> Graph:
             stage = self.cur
         artifact = self.get_artifact(stage)
         if (
-            self.cur == self.stage_name(tester.ToEdge)
-            or self.cur == self.stage_name(Partition)
-            or self.cur == self.stage_name(ToEdgeTransformAndLower)
+            self.cur == StageType.TO_EDGE
+            or self.cur == StageType.PARTITION
+            or self.cur == StageType.TO_EDGE_TRANSFORM_AND_LOWER
         ):
             graph = artifact.exported_program().graph
-        elif self.cur == self.stage_name(tester.Export) or self.cur == self.stage_name(
-            tester.Quantize
-        ):
+        elif self.cur == StageType.EXPORT or self.cur == StageType.QUANTIZE:
             graph = artifact.graph
         else:
             raise RuntimeError(
@@ -533,8 +533,8 @@ def dump_operator_distribution(
         if (
             self.cur
             in (
-                self.stage_name(tester.Partition),
-                self.stage_name(ToEdgeTransformAndLower),
+                StageType.PARTITION,
+                StageType.TO_EDGE_TRANSFORM_AND_LOWER,
             )
             and print_table
         ):
@@ -625,7 +625,7 @@ def run_transform_for_annotation_pipeline(
             stage = self.cur
         # We need to clone the artifact in order to ensure that the state_dict is preserved after passes are run.
         artifact = self.get_artifact(stage)
-        if self.cur == self.stage_name(tester.Export):
+        if self.cur == StageType.EXPORT:
             new_gm = ArmPassManager(get_tosa_spec(self.compile_spec)).transform_for_annotation_pipeline(  # type: ignore[arg-type]
                 graph_module=artifact.graph_module
             )

backends/test/harness/TARGETS

@@ -0,0 +1,18 @@
+load("@fbsource//xplat/executorch/build:runtime_wrapper.bzl", "runtime")
+
+oncall("executorch")
+
+runtime.python_library(
+    name = "tester",
+    srcs = [
+        "__init__.py",
+        "tester.py",
+    ] + native.glob(["stages/*.py"]),
+    visibility = [
+        "//executorch/...",
+        "@EXECUTORCH_CLIENTS",
+    ],
+    deps = [
+        "//executorch/exir:graph_module",
+    ],
+)

backends/test/harness/__init__.py

@@ -0,0 +1 @@
+from .tester import Tester

backends/test/harness/stages/__init__.py

@@ -0,0 +1,10 @@
+from .export import Export
+from .partition import Partition
+from .quantize import Quantize
+from .run_passes import RunPasses
+from .serialize import Serialize
+from .stage import Stage, StageType
+from .to_edge import ToEdge
+from .to_edge_transform_and_lower import ToEdgeTransformAndLower
+from .to_executorch import ToExecutorch
+

backends/test/harness/stages/export.py

@@ -0,0 +1,32 @@
+from typing import Any, Optional, Sequence, Tuple
+
+import torch
+
+from executorch.backends.test.harness.stages.stage import Stage, StageType
+from torch.export import export, ExportedProgram
+
+class Export(Stage):
+    def __init__(self, dynamic_shapes: Optional[Tuple[Any]] = None):
+        self.exported_program = None
+        self.dynamic_shapes = dynamic_shapes
+
+    def stage_type(self) -> StageType:
+        return StageType.EXPORT
+
+    def run(
+        self,
+        artifact: torch.nn.Module,
+        inputs: Tuple[torch.Tensor],
+    ) -> None:
+        self.exported_program = export(
+            artifact, inputs, dynamic_shapes=self.dynamic_shapes, strict=True
+        )
+
+    @property
+    def artifact(self) -> ExportedProgram:
+        return self.exported_program
+
+    @property
+    def graph_module(self) -> str:
+        return self.exported_program.graph_module
+

backends/test/harness/stages/partition.py

@@ -0,0 +1,27 @@
+from executorch.backends.test.harness.stages.stage import Stage, StageType
+from executorch.exir import (
+    EdgeProgramManager,
+)
+from executorch.exir.backend.backend_api import validation_disabled
+from executorch.exir.backend.partitioner import Partitioner
+
+class Partition(Stage):
+    def __init__(self, partitioner: Partitioner):
+        self.partitioner = partitioner
+        self.delegate_module = None
+    
+    def stage_type(self) -> StageType:
+        return StageType.PARTITION
+
+    def run(self, artifact: EdgeProgramManager, inputs=None):
+        with validation_disabled():
+            self.delegate_module = artifact
+            self.delegate_module = self.delegate_module.to_backend(self.partitioner)
+
+    @property
+    def artifact(self) -> EdgeProgramManager:
+        return self.delegate_module
+
+    @property
+    def graph_module(self) -> str:
+        return self.delegate_module.exported_program().graph_module

backends/test/harness/stages/quantize.py

@@ -0,0 +1,78 @@
+from typing import Any, Optional, Sequence, Tuple
+
+import torch
+
+from executorch.backends.test.harness.stages.stage import Stage, StageType
+from executorch.backends.transforms.duplicate_dynamic_quant_chain import (
+    DuplicateDynamicQuantChainPass,
+)
+
+from torch.export import export_for_training
+
+from torchao.quantization.pt2e.quantize_pt2e import (
+    convert_pt2e,
+    prepare_pt2e,
+    prepare_qat_pt2e,
+)
+from torchao.quantization.pt2e.quantizer import Quantizer
+
+class Quantize(Stage):
+    def __init__(
+        self,
+        quantizer: Optional[Quantizer] = None,
+        quantization_config: Optional[Any] = None,
+        calibrate: bool = True,
+        calibration_samples: Optional[Sequence[Any]] = None,
+        is_qat: Optional[bool] = False,
+    ):
+        self.quantizer = quantizer
+        self.quantization_config = quantization_config
+        self.calibrate = calibrate
+        self.calibration_samples = calibration_samples
+
+        self.quantizer.set_global(self.quantization_config)
+
+        self.converted_graph = None
+        self.is_qat = is_qat
+    
+    def stage_type(self) -> str:
+        return StageType.QUANTIZE
+
+    def run(
+        self, artifact: torch.nn.Module, inputs: Optional[Tuple[torch.Tensor]]
+    ) -> None:
+        assert inputs is not None
+        if self.is_qat:
+            artifact.train()
+        captured_graph = export_for_training(artifact, inputs, strict=True).module()
+
+        assert isinstance(captured_graph, torch.fx.GraphModule)
+
+        if self.is_qat:
+            prepared = prepare_qat_pt2e(captured_graph, self.quantizer)
+        else:
+            prepared = prepare_pt2e(captured_graph, self.quantizer)
+
+        if self.calibrate:
+            # Calibrate prepared model to provide data to quantization observers.
+            if self.calibration_samples is not None:
+                for inp in self.calibration_samples:
+                    prepared(*inp)
+            else:
+                prepared(*inputs)
+
+        converted = convert_pt2e(prepared)
+        DuplicateDynamicQuantChainPass()(converted)
+
+        self.converted_graph = converted
+
+    @property
+    def artifact(self) -> torch.fx.GraphModule:
+        return self.converted_graph
+
+    @property
+    def graph_module(self) -> str:
+        return self.converted_graph
+
+    def run_artifact(self, inputs):
+        return self.converted_graph.forward(*inputs)