Align aot_arm_compiler to latest export flow

- Update to_edge_transform_and_lower rather than export_to_edge - fix channel last on some models Signed-off-by: Rob Elliott <[email protected]> Change-Id: I0f8e9206aa1ff3004a746955010c7bf01c896347
pytorch · Oct 17, 2024 · 3027be6 · 3027be6
1 parent 6b858f2
commit 3027be6
Showing 1 changed file with 35 additions and 29 deletions.
diff --git a/examples/arm/aot_arm_compiler.py b/examples/arm/aot_arm_compiler.py
@@ -25,7 +25,11 @@
 from executorch.backends.arm.util.arm_model_evaluator import GenericModelEvaluator
 
 from executorch.devtools.backend_debug import get_delegation_info
-from executorch.exir import EdgeCompileConfig, ExecutorchBackendConfig
+from executorch.exir import (
+    EdgeCompileConfig,
+    ExecutorchBackendConfig,
+    to_edge_transform_and_lower,
+)
 from executorch.extension.export_util.utils import export_to_edge, save_pte_program
 from tabulate import tabulate
 
@@ -185,7 +189,7 @@ def get_compile_spec(target: str, intermediates: bool) -> ArmCompileSpecBuilder:
                 memory_mode="Shared_Sram",
                 extra_flags="--debug-force-regor --output-format=raw",
             )
-            .set_permute_memory_format(args.model_name in MODEL_NAME_TO_MODEL.keys())
+            .set_permute_memory_format(True)
             .set_quantize_io(True)
         )
     elif "ethos-u85" in target:
@@ -202,7 +206,7 @@ def get_compile_spec(target: str, intermediates: bool) -> ArmCompileSpecBuilder:
         )
 
     if intermediates is not None:
-        spec_builder.dump_intermediate_artifacts_to(args.intermediates)
+        spec_builder.dump_intermediate_artifacts_to(intermediates)
 
     return spec_builder.build()
 
@@ -356,40 +360,42 @@ def get_args():
     model, example_inputs = get_model_and_inputs_from_name(args.model_name)
     model = model.eval()
 
+    # export_for_training under the assumption we quantize, the exported form also works
+    # in to_edge if we don't quantize
+    exported_program = torch.export.export_for_training(model, example_inputs)
+    model = exported_program.module()
     model_fp32 = model
 
-    # pre-autograd export. eventually this will become torch.export
-    model = torch.export.export_for_training(model, example_inputs).module()
-
     # Quantize if required
     model_int8 = None
     if args.quantize:
         model = quantize(model, example_inputs)
         model_int8 = model
+        # Wrap quantized model back into an exported_program
+        exported_program = torch.export.export_for_training(model, example_inputs)
+
+    if args.delegate:
+        # As we can target multiple output encodings from ArmBackend, one must
+        # be specified.
+        compile_spec = get_compile_spec(args.target, args.intermediates)
+        edge = to_edge_transform_and_lower(
+            exported_program,
+            partitioner=[ArmPartitioner(compile_spec)],
+            compile_config=EdgeCompileConfig(
+                _check_ir_validity=False,
+                _skip_dim_order=True,
+            ),
+        )
+    else:
+        edge = to_edge_transform_and_lower(
+            exported_program,
+            compile_config=EdgeCompileConfig(
+                _check_ir_validity=False,
+                _skip_dim_order=True,
+            ),
+        )
 
-    edge = export_to_edge(
-        model,
-        example_inputs,
-        edge_compile_config=EdgeCompileConfig(
-            _check_ir_validity=False,
-        ),
-    )
-
-    # As we can target multiple output encodings from ArmBackend, one must
-    # be specified.
-    compile_spec = (
-        get_compile_spec(args.target, args.intermediates)
-        if args.delegate is True
-        else None
-    )
-
-    logging.debug(f"Exported graph:\n{edge.exported_program().graph}")
-    if args.delegate is True:
-        edge = edge.to_backend(ArmPartitioner(compile_spec))
-
-        dump_delegation_info(edge, args.intermediates)
-
-        logging.debug(f"Lowered graph:\n{edge.exported_program().graph}")
+    dump_delegation_info(edge, args.intermediates)
 
     try:
         exec_prog = edge.to_executorch(