up

Author: metascroy

exir/passes/constant_prop_pass.py

@@ -30,10 +30,6 @@
 # Propagating aten.full can significantly increase compiled model size.
 _DEFAULT_SKIP_TARGETS = {exir_ops.edge.aten.full.default}
 
-# Skipping transpose/permute for now because https://github.com/pytorch/executorch/issues/10499
-_DEFAULT_SKIP_TARGETS.add(exir_ops.edge.aten.transpose.int)
-_DEFAULT_SKIP_TARGETS.add(exir_ops.edge.aten.permute.default)
-
 # Do not const prop quantization primitives
 _QDQ_OPS = [
     exir_ops.edge.quantized_decomposed.dequantize_per_channel.default,

extension/llm/export/builder.py

@@ -503,23 +503,23 @@ def to_executorch(
         """
         Lower the model to executorch and get an ExecutorchProgram.
         """
-        # QuantFusionPass is not necessary because we set do_quant_fusion_and_const_prop=True
-        # in ExecutorchBackendConfig
-        to_executorch_passes = [
-            # If there are Linear operations left in the graph, let's execute
-            # them with the optimized op_linear rather than materializing a
-            # transpose followed by a regular op_mm.
-            # TODO: ConvertToLinearPass is not a sound pass and we should fix it
-            # https://github.com/pytorch/executorch/issues/10499
-            ConvertToLinearPass(),
-        ]
+        to_executorch_passes = []
         if passes:
             # pyre-fixme[6]: In call `list.extend`, for 1st positional argument,
             # expected `Iterable[Union[ConvertToLinearPass, QuantFusionPass]]` but
             # got `List[ExportPass]
             to_executorch_passes.extend(passes)
 
         assert self.edge_manager, "Need to run export_to_edge() first"
+
+        # If there are Linear operations left in the graph, let's execute
+        # them with the optimized op_linear rather than materializing a
+        # transpose followed by a regular op_mm.
+        # TODO: ConvertToLinearPass is not a sound pass and must be called before
+        # const propagation.  It requires fixing:
+        # https://github.com/pytorch/executorch/issues/10499
+        self.edge_manager.transform([ConvertToLinearPass()])
+
         self.export_program = self.edge_manager.to_executorch(
             ExecutorchBackendConfig(
                 extract_delegate_segments=True,