Arm backend: Remove DecomposeLinearVectorNormPass from INT pipeline

DecomposeLinearVectorNormPass is listed in both the
transform_for_annotation_pipeline and _tosa_INT_pipeline stages. The
latter is redundant because torch.linalg.vector_norm can only run with
floating point inputs, i.e., we should always be in a quantized setting
when we enter _tosa_INT_pipeline if the operator is present and
_transform_for_annotation_pipeline will always run; therefore, remove
DecomposeLinearVectorNormPass from _tosa_INT_pipeline.

Signed-off-by: Martin Lindström <[email protected]>
Change-Id: I647687b51298bbb98087914fbbee053436ffb79f

Author: Martin Lindström

backends/arm/_passes/arm_pass_manager.py

@@ -155,7 +155,6 @@ def _tosa_INT_pipeline(self, exported_program: ExportedProgram) -> GraphModule:
         self.add_pass(RemoveGetItemPass())
         self.add_pass(ConvertSplitToSlicePass())
         self.add_pass(ConvertMmToBmmPass())
-        self.add_pass(DecomposeLinearVectorNormPass())
         self.add_pass(
             DecomposeMeanDimPass(exported_program.graph_module, self.tosa_spec)
         )