Performance - No constant folding/propagation happens on final .pte

🐛 Describe the bug

After exporting a model with torch.export.export, to_edge_ and even applying several passes to propagate/fold constants, the final .pte still containing several nodes operating only on constants.

I find, specifically, nodes like:

t1 = aten::add(const, const, const)
t2 = aten::mul(t1, const)
t3 = aten::permute_copy(t2)
t4 = aten::unsqueeze_copy(t3)

t5 = aten::add(const, const, const)
t6 = aten::mul(t5, const)
t7 = aten::permute_copy(t6)

t8 = aten::index(tensor, t4, t7)

Why all these operations are not folded even when applying the passes?

This is how I exported and applied the passes:

from executorch.exir.passes.constant_prop_pass import constant_prop_pass
from executorch.exir.passes.const_prop_pass import ConstPropPass
from executorch import exir

self.model.eval()
aten_dialect_program = torch.export.export(self.model, (self.input,))
aten_dialect_program = constant_prop_pass(aten_dialect_program)
edge_dialect_program = exir.to_edge_transform_and_lower(aten_dialect_program, transform_passes=[ConstPropPass()])
executorch_program = edge_dialect_program.to_executorch(
    exir.ExecutorchBackendConfig(
        passes=[]
    )
)

Is there a way of making sure no constants are operations are left on the final .pte before inference? For performance reasons.

Versions

Versions of relevant libraries:
[pip3] executorch==0.6.0a0+4e3a8bd
[pip3] numpy==2.0.0
[pip3] torch==2.7.0.dev20250131+cpu
[pip3] torchao==0.8.0+git11333ba2
[pip3] torchaudio==2.6.0.dev20250131+cpu
[pip3] torchsr==1.0.4
[pip3] torchvision==0.22.0.dev20250131+cpu

cc @JacobSzwejbka @angelayi