20314 (opens in new tab)

#Python Quantize moveaxis/movedim so they delegate to Ethos-U Summary: The ARM PT2 quantizer's pass-through shared-qspec set in quantization_annotator.py (_one_to_one_shared_input_qspec) covers permute/permute_copy/transpose/view/squeeze etc., but omits aten.moveaxis/aten.movedim. A model that uses torch.moveaxis therefore leaves those ops unquantized: the quantizer brackets each one with dequantize -> moveaxis(float) -> quantize. On lowering, moveaxis decomposes to a float permute_copy. The ...

Read the original article