Does Executorch support int16x8 quantization?

ExecuTorch
1

Hi,

I’m currently working on getting audio models to run on HiFi 4 DSPs, coming from TFLite Micro.

TFLM supports int16x8 quantization and I’m wondering if Executorch does too.

I just started looking into Executorch, so I’m pretty ignorant about PyTorch stuff at the moment. I read something about Mixed Precision support in PyTorch, but I couldn’t find anything about that in Executorch.

2

Hey @andresovela, thanks for taking a look at ExecuTorch!

@jerryzh168, is there support for int16x8 quantization?

ExecuTorch supports various quantization schemes - see Quantization Overview — ExecuTorch 0.3 documentation

3

it’s not right now probably, but it’s possible to express in quantizer API, and also will require xnnpack or some backend to implement an efficient op for it I think

1 Like
4

I see. Thanks for the response!