Enabling fp_quantizer extension unit tests on ROCm

[jithunnair-amd]

Hi @loadams,

We would like to enable the fp_quantizer extension for ROCm. However, the unit tests for it require qtorch. This project seems to not be actively maintained.

We could create a ROCm fork of QPyTorch to tackle this issue, but I'm not sure how we would be able to handle the requirements-dev.txt dependency being different for ROCm.

Any suggestions to resolve this?

cc @rraminen @pruthvistony

[loadams]

Hi @jithunnair-amd - thanks for bringing this up, let me take a look since I had forgotten about the dependency on qtorch.

[jithunnair-amd]

Another way around could be for us to create a deepspeedai/QPyTorch fork if you really want to use it for the fp_quantizer tests, so we can integrate the ROCm fixes into that fork and use it as the common dependency instead of the original QPyTorch repo

[loadams]

Agreed, that would be the best way to resolve the conflict in requirements for the qtorch dependency.

cc: @tjruwase as well for awareness on the package we rely on being actively updated/maintained?

[loadams]

@jithunnair-amd - there do not appear to be any other replacement or equivalents for QTorch that you've seen?