How Built a CM Script

[zixianwang2022]

Hi,

I have a customized inference script to run SDXL Datacenter Category code. How can I incorporate it into my local CM directory as part of a CM Run script?

Thank you

[arjunsuresh]

If you are talking about modifications to reference implementation, you can do the changes, push them to your fork in github and then add the below option to your CM run commands.

--adr.mlperf-inference-implementation.repo=https://github/arjunsuresh/inference,_branch.master.

If you want to do some customizations to the options, you can do it in the below file
https://github.com/mlcommons/cm4mlops/blob/main/script/app-mlperf-inference-mlcommons-python/customize.py#L287

which will be mapped to $HOME/CM/repos/mlcommons@cm4mlops/script/app-mlperf-inference-mlcommons-python/customize.py locally.

[zixianwang2022]

Thank you Arjun,

Which of the methods do you recommend if this is for running the SCC24 SDXL Coco dataset scripts? We have a customized and new backend_sut. We looked into the cm4mlops/script, it seems that each new script requires a customize.py with preprocess and postprocess function. Is there a clear instructions on how to incorporate customized backend_sut.py into customize.py ?

[arjunsuresh]

If you are adding a custom backend for the reference implementation, you can do the first option given here. There should be no need to modify any CM script.