Results of IndySCC24 Dreambrook Team

[NeverGpDzy]

No description provided.

[github-actions]

MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅

[NeverGpDzy]

Hi！I am a member of Dreambrook Team at IndySCC24.

I have completed the basic tests on both CPU and GPU, and I have signed the CLA and submitted the results via a pull request.

Furthermore, I am curious about how to conduct performance tuning.
I've observed that when benchmarking with BERT99, my GPU utilization hovers around 54%, which indicates that the full computational capacity of the GPU is not being utilized. I am considering enhancing the computation speed by adjusting the batch size and switching from fp32 to fp16, but I am unsure about where to make these modifications in the scripts and how to alter the commands accordingly.

[arjunsuresh]

Hi @NeverGpDzy , it looks good and let me merge them now. You are using the PyTorch backend and the source code being run can be seen at Bert PyTorch implementation

You can fork the MLCommons inference repository and then to run the code from the fork just add the below option to your run command.

--adr.mlperf-implementation.tags=_repo.https://github.com/NeverGpDzy/inference,_branch.master

You can also find the local file location in the CM cache using the below command

cm show cache --tags=inference,src -j

But AFAIK optimization work is not expected in IndySCC24 and your current submission itself should be enough to get max score - please confirm this though with the IndySCC team.