Is there any chance to share the workload trace for the fine-tune model?

[chenhongyu2048]

Hello, I'm currently studying on multi-lora inference system and looking for any data about the trace of fine-tune model useage.
The paper said "Its model repository contains over 200 foundation models and thousands of fine-tuned models(Bailian).", so I'm wondering is there any chance that Alibaba can share the workload trace for the fine-tune model?

We appreciate any offers of help and are looking for opportunities for collaboration in industry/academia.

[lixueclaire]

Hello, I'm currently studying on multi-lora inference system and looking for any data about the trace of fine-tune model useage. The paper said "Its model repository contains over 200 foundation models and thousands of fine-tuned models(Bailian).", so I'm wondering is there any chance that Alibaba can share the workload trace for the fine-tune model?

We appreciate any offers of help and are looking for opportunities for collaboration in industry/academia.

Thank you for your interest in our fine-tuned models. The Multi-Lora inference system is a vital area for us, and we appreciate your work in this field. We are open to evaluating the possibility of sharing workload traces; however, we must first ensure compliance with privacy and security regulations.

If you could share more details about your research objectives and how the workload trace data would enhance your work, we would be better positioned to explore potential collaboration or provide relevant insights.

[chenhongyu2048]

Thanks for the reply!
My current work is to try to provide the same latency and SLO for long-tail LoRA requests as for requests for hot LoRA. Therefore, I hope to obtain traces of invocation frequency of fine-tuned models on the platform, which can help us define the skewness of LoRA request distribution.
Our current work is based on the MLaas trace in the paper by Alibaba in NSDI22. Currently, multi-LoRA inference systems often use MAF from Microsoft as the real load in their experiments. However, both of the above methods are probably different from the actual fine-tuned model invacation trace. This is why we want to obtain some real traces.