LoRA on a Budget: Function-Aware Post-Hoc Compression for Adapter Fleets
Summary
- The paper introduces Functional Rank Allocation (FRA), an unlabeled method for post-hoc compression of LoRA adapter fleets under shared resource budgets.
- Serving multiple LoRA adapters requires distributing ranks across layers within each adapter and across adapters within the fleet.
- FRA measures isolated module truncations end to end, uses dynamic programming for an additive search surrogate, and combines a spectral proposal with exact minimax allocation.
- Across 78 adapters from three public fleets, FRA improves worst-case and tail utility compared to uniform spectral compression under constrained budgets.
Why it matters
Efficiently compressing and managing fleets of LoRA adapters is crucial for scalable, multi-tenant foundation model deployments operating under strict resource budgets.