LoRAForge: Co-Designing Post-Hoc Compression and Execution for Efficient Multi-LoRA Serving
Multi-LoRA serving allows many task-specific adapters to share a common base model, but scaling to large adapter fleets introduces substantial storage, loading, and execution overheads. Post-hoc compression can reduce the memory footprint of trained LoRA adapters, yet smaller ranks do not automatically translate into f...