LoRAForge
An efficient system for multi-LoRA serving that co-designs batched factor construction with rank-specialized execution through explicit rank representation. It supports two compression strategies, a unified retention ratio and a spectral energy budget, improving adapter storage, loading, and inference throughput.
Timeline (last 90 days)
LoRAForge is a system that connects batched factor construction with rank-specialized execution through explicit rank-preserving representations, supporting two strategies: a uniform retention ratio (LoRAForge-U) and spectral-energy budget allocation (LoRAForge-S).
In a fixed-workload Qwen controlled experiment, LoRAForge's shared runtime brought 1.17–1.22x throughput improvements to both compression strategies.
The same rank-lower-bound 1/12 spectral export configuration achieved substantial token throughput improvements on both AMD and NVIDIA GPUs.
On original adapters, the throughput gain of the LoRAForge shared runtime is negligible, indicating that the speedup comes from exploiting the compressed rank structure rather than from general optimizations of the runtime itself.
All Facts (4)
LoRAForge is a system that connects batched factor construction with rank-specialized execution through explicit rank-preserving representations, supporting two strategies: a uniform retention ratio (LoRAForge-U) and spectral-energy budget allocation (LoRAForge-S).
50%UnverifiedIn a fixed-workload Qwen controlled experiment, LoRAForge's shared runtime brought 1.17–1.22x throughput improvements to both compression strategies.
50%UnverifiedThe same rank-lower-bound 1/12 spectral export configuration achieved substantial token throughput improvements on both AMD and NVIDIA GPUs.
50%UnverifiedOn original adapters, the throughput gain of the LoRAForge shared runtime is negligible, indicating that the speedup comes from exploiting the compressed rank structure rather than from general optimizations of the runtime itself.
50%