Why it matters

S-LoRA is theoretical basis for multi-LoRA. Understanding shapes design.

Advertisement

The architecture

Unified paged memory for LoRA weights.

Kernels handle multiple adapters per batch.

S-LoRA designUnified pagingLoRA weight pagesCustom kernelsmixed batchSchedulingadapter-awarePunica paper similar ideas; vLLM + LoRAX implement in practice
S-LoRA.
Advertisement

How it works end to end

SGMV kernel: segment matmul.

Paging for LoRA weights.

Scheduler considers adapter demand.