Search for: All records

The "pretrain-then-finetune" paradigm is commonly adopted in the deployment of large language models. Low-Rank Adaptation (LoRA), a parameter-efficient fine-tuning method, is often employed to adapt a base model to a multitude of tasks, resulting in a substantial collection of LoRA adapters derived from one base model. We observe that this paradigm presents significant opportunities for batched inference during serving. To capitalize on these opportunities, we present S-LoRA, a system designed for the scalable serving of many LoRA adapters. S-LoRA stores all adapters in the main memory and fetches the adapters used by the currently running queries to the GPU memory. To efficiently use the GPU memory and reduce fragmentation, S-LoRA proposes Unified Paging. Unified Paging uses a unified memory pool to manage dynamic adapter weights with different ranks and KV cache tensors with varying sequence lengths. Additionally, S-LoRA employs a novel tensor parallelism strategy and highly optimized custom CUDA kernels for heterogeneous batching of LoRA computation. Collectively, these features enable S-LoRA to serve thousands of LoRA adapters on a single GPU or across multiple GPUs with a small overhead. Compared to state-of-the-art libraries such as HuggingFace PEFT and vLLM (with naive support of LoRA serving), S-LoRA can improve the throughput by up to 4 times and increase the number of served adapters by several orders of magnitude. As a result, S-LoRA enables scalable serving of many task-specific fine-tuned models and offers the potential for large-scale customized fine-tuning services. The code is available at this https URL 

Large language models have substantially advanced nuance and context understanding in natural language processing (NLP), further fueling the growth of intelligent conversational interfaces and virtual assistants. However, their hefty computational and memory demands make them potentially expensive to deploy on cloudless edge platforms with strict latency and energy requirements. For example, an inference pass using the state-of-the-art BERT-base model must serially traverse through 12 computationally intensive transformer layers, each layer containing 12 parallel attention heads whose outputs concatenate to drive a large feed-forward network. To reduce computation latency, several algorithmic optimizations have been proposed, e.g., a recent algorithm dynamically matches linguistic complexity with model sizes via entropy-based early exit. Deploying such transformer models on edge platforms requires careful co-design and optimizations from algorithms to circuits, where energy consumption is a key design consideration.