Low-Rank Adaptation (LoRA) has recently gained attention for fine-tuning foundation models by incorporating trainable low-rank matrices, thereby reducing the number of trainable parameters. While LoRA offers numerous advantages, its applicability for real-time serving to a diverse and global user base is constrained by its incapability to handle multiple task-specific adapters efficiently. This imposes a performance bottleneck in scenarios requiring personalized, task-specific adaptations for each incoming request. To mitigate this constraint, we introduce Fast LoRA (FLoRA), a framework in which each input example in a minibatch can be associated with its unique low-rank adaptation weights, allowing for efficient batching of heterogeneous requests. We empirically demonstrate that FLoRA retains the performance merits of LoRA, showcasing competitive results on the MultiPL-E code generation benchmark spanning over 8 languages and a multilingual speech recognition task across 6 languages.
more »
« less
This content will become publicly available on May 12, 2025
SLoRA: Scalable Serving of Thousands of LoRA Adapters
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
more »
« less
- Award ID(s):
- 1846431
- PAR ID:
- 10552444
- Publisher / Repository:
- Proceedings of Machine Learning and Systems 6 (MLSys 2024)
- Date Published:
- Format(s):
- Medium: X
- Location:
- Santa Clara, CA
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Training Large Language Models (LLMs) presents significant memory challenges, predominantly due to the growing size of weights and optimizer states. Common memory-reduction approaches, such as low-rank adaptation (LoRA), add a trainable low-rank matrix to the frozen pre-trained weight in each layer, reducing trainable parameters and optimizer states. However, such approaches typically underperform training with full-rank weights in both pre-training and fine-tuning stages since they limit the parameter search to a low-rank subspace and alter the training dynamics, and further, may require full-rank warm start. In this work, we propose Gradient Low-Rank Projection (GaLore), a training strategy that allows full-parameter learning but is more memory-efficient than common low-rank adaptation methods such as LoRA. Our approach reduces memory usage by up to 65.5% in optimizer states while maintaining both efficiency and performance for pre-training on LLaMA 1B and 7B architectures with C4 dataset with up to 19.7B tokens, and on fine-tuning RoBERTa on GLUE tasks. Our 8-bit GaLore further reduces optimizer memory by up to 82.5% and total training memory by 63.3%, compared to a BF16 baseline. Notably, we demonstrate, for the first time, the feasibility of pre-training a 7B model on consumer GPUs with 24GB memory (e.g., NVIDIA RTX 4090) without model parallel, checkpointing, or offloading strategies.more » « less
-
Training Large Language Models (LLMs) presents significant memory challenges, predominantly due to the growing size of weights and optimizer states. Common memory-reduction approaches, such as low-rank adaptation (LoRA), add a trainable low-rank matrix to the frozen pre-trained weight in each layer, reducing trainable parameters and optimizer states. However, such approaches typically underperform training with full-rank weights in both pre-training and fine-tuning stages since they limit the parameter search to a low-rank subspace and alter the training dynamics, and further, may require full-rank warm start. In this work, we propose Gradient Low-Rank Projection (GaLore), a training strategy that allows full-parameter learning but is more memory-efficient than common low-rank adaptation methods such as LoRA. Our approach reduces memory usage by up to 65.5% in optimizer states while maintaining both efficiency and performance for pre-training on LLaMA 1B and 7B architectures with C4 dataset with up to 19.7B tokens, and on fine-tuning RoBERTa on GLUE tasks. Our 8-bit GaLore further reduces optimizer memory by up to 82.5% and total training memory by 63.3%, compared to a BF16 baseline. Notably, we demonstrate, for the first time, the feasibility of pre-training a 7B model on consumer GPUs with 24GB memory (e.g., NVIDIA RTX 4090) without model parallel, checkpointing, or offloading strategies.more » « less
-
Concurrent kernel execution on GPU has proven an effective technique to improve system throughput by maximizing the resource utilization. In order to increase programmability and meet the increasing memory requirements of data-intensive applications, current GPUs support Unified Virtual Memory (UVM), which provides a virtual memory abstraction with demand paging. By allowing applications to oversubscribe GPU memory, UVM provides increased opportunities to share GPU resources across applications. However, in the presence of applications with competing memory requirements, GPU sharing can lead to performance degradation due to thrashing. NVIDIA's Multiple Process Service (MPS) offers the capability to space share bare metal GPUs, thereby enabling cluster workload managers, such as Slurm, to share a single GPU across MPI ranks with limited control over resource partitioning. However, it is not possible to preempt, schedule, or throttle a running GPU process through MPS. These features would enable new OS-managed scheduling policies to be implemented for GPU kernels to dynamically handle resource contention and offer consistent performance. The contribution of this paper is two-fold. We first show how memory oversubscription can impact the performance of concurrent GPU applications. Then, we propose three methods to transparently mitigate memory interference through kernel preemption and scheduling policies. To implement our policies, we develop our own runtime system (PILOT) to serve as an alternative to NVIDIA's MPS. In the presence of memory over-subscription, we noticed a dramatic improvement in the overall throughput when using our scheduling policies and runtime hints.more » « less
-
Begnum, Kyrre ; Border, Charles (Ed.)With the increasing popularity of large deep learning model serving workloads, there is a pressing need to reduce the energy consumption of a model-serving cluster while maintaining satisfied throughput or model-serving latency requirements. Model multiplexing approaches such as model parallelism, model placement, replication, and batching aim to optimize the model-serving performance. However, they fall short of leveraging the GPU frequency scaling opportunity for power saving. In this paper, we demonstrate (1) the benefits of GPU frequency scaling in power saving for model serving; and (2) the necessity for co-design and optimization of fine grained model multiplexing and GPU frequency scaling. We explore the co-design space and present a novel power-aware model-serving system, μ-Serve. μ-Serve is a model-serving framework that optimizes the power consumption and model serving latency/throughput of serving multiple ML models efficiently in a homogeneous GPU cluster. Evaluation results on production workloads show that μ-Serve achieves 1.2–2.6× power saving by dynamic GPU frequency scaling (up to 61% reduction) without SLO attainment violations.more » « less
