High throughput serving of large language models (LLMs) requires batching sufficiently many requests at a time. However, existing systems struggle because the key-value cache (KV cache) memory for each request is huge and grows and shrinks dynamically. When managed inefficiently, this memory can be significantly wasted by fragmentation and redundant duplication, limiting the batch size. To address this problem, we propose PagedAttention, an attention algorithm inspired by the classical virtual memory and paging techniques in operating systems. On top of it, we build vLLM, an LLM serving system that achieves (1) near-zero waste in KV cache memory and (2) flexible sharing of KV cache within and across requests to further reduce memory usage. Our evaluations show that vLLM improves the throughput of popular LLMs by 2--4× with the same level of latency compared to the state-of-the-art systems, such as FasterTransformer and Orca. The improvement is more pronounced with longer sequences, larger models, and more complex decoding algorithms. vLLM's source code is publicly available at https://github.com/vllm-project/vllm.

Efficient deployment of transformer models in practice is challenging due to their inference cost including memory footprint, latency, and power consumption, which scales quadratically with input sequence length. To address this, we present a novel token reduction method dubbed Learned Token Pruning (LTP) which adaptively removes unimportant tokens as an input sequence passes through transformer layers. In particular, LTP prunes tokens with an attention score below a threshold, whose value is learned for each layer during training. Our threshold-based method allows the length of the pruned sequence to vary adaptively based on the input sequence, and avoids algorithmically expensive operations such as top-k token selection. We extensively test the performance of LTP on GLUE and SQuAD tasks and show that our method outperforms the prior state-of-the-art token pruning methods by up to ∽2.5% higher accuracy with the same amount of FLOPs. In particular, LTP achieves up to 2.1× FLOPs reduction with less than 1% accuracy drop, which results in up to 1.9× and 2.0× throughput improvement on Intel Haswell CPUs and NVIDIA V100 GPUs. Furthermore, we demonstrate that LTP is more robust than prior methods to variations in input sequence lengths. Our code has been developed in PyTorch and open-sourced

Pruning is an effective way to reduce the huge inference cost of Transformer models. However, prior work on pruning Transformers requires retraining the models. This can add high training cost and high complexity to model deployment, making it difficult to use in many practical situations. To address this, we propose a fast post-training pruning framework for Transformers that does not require any retraining. Given a resource constraint and a sample dataset, our framework automatically prunes the Transformer model using structured sparsity methods. To retain high accuracy without retraining, we introduce three novel techniques: (i) a lightweight mask search algorithm that finds which heads and filters to prune based on the Fisher information; (ii) mask rearrangement that complements the search algorithm; and (iii) mask tuning that reconstructs the output activations for each layer. We apply our method to BERT-base and DistilBERT, and we evaluate its effectiveness on GLUE and SQuAD benchmarks. Our framework achieves up to 2.0x reduction in FLOPs and 1.56x speedup in inference latency, while maintaining < 1% loss in accuracy. Importantly, our framework prunes Transformers in less than 3 minutes on a single GPU, which is over two orders of magnitude faster than existing pruning approaches that retrain the models.

Row Hammer is a serious security threat to modern computing systems using DRAM as main memory. It causes charge loss in DRAM cells adjacent to a frequently activated aggressor row and eventually leads to data bit flips in those cells. Even with countermeasures from hardware vendors for years, many latest DDR4 DRAM-based systems are still vulnerable to Row Hammer. Furthermore, technology scaling continues to reduce the Row Hammer threshold, hence posing even greater challenges than before. Although many architectural solutions for Row Hammer have been proposed in both industry and academia, they still incur substantial overhead in terms of chip area, energy, and performance, fail to provide a sufficient level of protection, or both. Thus, we propose Graphene, a low-cost Row Hammer prevention technique based on a space-efficient algorithm that identifies frequent elements from an incoming data stream. Graphene is provably secure without false negatives and with tightly bounded false positives. Furthermore, Graphene has an order of magnitude smaller area overhead compared to a state-of-the-art counter-based scheme. This makes Graphene a scalable solution to Row Hammer attacks for the memory systems of today and the future. Our evaluation shows that Graphene features nearly zero performance and energy overhead when running realistic workloads. Even for the most adversarial memory access patterns, Graphene increases refresh energy only by 0.34%.

Deep learning (DL) frameworks take advantage of GPUs to improve the speed of DL inference and training. Ideally, DL frameworks should be able to fully utilize the computation power of GPUs such that the running time depends on the amount of computation assigned to GPUs. Yet, we observe that in scheduling GPU tasks, existing DL frameworks suffer from inefficiencies such as large scheduling overhead and unnecessary serial execution. To this end, we propose Nimble, a DL execution engine that runs GPU tasks in parallel with minimal scheduling overhead. Nimble introduces a novel technique called ahead-of-time (AoT) scheduling. Here, the scheduling procedure finishes before executing the GPU kernel, thereby removing most of the scheduling overhead during run time. Furthermore, Nimble automatically parallelizes the execution of GPU tasks by exploiting multiple GPU streams in a single GPU. Evaluation on a variety of neural networks shows that compared to PyTorch, Nimble speeds up inference and training by up to 22.34× and 3.61×, respectively. Moreover, Nimble outperforms state-of-the-art inference systems, TensorRT and TVM, by up to 2.81× and 1.70×, respectively.