Large Language Model (LLM) serving is increasingly constrained by the growing size of the key-value (KV) cache, which scales with both context length and generation length. Prior work shows that attention is dominated by a small subset of critical tokens, yet existing systems struggle to exploit this efficiently without degrading accuracy, especially in long generation. We make a key observation: the temporal stability of these critical tokens varies significantly across KV heads: some heads consistently focus on the same tokens, while others shift frequently. Building on this insight, we introduce FlexiCache, a hierarchical KV-cache management system that leverages the temporal stability of KV heads to reduce GPU memory usage and computation overhead, while preserving model accuracy. FlexiCache classifies KV heads as stable or unstable: it retains all KV-cache pages from unstable heads in GPU memory, whereas for stable heads, it keeps only the top-K pages on the GPU and offloads the rest to host memory. By exploiting temporal stability, FlexiCache performs periodic reranking for stable heads to fetch newly promoted top pages. Implemented atop vLLM, FlexiCache reduces GPU memory footprint for long-context requests by up to 70%, improves offline serving throughput by 1.38-1.55x, and lowers online token latency by 1.6-2.1x, all while maintaining accuracy in long-context, long-generation scenarios.

With the surging growth of model parameters, foundation models pose unprecedented challenges to traditional computational infrastructures.

Training LLMs larger than the aggregated memory of multiple GPUs is increasingly necessary due to the faster growth of LLM sizes compared to GPU memory. To this end, multi-tier host memory or disk offloading techniques are proposed by state of art. Despite advanced asynchronous multi-tier read/write strategies, such offloading strategies result in significant I/O overheads in the critical path of training, resulting in slower iterations. To this end, we propose MLP-Offload, a novel multi-level, multi-path offloading engine specifically designed for optimizing LLM training on resource-constrained setups by mitigating I/O bottlenecks. We make several key observations that drive the design of MLP-Offload, such as I/O overheads during the update dominate the iteration time; I/O bandwidth of the third-level remote storage tier remains unutilized; and, contention due to concurrent offloading amplifies I/O bottlenecks. Driven by these insights, we design and implement MLP-Offload to offload the optimizer states across multiple tiers in a cache-efficient and concurrency-controlled fashion to mitigate I/O bottlenecks during the backward and update phases. Evaluations on models up to 280B parameters shows that MLP-Offload achieves 2.5$\times$ faster iterations compared to the state-of-the-art LLM training runtimes.

Graph neural networks (GNNs) are widely used for learning node embeddings in graphs, typically adopting a message-passing scheme. This approach, however, leads to the neighbor explosion problem, with exponentially growing computational and memory demands as layers increase. Graph sampling has become the predominant method for scaling GNNs to large graphs, mitigating but not fully solving the issue. Pre-propagation GNNs (PP-GNNs) represent a new class of models that decouple feature propagation from training through pre-processing, addressing neighbor explosion in theory. Yet, their practical advantages and system-level optimizations remain underexplored. This paper provides a comprehensive characterization of PP-GNNs, comparing them with graph-sampling-based methods in training efficiency, scalability, and accuracy. While PP-GNNs achieve comparable accuracy, we identify data loading as the key bottleneck for training efficiency and input expansion as a major scalability challenge. To address these issues, we propose optimized data loading schemes and tailored training methods that improve PP-GNN training throughput by an average of 15$\times$ over the PP-GNN baselines, with speedup of up to 2 orders of magnitude compared to sampling-based GNNs on large graph benchmarks. Our implementation is publicly available at https://github.com/cornell-zhang/preprop-gnn.

As large language models (LLMs) continue to support increasingly longer contexts, the memory demand for key-value (KV) caches during decoding grows rapidly, becoming a critical bottleneck in both GPU memory capacity and PCIe bandwidth. Sparse attention mechanisms alleviate this issue by computing attention weights only for selected key-value pairs. However, their indexing computation typically requires traversing all key vectors, resulting in significant computational and data transfer overhead. To reduce the cost of index retrieval, existing methods often treat each decoding step as an independent process, failing to exploit the temporal correlations embedded in historical decoding information. To this end, we propose LFPS(Learn From the Past for Sparse Indexing), an acceleration method that dynamically constructs sparse indexing candidates based on historical attention patterns. LFPS captures two prevalent trends in decoder attention -vertical patterns (attending to fixed positions) and slash patterns (attending to relative positions) -and incorporates a positional expansion strategy to effectively predict the Top-k indices for the current step. We validate LFPS on challenging long-context benchmarks such as LongBench-RULER, using Llama-3.1-8B-Instruct as the base model. Experimental results show that LFPS achieves up to 22.8$\times$ speedup over full attention and 9.6$\times$ speedup over exact Top-k retrieval on an RTX 4090 GPU and a single CPU core of a Xeon Gold 6430, respectively, while preserving generation accuracy. These results demonstrate that LFPS offers a practical and efficient solution for decoding optimization in long-context LLM inference.

We present the design and implementation of a new lifetime-aware tensor offloading framework for GPU memory expansion using low-cost PCIe-based solid-state drives (SSDs). Our framework, TERAIO, is developed explicitly for large language model (LLM) training with multiple GPUs and multiple SSDs. Its design is driven by our observation that the active tensors take only a small fraction (1.7% on average) of allocated GPU memory in each LLM training iteration, the inactive tensors are usually large and will not be used for a long period of time, creating ample opportunities for offloading/prefetching tensors to/from slow SSDs without stalling the GPU training process. TERAIO accurately estimates the lifetime (active period of time in GPU memory) of each tensor with the profiling of the first few iterations in the training process. With the tensor lifetime analysis, TERAIO will generate an optimized tensor offloading/prefetching plan and integrate it into the compiled LLM program via PyTorch. TERAIO has a runtime tensor migration engine to execute the offloading/prefetching plan via GPUDirect storage, which allows direct tensor migration between GPUs and SSDs for alleviating the CPU bottleneck and maximizing the SSD bandwidth utilization. In comparison with state-of-the-art studies such as ZeRO-Offload and ZeRO-Infinity, we show that TERAIO improves the training performance of various LLMs by 1.47x on average, and achieves 80.7% of the ideal performance assuming unlimited GPU memory.

Transformers and large language models~(LLMs) have seen rapid adoption in all domains. Their sizes have exploded to hundreds of billions of parameters and keep increasing. Under these circumstances, the training of transformers is very expensive and often hits a ``memory wall'', i.e., even when using 3D parallelism (pipeline, tensor, data) and aggregating the memory of many GPUs, it is still not enough to hold the necessary data structures (model parameters, optimizer state, gradients, activations) in GPU memory. To compensate, state-of-the-art approaches offload the optimizer state, at least partially, to the host memory and perform hybrid CPU-GPU computations. However, the management of the combined host-GPU memory is often suboptimal and results in poor overlapping between data movements and computations. This leads to missed opportunities to simultaneously leverage the interconnect bandwidth and computational capabilities of CPUs and GPUs. In this paper, we leverage a key observation that the interleaving of the forward, backward and update phases generate fluctuations in the GPU memory utilization, which can be exploited to dynamically move a part of the optimizer state between the host and the GPU memory at each iteration. To this end, we design and implement \proj, a novel technique to split the LLM into subgroups, whose update phase is scheduled on either the CPU or the GPU based on our proposed performance model that addresses the trade-off between data movement cost, acceleration on the GPUs vs the CPUs, and competition for shared resources. We integrate our approach with DeepSpeed and demonstrate 2.5$\times$ faster iterations over state-of-the-art approaches using extensive experiments.

Large Language Models (LLMs) with long context capabilities are integral to complex tasks in natural language processing and computational biology, such as text generation and protein sequence analysis. However, training LLMs directly on extremely long contexts demands considerable GPU resources and increased memory, leading to higher costs and greater complexity. Alternative approaches that introduce long context capabilities via downstream finetuning or adaptations impose significant design limitations. In this paper, we propose Fully Pipelined Distributed Transformer (FPDT) for efficiently training long-context LLMs with extreme hardware efficiency. For GPT and Llama models, we achieve a 16x increase in sequence length that can be trained on the same hardware compared to current state-of-the-art solutions. With our dedicated sequence chunk pipeline design, we can now train 8B LLM with 2 million sequence length on only 4 GPUs, while also maintaining over 55% of MFU. Our proposed FPDT is agnostic to existing training techniques and is proven to work efficiently across different LLM models.

Interacting with humans through multi-turn conversations is a fundamental feature of large language models (LLMs). However, existing LLM serving engines executing multi-turn conversations are inefficient due to the need to repeatedly compute the key-value (KV) caches of historical tokens, incurring high serving costs. To address the problem, this paper proposes CachedAttention, a new attention mechanism that enables reuse of KV caches across multi-turn conversations, significantly reducing the repetitive computation overheads. CachedAttention maintains a hierarchical KV caching system that leverages cost-effective memory/storage mediums to save KV caches for all requests. To reduce KV cache access overheads from slow mediums, CachedAttention employs layer-wise pre-loading and asynchronous saving schemes to overlap the KV cache access with the GPU computation. To ensure that the KV caches to be accessed are placed in the fastest hierarchy, CachedAttention employs scheduler-aware fetching and eviction schemes to consciously place the KV caches in different layers based on the hints from the inference job scheduler. To avoid the invalidation of the saved KV caches incurred by context window overflow, CachedAttention enables the saved KV caches to remain valid via decoupling the positional encoding and effectively truncating the KV caches. Extensive experimental results demonstrate that CachedAttention significantly decreases the time to the first token (TTFT) by up to 87%, improves the prompt prefilling throughput by up to 7.8$\times$ for multi-turn conversations, and reduces the end-to-end inference cost by up to 70%.