KV cache compression promises increased throughput and efficiency with negligible loss in performance. While the gains in throughput are indisputable and recent literature has indeed shown minimal degradation on particular benchmarks, in general the consequences of compression in realistic scenarios such as multi-instruction prompting have been insufficiently studied. In this paper, we identify several pitfalls practitioners should be aware of when deploying KV cache compressed LLMs. Importantly, we show that certain instructions degrade much more rapidly with compression, effectively causing them to be completely ignored by the LLM. As a practical example of that, we highlight system prompt leakage as a case study, empirically showing the impact of compression on leakage and general instruction following. We show several factors that play a role in prompt leakage: compression method, instruction order, and KV eviction bias. We then propose simple changes to KV cache eviction policies that can reduce the impact of these factors and improve the overall performance in multi-instruction tasks. KV cache compression offers a compelling trade-off: sacrifice a small amount of model performance for substantial gains in inference efficiency. The technique addresses the main bottleneck in serving large language models (LLMs): the memory required to store the Key-V alue (KV) cache (Pope et al., 2023). During autoregressive generation, this cache grows linearly with context length, making inference a memory-bounded operation that limits server throughput and increases latency (Y uan et al., 2024b). Recently, many compression methods have emerged, each with various KV eviction techniques (Shi et al., 2024a).

Serving large language models (LLMs) is important for cloud providers, and caching intermediate results (KV\$) after processing each request substantially improves serving throughput and latency. However, there is limited understanding of how LLM serving benefits from KV\$ caching, where system design decisions like cache eviction policies are highly workload-dependent. In this paper, we present the first systematic characterization of the KV\$ workload patterns from one of the leading LLM service providers. We draw observations that were not covered by previous studies focusing on synthetic workloads, including: KV\$ reuses are skewed across requests, where reuses between single-turn requests are equally important as multi-turn requests; the reuse time and probability are diverse considering all requests, but for a specific request category, the pattern tends to be predictable; and the overall cache size required for an ideal cache hit ratio is moderate. Based on the characterization, we further propose a workload-aware cache eviction policy that improves the serving performance under real-world traces, especially with limited cache capacity.

Large language models (LLMs) have shown remarkable performance in complex reasoning tasks, but their efficiency is hindered by the substantial memory and computational costs associated with generating lengthy tokens. In this paper, we propose LightThinker, a novel method that enables LLMs to dynamically compress intermediate thoughts during reasoning. Inspired by human cognitive processes, LightThinker compresses verbose thought steps into compact representations and discards the original reasoning chains, thereby significantly reducing the number of tokens stored in the context window. This is achieved by training the model on when and how to perform compression through data construction, mapping hidden states to condensed gist tokens, and creating specialized attention masks. Additionally, we introduce the Dependency (Dep) metric to quantify the degree of compression by measuring the reliance on historical tokens during generation. Extensive experiments on four datasets and two models show that LightThinker reduces peak memory usage and inference time, while maintaining competitive accuracy. Our work provides a new direction for improving the efficiency of LLMs in complex reasoning tasks without sacrificing performance. Code will be released at https://github.com/zjunlp/LightThinker.

Large Language Models (LLMs) have ignited an innovative surge of AI applications, marking a new era of exciting possibilities equipped with extended context windows. However, hosting these models is cost-prohibitive mainly due to the extensive memory consumption of KV Cache involving long-context modeling. Despite several works proposing to evict unnecessary tokens from the KV Cache, most of them rely on the biased local statistics of accumulated attention scores and report performance using unconvincing metric like perplexity on inadequate short-text evaluation. In this paper, we propose NACL, a general framework for long-context KV cache eviction that achieves more optimal and efficient eviction in a single operation during the encoding phase. Due to NACL's efficiency, we combine more accurate attention score statistics in PROXY TOKENS EVICTION with the diversified random eviction strategy of RANDOM EVICTION, aiming to alleviate the issue of attention bias and enhance the robustness in maintaining pivotal tokens for long-context modeling tasks. Notably, our method significantly improves the performance on short- and long-text tasks by 80% and 76% respectively, reducing KV Cache by up to 50% with over 95% performance maintenance. The code is available at https://github.com/PaddlePaddle/Research/tree/master/NLP/ACL2024-NACL.

Graph Neural Networks (GNNs) are widely used today in recommendation systems, fraud detection, and node/link classification tasks. Real world GNNs continue to scale in size and require a large memory footprint for storing graphs and embeddings that often exceed the memory capacities of the target GPUs used for training. To address limited memory capacities, traditional GNN training approaches use graph partitioning and sharding techniques to scale up across multiple GPUs within a node and/or scale out across multiple nodes. However, this approach suffers from the high computational costs of graph partitioning algorithms and inefficient communication across GPUs. To address these overheads, we propose Large-scale Storage-based Multi-GPU GNN framework (LSM-GNN), a storagebased approach to train GNN models that utilizes a novel communication layer enabling GPU software caches to function as a system-wide shared cache with low overheads.LSM-GNN incorporates a hybrid eviction policy that intelligently manages cache space by using both static and dynamic node information to significantly enhance cache performance. Furthermore, we introduce the Preemptive Victim-buffer Prefetcher (PVP), a mechanism for prefetching node feature data from a Victim Buffer located in CPU pinned-memory to further reduce the pressure on the storage devices. Experimental results show that despite the lower compute capabilities and memory capacities, LSM-GNN in a single node with two GPUs offers superior performance over two-node-four-GPU Dist-DGL baseline and provides up to 3.75x speed up on end-to-end epoch time while running large-scale GNN training

Despite the recent success associated with Large Language Models~(LLMs), they are notably cost-prohibitive to deploy in resource-constrained environments due to their excessive memory and computational demands. In addition to model parameters, the key-value cache is also stored in GPU memory, growing linearly with batch size and sequence length. As a remedy, recent works have proposed various eviction policies for maintaining the overhead of key-value cache under a given budget. This paper embarks on the efficacy of existing eviction policies in terms of \textit{importance score calculation} and \textit{eviction scope construction}. We identify the deficiency of prior policies in these two aspects and introduce RoCo, a \underline{r}\underline{o}bust \underline{c}ache \underline{o}mission policy based on temporal attention scores and robustness measures. Extensive experimentation spanning prefilling and auto-regressive decoding stages validates the superiority of RoCo. Finally, we release EasyKV, a versatile software package dedicated to user-friendly key-value constrained generative inference. Code available at \url{https://github.com/DRSY/EasyKV}.

Large Language Models (LLMs), despite their recent impressive accomplishments, are notably cost-prohibitive to deploy, particularly for applications involving long-content generation, such as dialogue systems and story writing. Often, a large amount of transient state information, referred to as the KV cache, is stored in GPU memory in addition to model parameters, scaling linearly with the sequence length and batch size. In this paper, we introduce a novel approach for implementing the KV cache which significantly reduces its memory footprint. Our approach is based on the noteworthy observation that a small portion of tokens contributes most of the value when computing attention scores. We call these tokens Heavy Hitters (H$_2$). Through a comprehensive investigation, we find that (i) the emergence of H$_2$ is natural and strongly correlates with the frequent co-occurrence of tokens in the text, and (ii) removing them results in significant performance degradation. Based on these insights, we propose Heavy Hitter Oracle (H$_2$O), a KV cache eviction policy that dynamically retains a balance of recent and H$_2$ tokens. We formulate the KV cache eviction as a dynamic submodular problem and prove (under mild assumptions) a theoretical guarantee for our novel eviction algorithm which could help guide future work. We validate the accuracy of our algorithm with OPT, LLaMA, and GPT-NeoX across a wide range of tasks. Our implementation of H$_2$O with 20% heavy hitters improves the throughput over three leading inference systems DeepSpeed Zero-Inference, Hugging Face Accelerate, and FlexGen by up to 29$\times$, 29$\times$, and 3$\times$ on OPT-6.7B and OPT-30B. With the same batch size, H2O can reduce the latency by up to 1.9$\times$. The code is available at https://github.com/FMInference/H2O.

As the accuracy of machine learning models increases at a fast rate, so does their demand for energy and compute resources. On a low level, the major part of these resources is consumed by data movement between different memory units. Modern hardware architectures contain a form of fast memory (e.g., cache, registers), which is small, and a slow memory (e.g., DRAM), which is larger but expensive to access. We can only process data that is stored in fast memory, which incurs data movement (input/output-operations, or I/Os) between the two units. In this paper, we provide a rigorous theoretical analysis of the I/Os needed in sparse feedforward neural network (FFNN) inference. We establish bounds that determine the optimal number of I/Os up to a factor of 2 and present a method that uses a number of I/Os within that range. Much of the I/O-complexity is determined by a few high-level properties of the FFNN (number of inputs, outputs, neurons, and connections), but if we want to get closer to the exact lower bound, the instance-specific sparsity patterns need to be considered. Departing from the 2-optimal computation strategy, we show how to reduce the number of I/Os further with simulated annealing. Complementing this result, we provide an algorithm that constructively generates networks with maximum I/O-efficiency for inference. We test the algorithms and empirically verify our theoretical and algorithmic contributions. In our experiments on real hardware we observe speedups of up to 45$\times$ relative to the standard way of performing inference.