Training large language models (LLMs) in the cloud faces growing memory bottlenecks due to the limited capacity and high cost of GPUs. While GPU memory offloading to CPU and NVMe has made large-scale training more feasible, existing approaches suffer from high tensor migration latency and suboptimal device memory utilization, ultimately increasing training time and cloud costs. To address these challenges, we present 10Cache, a resource-aware tensor caching and migration system that accelerates LLM training by intelligently coordinating memory usage across GPU, CPU, and NVMe tiers. 10Cache profiles tensor execution order to construct prefetch policies, allocates memory buffers in pinned memory based on tensor size distributions, and reuses memory buffers to minimize allocation overhead. Designed for cloud-scale deployments, 10Cache improves memory efficiency and reduces reliance on high-end GPUs. Across diverse LLM workloads, it achieves up to 2x speedup in training time, improves GPU cache hit rate by up to 86.6x, and increases CPU/GPU memory utilization by up to 2.15x and 1.33x, respectively, compared to state-of-the-art offloading methods. These results demonstrate that 10Cache is a practical and scalable solution for optimizing LLM training throughput and resource efficiency in cloud environments.

Large language models (LLMs) have demonstrated impressive capabilities, but their enormous size poses significant challenges for deployment in real-world applications. To address this issue, researchers have sought to apply network pruning techniques to LLMs. A critical challenge in pruning is allocation the sparsity for each layer. Recent sparsity allocation methods is often based on heuristics or search that can easily lead to suboptimal performance. In this paper, we conducted an extensive investigation into various LLMs and revealed three significant discoveries: (1) the layerwise pruning sensitivity (LPS) of LLMs is highly non-uniform, (2) the choice of pruning metric affects LPS, and (3) the performance of a sparse model is related to the uniformity of its layerwise redundancy level. Based on these observations, we propose that the layerwise sparsity of LLMs should adhere to three principles: \emph{non-uniformity}, \emph{pruning metric dependency}, and \emph{uniform layerwise redundancy level} in the pruned model. To this end, we proposed Maximum Redundancy Pruning (MRP), an iterative pruning algorithm that prunes in the most redundant layers (\emph{i.e.}, those with the highest non-outlier ratio) at each iteration. The achieved layerwise sparsity aligns with the outlined principles. We conducted extensive experiments on publicly available LLMs, including the LLaMA2 and OPT, across various benchmarks. Experimental results validate the effectiveness of MRP, demonstrating its superiority over previous methods.

The increasing size and complexity of Large Language Models (LLMs) pose challenges for their deployment on personal computers and mobile devices. Aggressive post-training model compression is necessary to reduce the models' size, but it often results in significant accuracy loss. To address this challenge, we propose a novel network pruning technology that utilizes over 0.7 sparsity and less than 8 bits of quantization. Our approach enables the compression of prevailing LLMs within a couple of hours while maintaining a relatively small accuracy loss. In experimental evaluations, our method demonstrates effectiveness and potential for practical deployment. By making LLMs available on domestic devices, our work can facilitate a new era of natural language processing applications with wide-ranging impacts.

As Large Language Models (LLMs) grow dramatically in size, there is an increasing trend in compressing and speeding up these models. Previous studies have highlighted the usefulness of gradients for importance scoring in neural network compressing, especially in pruning medium-size networks. However, the substantial memory requirements involved in calculating gradients with backpropagation impede the utilization of gradients in guiding LLM pruning. As a result, most pruning strategies for LLMs rely on gradient-free criteria, such as weight magnitudes or a mix of magnitudes and activations. In this paper, we devise a hybrid pruning criterion, which appropriately integrates magnitude, activation, and gradient to capitalize on feature map sensitivity for pruning LLMs. To overcome memory requirement barriers, we estimate gradients using only forward passes. Based on this, we propose a Memory-effIcieNt structured prunIng procedure for LLMs (MINI-LLM) to remove no-critical channels and multi-attention heads. Experimental results demonstrate the superior performance of MINI-LLM over existing gradient-free methods on three LLMs: LLaMA, BLOOM, and OPT across various downstream tasks (classification, multiple-choice, and generation), while MINI-LLM maintains a GPU memory footprint akin to gradient-free methods.

Large Language Models (LLMs) are trained on large-scale web data, which makes it difficult to grasp the contribution of each text. This poses the risk of leaking inappropriate data such as benchmarks, personal information, and copyrighted texts in the training data. Membership Inference Attacks (MIA), which determine whether a given text is included in the model's training data, have been attracting attention. Previous studies of MIAs revealed that likelihood-based classification is effective for detecting leaks in LLMs. However, the existing methods cannot be applied to some proprietary models like ChatGPT or Claude 3 because the likelihood is unavailable to the user. In this study, we propose a Sampling-based Pseudo-Likelihood (\textbf{SPL}) method for MIA (\textbf{SaMIA}) that calculates SPL using only the text generated by an LLM to detect leaks. The SaMIA treats the target text as the reference text and multiple outputs from the LLM as text samples, calculates the degree of $n$-gram match as SPL, and determines the membership of the text in the training data. Even without likelihoods, SaMIA performed on par with existing likelihood-based methods.

Large language models (LLMs) show an innate skill for solving language based tasks. But insights have suggested an inability to adjust for information or task-solving skills becoming outdated, as their knowledge, stored directly within their parameters, remains static in time. Tool use helps by offloading work to systems that the LLM can access through an interface, but LLMs that use them still must adapt to nonstationary environments for prolonged use, as new tools can emerge and existing tools can change. Nevertheless, tools require less specialized knowledge, therefore we hypothesize they are better suited for continual learning (CL) as they rely less on parametric memory for solving tasks and instead focus on learning when to apply pre-defined tools. To verify this, we develop a synthetic benchmark and follow this by aggregating existing NLP tasks to form a more realistic testing scenario. While we demonstrate scaling model size is not a solution, regardless of tool usage, continual learning techniques can enable tool LLMs to both adapt faster while forgetting less, highlighting their potential as continual learners.

Recently, considerable efforts have been directed towards compressing Large Language Models (LLMs), which showcase groundbreaking capabilities across diverse applications but entail significant deployment costs due to their large sizes. Meanwhile, much less attention has been given to mitigating the costs associated with deploying multiple LLMs of varying sizes despite its practical significance. Thus, this paper introduces \emph{any-precision LLM}, extending the concept of any-precision DNN to LLMs. Addressing challenges in any-precision LLM, we propose a lightweight method for any-precision quantization of LLMs, leveraging a post-training quantization framework, and develop a specialized software engine for its efficient serving. As a result, our solution significantly reduces the high costs of deploying multiple, different-sized LLMs by overlaying LLMs quantized to varying bit-widths, such as 3, 4, ..., $n$ bits, into a memory footprint comparable to a single $n$-bit LLM. All the supported LLMs with varying bit-widths demonstrate state-of-the-art model quality and inference throughput, proving itself to be a compelling option for deployment of multiple, different-sized LLMs. The source code will be publicly available soon.

Speculative Decoding is a widely used technique to speed up inference for Large Language Models (LLMs) without modifying its outcome. When performing inference on an LLM, speculative decoding uses a smaller draft model which generates speculative tokens and then uses the target LLM to verify those draft tokens. The speedup provided by speculative decoding heavily depends on the choice of the draft model. It has been widely suggested to select a draft model that provides a high probability of the generated token being accepted by the LLM to achieve the highest throughput. However, our experiments indicate the contrary with throughput diminishing as the probability of generated tokens to be accepted by the target model increases. To understand this phenomenon, we perform extensive experiments to characterize the different factors that affect speculative decoding and how those factors interact and affect the speedups. Based on our experiments we describe an analytical model which can be used to decide the right draft model for a given workload. Further, using our insights we design a new draft model for LLaMA-65B which can provide 30% higher throughput than existing draft models.

Many applications must provide low-latency LLM service to users or risk unacceptable user experience. However, over-provisioning resources to serve fluctuating request patterns is often prohibitively expensive. In this work, we present a best-effort serving system that employs deep reinforcement learning to adjust service quality based on the task distribution and system load. Our best-effort system can maintain availability with over 10x higher client request rates, serves above 96% of peak performance 4.1x more often, and serves above 98% of peak performance 2.3x more often than static serving on unpredictable workloads. Our learned router is robust to shifts in both the arrival and task distribution. Compared to static serving, learned best-effort serving allows for cost-efficient serving through increased hardware utility. Additionally, we argue that learned best-effort LLM serving is applicable in wide variety of settings and provides application developers great flexibility to meet their specific needs.

Aligning large language models (LLMs) with human values is a vital task for LLM practitioners. Current alignment techniques have several limitations: (1) requiring a large amount of annotated data; (2) demanding heavy human involvement; (3) lacking a systematic mechanism to continuously improve. In this work, we study aligning LLMs to a new domain with limited samples (e.g. < 100). We propose an algorithm that can self-align LLMs iteratively without active human involvement. Unlike existing works, our algorithm relies on neither human-crafted instructions nor labeled rewards, significantly reducing human involvement. In addition, our algorithm can self-improve the alignment continuously. The key idea is to first retrieve high-quality samples related to the target domain and use them as In-context Learning examples to generate more samples. Then we use the self-generated samples to finetune the LLM iteratively. We show that our method can unlock the LLMs' self-generalization ability to perform alignment with near-zero human supervision. We test our algorithm on three benchmarks in safety, truthfulness, and instruction-following, and show good performance in alignment, domain adaptability, and scalability.