Daily Papers

Reinforcement Learning (RL) has become the most effective post-training approach for improving the capabilities of Large Language Models (LLMs). In practice, because of the high demands on latency and memory, it is particularly challenging to develop an efficient RL framework that reliably manages policy models with hundreds to thousands of billions of parameters. In this paper, we present LlamaRL, a fully distributed, asynchronous RL framework optimized for efficient training of large-scale LLMs with various model sizes (8B, 70B, and 405B parameters) on GPU clusters ranging from a handful to thousands of devices. LlamaRL introduces a streamlined, single-controller architecture built entirely on native PyTorch, enabling modularity, ease of use, and seamless scalability to thousands of GPUs. We also provide a theoretical analysis of LlamaRL's efficiency, including a formal proof that its asynchronous design leads to strict RL speed-up. Empirically during the Llama 3 post-training, by leveraging best practices such as colocated model offloading, asynchronous off-policy training, and distributed direct memory access for weight synchronization, LlamaRL achieves significant efficiency gains -- up to 10.7x speed-up compared to DeepSpeed-Chat-like systems on a 405B-parameter policy model. Furthermore, the efficiency advantage continues to grow with increasing model scale, demonstrating the framework's suitability for future large-scale RL training.

The increasing size of large language models (LLMs) challenges their usage on resource-constrained platforms. For example, memory on modern GPUs is insufficient to hold LLMs that are hundreds of Gigabytes in size. Offloading is a popular method to escape this constraint by storing weights of an LLM model to host CPU memory and SSD, then loading each weight to GPU before every use. In our case study of offloaded inference, we found that due to the low bandwidth between storage devices and GPU, the latency of transferring large model weights from its offloaded location to GPU memory becomes the critical bottleneck with actual compute taking nearly 0% of runtime. To effectively reduce the weight transfer latency, we propose a novel sparse format that compresses the unstructured sparse pattern of pruned LLM weights to non-zero values with high compression ratio and low decompression overhead. Endor achieves this by expressing the positions of non-zero elements with a bitmap. Compared to offloaded inference using the popular Huggingface Accelerate, applying Endor accelerates OPT-66B by 1.70x and Llama2-70B by 1.78x. When direct weight transfer from SSD to GPU is leveraged, Endor achieves 2.25x speedup on OPT-66B and 2.37x speedup on Llama2-70B.

The immense model sizes of large language models (LLMs) challenge deployment on memory-limited consumer GPUs. Although model compression and parameter offloading are common strategies to address memory limitations, compression can degrade quality, and offloading maintains quality but suffers from slow inference. Speculative decoding presents a promising avenue to accelerate parameter offloading, utilizing a fast draft model to propose multiple draft tokens, which are then verified by the target LLM in parallel with a single forward pass. This method reduces the time-consuming data transfers in forward passes that involve offloaded weight transfers. Existing methods often rely on pretrained weights of the same family, but require additional training to align with custom-trained models. Moreover, approaches that involve draft model training usually yield only modest speedups. This limitation arises from insufficient alignment with the target model, preventing higher token acceptance lengths. To address these challenges and achieve greater speedups, we propose SubSpec, a plug-and-play method to accelerate parameter offloading that is lossless and training-free. SubSpec constructs a highly aligned draft model by generating low-bit quantized substitute layers from offloaded target LLM portions. Additionally, our method shares the remaining GPU-resident layers and the KV-Cache, further reducing memory overhead and enhance alignment. SubSpec achieves a high average acceptance length, delivering 9.1x speedup for Qwen2.5 7B on MT-Bench (8GB VRAM limit) and an average of 12.5x speedup for Qwen2.5 32B on popular generation benchmarks (24GB VRAM limit).

Running large language models on consumer devices such as laptops and desktops is challenging because model weights often exceed GPU memory capacity, making offloading inference necessary to extend effective model capacity with CPU memory. Existing offloading systems, however, typically rely on coarse layer-level or expert-level scheduling, which overlooks substantial heterogeneity among tensors within the same layer and adapts poorly to changing hardware load conditions on such devices. This paper presents ATSInfer, a hybrid CPU-GPU inference system for consumer devices that performs offloading at tensor granularity. ATSInfer combines static tensor placement with load-aware dynamic transfer, and introduces asynchronous CPU-GPU coordination to efficiently schedule hardware storage, data movement, and computation across heterogeneous backends. We implement ATSInfer and evaluate it on representative consumer platforms using both dense and MoE models. Compared with existing systems, ATSInfer improves prefill throughput by up to 1.94times and decode throughput by up to 3.29times, while also increasing GPU utilization and making more effective use of PCIe bandwidth. These results show that ATSInfer can substantially improve the user experience of local LLM deployment on personal consumer devices.

Sparse Mixture of Experts (MoE) models, while outperforming dense Large Language Models (LLMs) in terms of performance, face significant deployment challenges during inference due to their high memory demands. Existing offloading techniques, which involve swapping activated and idle experts between the GPU and CPU, often suffer from rigid expert caching mechanisms. These mechanisms fail to adapt to dynamic routing, leading to inefficient cache utilization, or incur prohibitive costs for prediction training. To tackle these inference-specific challenges, we introduce ExpertFlow, a comprehensive system specifically designed to enhance inference efficiency by accommodating flexible routing and enabling efficient expert scheduling between CPU and GPU. This reduces overhead and boosts system performance. Central to our approach is a predictive routing path-based offloading mechanism that utilizes a lightweight predictor to accurately forecast routing paths before computation begins. This proactive strategy allows for real-time error correction in expert caching, significantly increasing cache hit ratios and reducing the frequency of expert transfers, thereby minimizing I/O overhead. Additionally, we implement a dynamic token scheduling strategy that optimizes MoE inference by rearranging input tokens across different batches. This method not only reduces the number of activated experts per batch but also improves computational efficiency. Our extensive experiments demonstrate that ExpertFlow achieves up to 93.72\% GPU memory savings and enhances inference speed by 2 to 10 times compared to baseline methods, highlighting its effectiveness and utility as a robust solution for resource-constrained inference scenarios.

Large Language Models (LLMs) have demonstrated impressive performance across various tasks, and their application in edge scenarios has attracted significant attention. However, sparse-activated Mixture-of-Experts (MoE) models, which are well suited for edge scenarios, have received relatively little attention due to their high memory demands. Offload-based methods have been proposed to address this challenge, but they face difficulties with expert prediction. Inaccurate expert predictions can result in prolonged inference delays. To promote the application of MoE models in edge scenarios, we propose Fate, an offloading system designed for MoE models to enable efficient inference in resource-constrained environments. The key insight behind Fate is that gate inputs from adjacent layers can be effectively used for expert prefetching, achieving high prediction accuracy without additional GPU overhead. Furthermore, Fate employs a shallow-favoring expert caching strategy that increases the expert hit rate to 99\%. Additionally, Fate integrates tailored quantization strategies for cache optimization and IO efficiency. Experimental results show that, compared to Load on Demand and Expert Activation Path-based method, Fate achieves up to 4.5x and 1.9x speedups in prefill speed and up to 4.1x and 2.2x speedups in decoding speed, respectively, while maintaining inference quality. Moreover, Fate's performance improvements are scalable across different memory budgets.

Large Language Models (LLMs) have gained immense success in revolutionizing various applications, including content generation, search and recommendation, and AI-assisted operation. To reduce high training costs, Mixture-of-Experts (MoE) architecture has become a popular backbone for modern LLMs. However, despite the benefits, serving MoE-based LLMs experience severe memory inefficiency due to sparsely activated experts. Recent studies propose to offload inactive experts from GPU memory to CPU memory to improve the serving efficiency of MoE models. However, they either incur high inference latency or high model memory footprints due to coarse-grained designs. To tame the latency-memory trade-off in MoE serving, we present fMoE, a fine-grained expert offloading system for MoE serving that achieves low inference latency with memory efficiency. We design fMoE to extract fine-grained expert selection patterns from MoE models and semantic hints from input prompts to efficiently guide expert prefetching, caching, and offloading decisions. fMoE is prototyped on top of HuggingFace Transformers and deployed on a six-GPU testbed. Experiments with open-source MoE models and real-world workloads show that fMoE reduces inference latency by 47% and improves expert hit rate by 36% over state-of-the-art solutions.

With the growth of model sizes and the scale of their deployment, their sheer size burdens the infrastructure requiring more network and more storage to accommodate these. While there is a vast model compression literature deleting parts of the model weights for faster inference, we investigate a more traditional type of compression - one that represents the model in a compact form and is coupled with a decompression algorithm that returns it to its original form and size - namely lossless compression. We present ZipNN a lossless compression tailored to neural networks. Somewhat surprisingly, we show that specific lossless compression can gain significant network and storage reduction on popular models, often saving 33% and at times reducing over 50% of the model size. We investigate the source of model compressibility and introduce specialized compression variants tailored for models that further increase the effectiveness of compression. On popular models (e.g. Llama 3) ZipNN shows space savings that are over 17% better than vanilla compression while also improving compression and decompression speeds by 62%. We estimate that these methods could save over an ExaByte per month of network traffic downloaded from a large model hub like Hugging Face.

We present Nemotron-Labs-3-Puzzle-75B-A9B, a compressed variant of Nemotron-3-Super optimized for interactive deployment. We designed the model to maximize server throughput under high user throughput constraints. In interactive serving workloads on a single 8xB200 node, Puzzle-75B-A9B achieves approximately 2x higher server throughput than Nemotron-3-Super at matched user throughput constraints. In ultra-long-context deployment on a single H100 GPU, the compressed model increases 1M-token concurrency from 1 request to 8 requests. Puzzle-75B-A9B is constructed using a multi-stage pipeline that combines the Iterative Puzzle compression framework with knowledge distillation, reinforcement learning, quantization, and a Multi-Token Prediction head. The compression process jointly optimizes heterogeneous MoE pruning, active parameter budget, and Mamba pruning to improve inference efficiency while preserving model quality. We evaluate Puzzle-75B-A9B on a broad suite of reasoning, coding, multilingual, long-context, and agentic benchmarks. Despite substantial compression, the model retains strong downstream accuracy relative to the parent model across a wide range of tasks. These results demonstrate that large hybrid MoE models can be substantially optimized for deployment efficiency while maintaining strong downstream capability.

With the widespread adoption of Large Language Models (LLMs), many deep learning practitioners are looking for strategies of running these models more efficiently. One such strategy is to use sparse Mixture-of-Experts (MoE) - a type of model architectures where only a fraction of model layers are active for any given input. This property allows MoE-based language models to generate tokens faster than their dense counterparts, but it also increases model size due to having multiple experts. Unfortunately, this makes state-of-the-art MoE language models difficult to run without high-end GPUs. In this work, we study the problem of running large MoE language models on consumer hardware with limited accelerator memory. We build upon parameter offloading algorithms and propose a novel strategy that accelerates offloading by taking advantage of innate properties of MoE LLMs. Using this strategy, we build can run Mixtral-8x7B with mixed quantization on desktop hardware and free-tier Google Colab instances.

When considering a model architecture, there are several ways to reduce its memory footprint. Historically, popular approaches included selecting smaller architectures and creating sparse networks through pruning. More recently, randomized parameter-sharing (RPS) methods have gained traction for model compression at start of training. In this paper, we comprehensively assess the trade-off between memory and accuracy across RPS, pruning techniques, and building smaller models. Our findings demonstrate that RPS, which is both data and model-agnostic, consistently outperforms/matches smaller models and all moderately informed pruning strategies, such as MAG, SNIP, SYNFLOW, and GRASP, across the entire compression range. This advantage becomes particularly pronounced in higher compression scenarios. Notably, even when compared to highly informed pruning techniques like Lottery Ticket Rewinding (LTR), RPS exhibits superior performance in high compression settings. This points out inherent capacity advantage that RPS enjoys over sparse models. Theoretically, we establish RPS as a superior technique in terms of memory-efficient representation when compared to pruning for linear models. This paper argues in favor of paradigm shift towards RPS based models. During our rigorous evaluation of RPS, we identified issues in the state-of-the-art RPS technique ROAST, specifically regarding stability (ROAST's sensitivity to initialization hyperparameters, often leading to divergence) and Pareto-continuity (ROAST's inability to recover the accuracy of the original model at zero compression). We provably address both of these issues. We refer to the modified RPS, which incorporates our improvements, as STABLE-RPS.

The high memory and computation demand of large language models (LLMs) makes them challenging to be deployed on consumer devices due to limited GPU memory. Offloading can mitigate the memory constraint but often suffers from low GPU utilization, leading to low inference efficiency. In this work, we propose a novel framework, called pipelined offloading (PIPO), for efficient inference on consumer devices. PIPO designs a fine-grained offloading pipeline, complemented with optimized data transfer and computation, to achieve high concurrency and efficient scheduling for inference. Experimental results show that compared with state-of-the-art baseline, PIPO increases GPU utilization from below 40% to over 90% and achieves up to 3.1times higher throughput, running on a laptop equipped with a RTX3060 GPU of 6GB memory.

Large Language Models (LLMs) present a critical trade-off between inference quality and computational cost: larger models offer superior capabilities but incur significant latency, while smaller models are faster but less powerful. Existing serving strategies often employ fixed model scales or static two-stage speculative decoding, failing to dynamically adapt to the varying complexities of user requests or fluctuations in system performance. This paper introduces , a novel framework that reimagines LLM inference as an adaptive routing problem solved through multi-level speculative decoding. dynamically constructs and optimizes inference "paths" (chains of models) based on real-time feedback, addressing the limitations of static approaches. Our contributions are threefold: (1) An adaptive model chain scheduling mechanism that leverages performance profiling (execution times) and predictive similarity metrics (derived from token distribution divergence) to continuously select the optimal sequence of draft and verifier models, minimizing predicted latency per generated token. (2) A multi-level collaborative verification framework where intermediate models within the selected chain can validate speculative tokens, reducing the verification burden on the final, most powerful target model. (3) A synchronized state management system providing efficient, consistent KV cache handling across heterogeneous models in the chain, including precise, low-overhead rollbacks tailored for asynchronous batch processing inherent in multi-level speculation. Preliminary experiments demonstrate the validity of our method.

Modern model hubs, such as Hugging Face, store tens of petabytes of LLMs, with fine-tuned variants vastly outnumbering base models and dominating storage consumption. Existing storage reduction techniques -- such as deduplication and compression -- are either LLM-oblivious or not compatible with each other, limiting data reduction effectiveness. Our large-scale characterization study across all publicly available Hugging Face LLM repositories reveals several key insights: (1) fine-tuned models within the same family exhibit highly structured, sparse parameter differences suitable for delta compression; (2) bitwise similarity enables LLM family clustering; and (3) tensor-level deduplication is better aligned with model storage workloads, achieving high data reduction with low metadata overhead. Building on these insights, we design BitX, an effective, fast, lossless delta compression algorithm that compresses XORed difference between fine-tuned and base LLMs. We build ZipLLM, a model storage reduction pipeline that unifies tensor-level deduplication and lossless BitX compression. By synergizing deduplication and compression around LLM family clustering, ZipLLM reduces model storage consumption by 54%, over 20% higher than state-of-the-art deduplication and compression approaches.

Mixture-of-Experts (MoE) enables efficient scaling of large language models (LLMs) with sparsely activated experts during inference. To effectively deploy large MoE models on memory-constrained devices, many systems introduce *expert offloading* that caches a subset of experts in fast memory, leaving others on slow memory to run on CPU or load on demand. While some research has exploited the locality of expert activations, where consecutive tokens activate similar experts, the degree of this **local routing consistency** varies across models and remains understudied. In this paper, we propose two metrics to measure local routing consistency of MoE models: (1) **Segment Routing Best Performance (SRP)**, which evaluates how well a fixed group of experts can cover the needs of a segment of tokens, and (2) **Segment Cache Best Hit Rate (SCH)**, which measures the optimal segment-level cache hit rate under a given cache size limit. We analyzed 20 MoE LLMs with diverse sizes and architectures and found that models that apply MoE on every layer and do not use shared experts exhibit the highest local routing consistency. We further showed that domain-specialized experts contribute more to routing consistency than vocabulary-specialized ones, and that most models can balance between cache effectiveness and efficiency with cache sizes approximately 2x the active experts. These findings pave the way for memory-efficient MoE design and deployment without compromising inference speed. We publish the code for replicating experiments at https://github.com/ljcleo/moe-lrc .

The growing prevalence of Large Language Models (LLMs) and their substantial memory requirements have prompted renewed interest in CPU offloading as a method to compensate for limited GPU memory. In particular, when CPU memory is leveraged to temporarily store intermediate states of LLMs, CPU memory becomes a new bottleneck and soon reaches the capacity limitation of commodity CPUs. In this work, we investigate the effectiveness of Compute Express Link (CXL) add-in card (AIC) memory as an extension to CPU memory, enabling larger model sizes and longer context lengths during fine-tuning. Through extensive benchmarking, this study quantifies the performance overhead introduced by transferring data between CXL memory, CPU, and GPUs, focusing on how concurrency and data volume influence bandwidth utilization and latency. This study also compares CPUbased optimizer steps when model parameters, gradients, and optimizer states reside in local memory versus CXL memory, revealing that naive adoption of CXL often degrades performance during the optimizer phase. To overcome these challenges, this study proposes a CXL-aware allocation to strategically partition CPU offloading workloads across both local and CXL memory. This study further demonstrates that employing multiple AICs significantly reduces bandwidth contention, thus improving scalability. Experimental results show that these optimizations enable efficient long-context LLM fine-tuning, underscoring CXL as a promising avenue for unlocking the full potential of CPU offloading in long-context LLM fine-tuning.

Recent advances in large language models have led to specialized models excelling in specific domains, creating a need for efficient model merging techniques. While traditional merging approaches combine parameters into a single static model, they often compromise task-specific performance. However, task-specific routing methods maintain accuracy but introduce substantial storage overhead. We present 1bit-Merging, a novel framework that integrates task-specific routing with 1-bit quantized task vectors to balance performance and storage efficiency. Our approach leverages the observation that different task-specific models store knowledge in distinct layers-chat models primarily in attention layers and math/code models in MLP layers-enabling targeted compression strategies. Through extensive experiments with LLaMA2 and Mistral model families across chat, mathematical reasoning, and code generation tasks, we demonstrate that 1bit-Merging achieves comparable or superior performance to existing methods while significantly reducing storage requirements. Our framework offers a practical solution for combining specialized models while maintaining their individual strengths and addressing the storage challenges of current approaches.

This paper presents MoE-Infinity, a cost-efficient mixture-of-expert (MoE) serving system that realizes activation-aware expert offloading. MoE-Infinity features sequence-level expert activation tracing, a new approach adept at identifying sparse activations and capturing the temporal locality of MoE inference. By analyzing these traces, MoE-Infinity performs novel activation-aware expert prefetching and caching, substantially reducing the latency overheads usually associated with offloading experts for improved cost performance. Extensive experiments in a cluster show that MoE-Infinity outperforms numerous existing systems and approaches, reducing latency by 4 - 20X and decreasing deployment costs by over 8X for various MoEs. MoE-Infinity's source code is publicly available at https://github.com/TorchMoE/MoE-Infinity

Large models have achieved remarkable performance across various tasks, yet they incur significant computational costs and privacy concerns during both training and inference. Distributed deployment has emerged as a potential solution, but it necessitates the exchange of intermediate information between model segments, with feature representations serving as crucial information carriers. To optimize information exchange, feature coding methods are applied to reduce transmission and storage overhead. Despite its importance, feature coding for large models remains an under-explored area. In this paper, we draw attention to large model feature coding and make three contributions to this field. First, we introduce a comprehensive dataset encompassing diverse features generated by three representative types of large models. Second, we establish unified test conditions, enabling standardized evaluation pipelines and fair comparisons across future feature coding studies. Third, we introduce two baseline methods derived from widely used image coding techniques and benchmark their performance on the proposed dataset. These contributions aim to advance the field of feature coding, facilitating more efficient large model deployment. All source code and the dataset are now available at https://github.com/chansongoal/FCM-LM/tree/master{https://github.com/chansongoal/FCM-LM/tree/master}.

Large-scale model training has been a playing ground for a limited few requiring complex model refactoring and access to prohibitively expensive GPU clusters. ZeRO-Offload changes the large model training landscape by making large model training accessible to nearly everyone. It can train models with over 13 billion parameters on a single GPU, a 10x increase in size compared to popular framework such as PyTorch, and it does so without requiring any model change from the data scientists or sacrificing computational efficiency. ZeRO-Offload enables large model training by offloading data and compute to CPU. To preserve compute efficiency, it is designed to minimize the data movement to/from GPU, and reduce CPU compute time while maximizing memory savings on GPU. As a result, ZeRO-Offload can achieve 40 TFlops/GPU on a single NVIDIA V100 GPU for 10B parameter model compared to 30TF using PyTorch alone for a 1.4B parameter model, the largest that can be trained without running out of memory. ZeRO-Offload is also designed to scale on multiple-GPUs when available, offering near linear speedup on up to 128 GPUs. Additionally, it can work together with model parallelism to train models with over 70 billion parameters on a single DGX-2 box, a 4.5x increase in model size compared to using model parallelism alone. By combining compute and memory efficiency with ease-of-use, ZeRO-Offload democratizes large-scale model training making it accessible to even data scientists with access to just a single GPU.

KV cache has traditionally been stored in GPU memory to accelerate the decoding phase of large language model (LLM) inference. However, it is increasingly necessary to move KV caches outside GPU devices, to enable cache reuse across different queries and inference engines. Our real-world usage statistics confirm this trend: over time, the total KV cache stored by users has grown rapidly, far exceeding the capacity of GPU memory. Despite this need, there lacks an efficient solution for offloading and transferring KV caches. We present LMCACHE, the first and so far the most efficient open-source KV caching solution, which extracts and stores KV caches generated by modern LLM engines (vLLM and SGLang) out of the GPU memory and shares them across engines and queries. LMCACHE supports both cache offloading (prefix reuse across queries) and prefill-decode (PD) disaggregation (cross-engine/GPU cache transfer). LMCACHE's high performance and wide adoption stem from the following contributions: (1) highly optimized KV cache data movement powered by batched data movement operations, compute and I/O pipelining; (2) a modular KV cache connector component, decoupling LMCACHE from the rapid evolution of inference engines; (3) a first-class control API for flexible cache orchestration across GPU, CPU, storage, and network layers. Our evaluation shows that combining LMCACHE with vLLM achieves up to 15x improvement in throughput across workloads such as multi-round question answering and document analysis. Large-scale adoption of LMCACHE in enterprise settings provides us valuable insights, for example, fetching KV cache from remote storage has unsurprisingly benefits to prefill delay, and that context truncation, which is a widely applied technique in industry, can greatly reduce prefix cache hit ratio by half. The source code of LMCACHE is at: https://github.com/LMCache/LMCache.

Efficient inference with Large Language Models (LLMs) increasingly relies on Key-Value (KV) caches to store previously computed key and value vectors at each layer. These caches are essential to minimize redundant computation during autoregressive token generation, lowering computational complexity from quadratic to linear. However, the growth of KV caches has posed significant system-level challenges, particularly as model sizes increase, context lengths grow, and concurrent requests compete for limited memory resources. Even though several recent frameworks for KV cache management have emerged, their comparative trade-offs in memory consumption and inference performance have not been fully understood, especially under varying request sizes and model configurations. In this work, we conduct an empirical study of three state-of-the-art KV cache management frameworks: vLLM, InfiniGen, and H2O. These frameworks employ techniques such as tensor offloading, token eviction heuristics, and speculative scheduling to balance memory usage and performance. We evaluate their performance in terms of a range of metrics such as latency, throughput, and memory usage across a spectrum of key parameters including request rates, model sizes, and sparsity levels. Our results pinpoint the conditions for each framework to perform the best, revealing the most suitable selection and configuration of KV cache strategies under memory and performance constraints.

The promising applications of large language models are often limited by the constrained GPU memory capacity available on edge devices. Mixture-of-Experts (MoE) models help address this issue by activating only a subset of the model's parameters during computation. This approach allows the unused parameters to be offloaded to host memory, thereby reducing the overall GPU memory demand. However, existing cache-based offloading solutions handle cache misses reactively, which significantly impacts system performance. In this paper, we introduce ProMoE, a novel proactive caching system that utilizes intermediate results to predict subsequent expert usage. By proactively fetching experts in advance, ProMoE eliminates passive cache misses, removes loading time from the critical path, and reduces the performance overhead associated with offloading. Our evaluations demonstrate that ProMoE achieves an average speedup of 2.20x (up to 3.21x) and 2.07x (up to 5.02x) in the prefill and decode stages, respectively, compared to existing offloading solutions.

Owing to the huge success of generative artificial intelligence (AI), large language models (LLMs) have emerged as a core subclass, underpinning applications such as question answering, text generation, and code completion. While fine-tuning these models on domain-specific data can yield significant performance gains, it also poses daunting computational challenges, especially for researchers and small organizations with limited hardware resources. Although SSD offloading (i.e., ZeRO-Infinity) has emerged as a viable strategy to overcome the GPU memory barrier via leveraging both system memory (i.e., CPU DRAM) and storage space (i.e., solid-state devices, SSDs), its design primarily targets model-centric performance issues. As a result, key system-level issues, including system memory fragmentation, inefficient pinned buffer allocation, peak CPU usage spikes, and file system overhead, remain unaddressed, stifling scalability and inflating costs. Such an observation motivates this paper to introduce MemAscend, a framework that systematically tackles the underexplored system memory bottlenecks in SSD-offloaded LLM training, with a focus on resource-constrained environments. By streamlining pinned-memory allocation, eradicating fragmentation, and mitigating peak overhead, MemAscend reclaims a substantial system memory budget, enabling larger models, longer context windows, and higher batch sizes without exceeding modest hardware limits. Across diverse LLM benchmarks, MemAscend reduces peak system-memory consumption by an average of 55.7% compared with standard SSD offloading techniques, lowering the hardware barrier for fine-tuning and unlocking new possibilities for cost-effective large-scale training on limited-resource machines.

Mixture-of-Experts (MoE) models enable scalable performance but face severe memory constraints on edge devices. Existing offloading strategies struggle with I/O bottlenecks due to the dynamic, low-information nature of autoregressive expert activation. In this paper, we propose to repurpose Speculative Decoding (SD) not merely as a compute accelerator, but as an informative lookahead sensor for memory management, supported by our theoretical and empirical analyses. Hence, we introduce MoE-SpAc, an MoE inference framework that integrates a Speculative Utility Estimator to track expert demand, a Heterogeneous Workload Balancer to dynamically partition computation via online integer optimization, and an Asynchronous Execution Engine to unify the prefetching and eviction in the same utility space. Extensive experiments on seven benchmarks demonstrate that MoE-SpAc achieves a 42% improvement in TPS over the SOTA SD-based baseline, and an average 4.04x speedup over all standard baselines. Code is available at https://github.com/lshAlgorithm/MoE-SpAc .

The Mixture-of-Experts (MoE) architecture has demonstrated significant advantages in the era of Large Language Models (LLMs), offering enhanced capabilities with reduced inference costs. However, deploying MoE-based LLMs on memoryconstrained edge devices remains challenging due to their substantial memory requirements. While existing expertoffloading methods alleviate the memory requirements, they often incur significant expert-loading costs or compromise model accuracy. We present HOBBIT, a mixed precision expert offloading system to enable flexible and efficient MoE inference. Our key insight is that dynamically replacing less critical cache-miss experts with low precision versions can substantially reduce expert-loading latency while preserving model accuracy. HOBBIT introduces three innovative techniques that map the natural hierarchy of MoE computation: (1) a token-level dynamic expert loading mechanism, (2) a layer-level adaptive expert prefetching technique, and (3) a sequence-level multidimensional expert caching policy. These innovations fully leverage the benefits of mixedprecision expert inference. By implementing HOBBIT on top of the renowned LLM inference framework Llama.cpp, we evaluate its performance across different edge devices with representative MoE models. The results demonstrate that HOBBIT achieves up to a 9.93x speedup in decoding compared to state-of-the-art MoE offloading systems.

We are entering a new era of composite model architectures that integrate diverse components such as vision encoders, language backbones, diffusion and flow heads, audio codecs, action generators, and world-model predictors. Such architectures underpin a broad class of multimodal models, including unified multimodal models, omni models, speech-language models, vision-language-action policies, and world models. However, existing model serving frameworks were built on narrow assumptions about model structure, making them ill-suited to accommodate this new architectural diversity. Here we present M*, a universal serving system for efficient serving of composite AI models. M* represents models as dataflow graphs, processing requests spanning diverse modalities and tasks as traversals over these graphs. The core insight is a modular abstraction that supports arbitrary composition of model components, flexible placement onto a physical cluster, and model-agnostic optimizations within a distributed runtime. We call this abstraction the Walk Graph and show how it can concisely capture composite models from a broad range of families. We instantiate M* on representative models and find that it achieves, on average, 20% lower end-to-end latency than vLLM-Omni for text-to-image workloads on BAGEL, while delivering up to 2.9x lower real-time factor and 2.7x higher throughput for text-to-speech workloads on Qwen3-Omni. M* also outperforms the V-JEPA 2-AC rollout baseline for robotic planning by up to 12.5x. Thus, our work paves the road towards more efficient serving of complex models with minimal developer effort.

Fine-tuning large language models (LLMs) for downstream tasks can greatly improve model quality, however serving many different fine-tuned LLMs concurrently for users in multi-tenant environments is challenging. Dedicating GPU memory for each model is prohibitively expensive and naively swapping large model weights in and out of GPU memory is slow. Our key insight is that fine-tuned models can be quickly swapped in and out of GPU memory by extracting and compressing the delta between each model and its pre-trained base model. We propose DeltaZip, an LLM serving system that efficiently serves multiple full-parameter fine-tuned models concurrently by aggressively compressing model deltas by a factor of 6times to 8times while maintaining high model quality. DeltaZip increases serving throughput by 1.5times to 3times and improves SLO attainment compared to a vanilla HuggingFace serving system.

Lossless model compression holds tremendous promise for alleviating the memory and bandwidth bottlenecks in bit-exact Large Language Model (LLM) serving. However, existing approaches often result in substantial inference slowdowns due to fundamental design mismatches with GPU architectures: at the kernel level, variable-length bitstreams produced by traditional entropy codecs break SIMT parallelism; at the system level, decoupled pipelines lead to redundant memory traffic. We present ZipServ, a lossless compression framework co-designed for efficient LLM inference. ZipServ introduces Tensor-Core-Aware Triple Bitmap Encoding (TCA-TBE), a novel fixed-length format that enables constant-time, parallel decoding, together with a fused decompression-GEMM (ZipGEMM) kernel that decompresses weights on-the-fly directly into Tensor Core registers. This "load-compressed, compute-decompressed" design eliminates intermediate buffers and maximizes compute intensity. Experiments show that ZipServ reduces the model size by up to 30%, achieves up to 2.21x kernel-level speedup over NVIDIA's cuBLAS, and expedites end-to-end inference by an average of 1.22x over vLLM. ZipServ is the first lossless compression system that provides both storage savings and substantial acceleration for LLM inference on GPUs.

Large Language Models (LLMs) and other large foundation models have achieved noteworthy success, but their size exacerbates existing resource consumption and latency challenges. In particular, the large-scale deployment of these models is hindered by the significant resource requirements during inference. In this paper, we study two approaches for mitigating these challenges: employing a cache to store previous queries and learning a model multiplexer to choose from an ensemble of models for query processing. Theoretically, we provide an optimal algorithm for jointly optimizing both approaches to reduce the inference cost in both offline and online tabular settings. By combining a caching algorithm, namely Greedy Dual Size with Frequency (GDSF) or Least Expected Cost (LEC), with a model multiplexer, we achieve optimal rates in both offline and online settings. Empirically, simulations show that the combination of our caching and model multiplexing algorithms greatly improves over the baselines, with up to 50times improvement over the baseline when the ratio between the maximum cost and minimum cost is 100. Experiments on real datasets show a 4.3times improvement in FLOPs over the baseline when the ratio for FLOPs is 10, and a 1.8times improvement in latency when the ratio for average latency is 1.85.

In recent times, the emergence of Large Language Models (LLMs) has resulted in increasingly larger model size, posing challenges for inference on low-resource devices. Prior approaches have explored offloading to facilitate low-memory inference but often suffer from efficiency due to I/O bottlenecks. To achieve low-latency LLMs inference on resource-constrained devices, we introduce HeteGen, a novel approach that presents a principled framework for heterogeneous parallel computing using CPUs and GPUs. Based on this framework, HeteGen further employs heterogeneous parallel computing and asynchronous overlap for LLMs to mitigate I/O bottlenecks. Our experiments demonstrate a substantial improvement in inference speed, surpassing state-of-the-art methods by over 317% at most.

The rapid growth of large language models (LLMs) with diverse capabilities, costs, and domains has created a critical need for intelligent model selection at inference time. While smaller models suffice for routine queries, complex tasks demand more capable models. However, static model deployment does not account for the complexity and domain of incoming queries, leading to suboptimal performance and increased costs. Dynamic routing systems that adaptively select models based on query characteristics have emerged as a solution to this challenge. We provide a systematic analysis of state-of-the-art multi-LLM routing and cascading approaches. In contrast to mixture-of-experts architectures, which route within a single model, we study routing across multiple independently trained LLMs. We cover diverse routing paradigms, including query difficulty, human preferences, clustering, uncertainty quantification, reinforcement learning, multimodality, and cascading. For each paradigm, we analyze representative methods and examine key trade-offs. Beyond taxonomy, we introduce a conceptual framework that characterizes routing systems along three dimensions: when decisions are made, what information is used, and how they are computed. This perspective highlights that practical systems are often compositional, integrating multiple paradigms under operational constraints. Our analysis demonstrates that effective multi-LLM routing requires balancing competing objectives. Choosing the optimal routing strategy depends on deployment and computational constraints. Well-designed routing systems can outperform even the most powerful individual models by strategically leveraging specialized capabilities across models while maximizing efficiency gains. Meanwhile, open challenges remain in developing routing mechanisms that generalize across diverse architectures, modalities, and applications.

Long-context inference is increasingly constrained by the KV cache: resident memory grows with context length, and decoding becomes limited by repeated High Bandwidth Memory (HBM) streaming rather than arithmetic. Existing methods such as eviction, windowing, quantization, and offloading reduce footprint, but often leave the critical-path bottleneck only partially addressed, especially when compressed states must still be reconstructed into dense vectors during decoding. We present Spherical KV, a long-context inference method that treats KV allocation as a rate-distortion problem grounded in attention geometry for efficient decoding. The method is built on two ideas: (i) represent directional information cheaply in the decode hot loop, and (ii) allocate retention and precision according to estimated future utility. Its first component, Angle-Domain Attention (ADA), stores keys in a spherical parameterization consisting of a scalar radius and compact angle codes, and computes attention logits directly from these codes without reconstructing dense keys. This preserves a paged, block-local, fusion-friendly decode path and directly targets HBM traffic in realistic serving settings. Its second component, Rate-Distortion Retention (RDR), jointly chooses keep/drop decisions and precision tiers per token and head under a fixed budget, producing tier-homogeneous pages with lightweight metadata and coalesced reads. Together, ADA and RDR provide a deployment-oriented mechanism for reducing KV residency while preserving decode efficiency.

Large Language Models (LLMs) have revolutionized natural language processing by achieving state-of-the-art results across a variety of tasks. However, the computational demands of LLM inference, including high memory consumption and slow processing speeds, pose significant challenges for real-world applications, particularly on resource-constrained devices. Efficient inference is crucial for scaling the deployment of LLMs to a broader range of platforms, including mobile and edge devices. This survey explores contemporary techniques in model compression that address these challenges by reducing the size and computational requirements of LLMs while maintaining their performance. We focus on model-level compression methods, including quantization, knowledge distillation, and pruning, as well as system-level optimizations like KV cache efficient design. Each of these methodologies offers a unique approach to optimizing LLMs, from reducing numerical precision to transferring knowledge between models and structurally simplifying neural networks. Additionally, we discuss emerging trends in system-level design that further enhance the efficiency of LLM inference. This survey aims to provide a comprehensive overview of current advancements in model compression and their potential to make LLMs more accessible and practical for diverse applications.

LLMs are widely adopted in production, pushing inference systems to their limits. Disaggregated LLM serving (e.g., PD separation and KV state disaggregation) improves scalability and cost efficiency, but it also turns KV into an explicit payload crossing network and storage boundaries, making KV a dominant end-to-end bottleneck. Existing KV compression are typically static runtime configurations, despite production service context varies over time in workload mix, bandwidth, and SLO/quality budgets. As a result, a fixed choice can be suboptimal or even increase latency. We present KVServe, the first service-aware and adaptive KV communication compression framework for disaggregated LLM serving: KVServe (1) unifies KV compression into a modular strategy space with new components and cross-method recomposition; (2) introduces Bayesian Profiling Engine that efficiently searches this space and distills a 3D Pareto candidate set, reducing 50times offline search overhead; and (3) deploys a Service-Aware Online Controller that combines an analytical latency model with a lightweight bandit to select profiles under constraints and correct offline-to-online mismatch. Integrated into vLLM and evaluated across datasets, models, GPUs and networks, KVServe achieves up to 9.13times JCT speedup in PD-separated serving and up to 32.8times TTFT reduction in KV-disaggregated serving.

The emergence of large-scale Mixture of Experts (MoE) models has marked a significant advancement in artificial intelligence, offering enhanced model capacity and computational efficiency through conditional computation. However, the deployment and inference of these models present substantial challenges in terms of computational resources, latency, and energy efficiency. This comprehensive survey systematically analyzes the current landscape of inference optimization techniques for MoE models across the entire system stack. We first establish a taxonomical framework that categorizes optimization approaches into model-level, system-level, and hardware-level optimizations. At the model level, we examine architectural innovations including efficient expert design, attention mechanisms, various compression techniques such as pruning, quantization, and knowledge distillation, as well as algorithm improvement including dynamic routing strategies and expert merging methods. At the system level, we investigate distributed computing approaches, load balancing mechanisms, and efficient scheduling algorithms that enable scalable deployment. Furthermore, we delve into hardware-specific optimizations and co-design strategies that maximize throughput and energy efficiency. This survey not only provides a structured overview of existing solutions but also identifies key challenges and promising research directions in MoE inference optimization. Our comprehensive analysis serves as a valuable resource for researchers and practitioners working on large-scale deployment of MoE models in resource-constrained environments. To facilitate ongoing updates and the sharing of cutting-edge advances in MoE inference optimization research, we have established a repository accessible at https://github.com/MoE-Inf/awesome-moe-inference/.

Prefill-decode (PD) disaggregation has become the standard architecture for large-scale LLM serving, but in practice its deployment boundary is still determined by KVCache transfer. In conventional dense-attention models, prefill generates huge KVCache traffics that keep prefill and decode tightly coupled within a single high-bandwidth network domain, limiting heterogeneous deployment and resource elasticity. Recent hybrid-attention architectures substantially reduce KVCache size, making cross-cluster KVCache transport increasingly plausible. However, smaller KVCache alone does not make heterogeneous cross-datacenter PD serving practical: real workloads remain bursty, request lengths are highly skewed, prefix caches are unevenly distributed, and inter-cluster bandwidth fluctuates. A naive design that fully externalizes prefill can therefore still suffer from congestion, unstable queueing, and poor utilization. We present Prefill-as-a-Service (PrfaaS), a cross-datacenter serving architecture that selectively offloads long-context prefill to standalone, compute-dense prefill clusters and transfers the resulting KVCache over commodity Ethernet to local PD clusters for decode. Rather than treating reduced KVCache as sufficient, PrfaaS combines model-side KV efficiency with system-side selective offloading, bandwidth-aware scheduling, and cache-aware request placement. This design removes the requirement that heterogeneous accelerators share the same low-latency RDMA fabric, enabling independent scaling of prefill and decode capacity across loosely coupled clusters. In a case study using an internal 1T-parameter hybrid model, a PrfaaS-augmented heterogeneous deployment achieves 54% higher serving throughput and 64% lower P90 TTFT than a homogeneous PD baseline, with approximately 15% throughput gain at equal cost, while consuming only modest cross-datacenter bandwidth.

Large language models (LLMs) are useful in many NLP tasks and become more capable with size, with the best open-source models having over 50 billion parameters. However, using these 50B+ models requires high-end hardware, making them inaccessible to most researchers. In this work, we investigate methods for cost-efficient inference and fine-tuning of LLMs, comparing local and distributed strategies. We observe that a large enough model (50B+) can run efficiently even on geodistributed devices in a consumer-grade network. This could allow running LLM efficiently by pooling together idle compute resources of multiple research groups and volunteers. We address two open problems: (1) how to perform inference and fine-tuning reliably if any device can disconnect abruptly and (2) how to partition LLMs between devices with uneven hardware, joining and leaving at will. In order to do that, we develop special fault-tolerant inference algorithms and load-balancing protocols that automatically assign devices to maximize the total system throughput. We showcase these algorithms in Petals - a decentralized system that runs Llama 2 (70B) and BLOOM (176B) over the Internet up to 10x faster than offloading for interactive generation. We evaluate the performance of our system in simulated conditions and a real-world setup spanning two continents.

Multi-modal Large Language Models (MLLMs) serving systems commonly employ KV-cache compression to reduce memory footprint. However, existing compression methods introduce significant processing overhead and queuing delays, particularly in concurrent serving scenarios. We present FastCache, a novel serving framework that effectively addresses these challenges through two key innovations: (1) a dynamic batching strategy that optimizes request scheduling across prefill, compression, and decode stages, and (2) an efficient KV-cache memory pool mechanism that eliminates memory fragmentation while maintaining high GPU utilization. Our comprehensive experiments on the GQA and MileBench datasets demonstrate that FastCache achieves up to 19.3times reduction in Time-To-First-Token (TTFT) and 12.1times improvement in throughput compared to state-of-the-art baselines. The system maintains stable performance under high-concurrency scenarios (up to 40 req/s) while reducing average memory consumption by 20\%. These results establish FastCache as an efficient solution for real-world LLM serving systems with KV-cache compression.

Model pruning is a popular approach to enable the deployment of large deep learning models on edge devices with restricted computational or storage capacities. Although sparse models achieve performance comparable to that of their dense counterparts at the level of the entire dataset, they exhibit high accuracy drops for some data sub-groups. Existing methods to mitigate this disparate impact induced by pruning (i) rely on surrogate metrics that address the problem indirectly and have limited interpretability; or (ii) scale poorly with the number of protected sub-groups in terms of computational cost. We propose a constrained optimization approach that directly addresses the disparate impact of pruning: our formulation bounds the accuracy change between the dense and sparse models, for each sub-group. This choice of constraints provides an interpretable success criterion to determine if a pruned model achieves acceptable disparity levels. Experimental results demonstrate that our technique scales reliably to problems involving large models and hundreds of protected sub-groups.

Large language models (LLMs) have shown remarkable advances in supporting long-context comprehension and processing tasks. However, scaling the generation inference of LLMs to such long contexts incurs significant additional computation load, and demands a substantial GPU memory footprint to maintain the key-value (KV) cache of transformer-based LLMs. Existing KV cache compression methods, such as quantization, face memory bottlenecks as context length increases, while static-sized caches, such as eviction, suffer from inefficient policies. These limitations restrict deployment on consumer-grade devices like a single Nvidia 4090 GPU. To overcome this, we propose Locret, a framework for long-context LLM inference that introduces retaining heads to evaluate the causal importance of KV cache units, allowing for more accurate eviction within a fixed cache size. Locret is fine-tuned on top of the frozen backbone LLM using a minimal amount of data from standard long-context SFT datasets. During inference, we evict low-importance cache units along with a chunked prefill pattern, significantly reducing peak GPU memory usage. We conduct an extensive empirical study to evaluate Locret, where the experimental results show that Locret outperforms the recent competitive approaches, including InfLLM, Quantization, SirLLM, and MInference, in terms of memory efficiency and the quality of generated contents -- Locret achieves over a 20x and 8x KV cache compression ratio compared to the full KV cache for Phi-3-mini-128K and Llama-3.1-8B-instruct. Additionally, Locret can be combined with other methods, such as quantization and token merging. To our knowledge, Locret is the first framework capable of deploying Llama-3.1-8B or similar models on a single Nvidia 4090 GPU, enabling 128K long-context inference without compromising generation quality, and requiring little additional system optimizations.

Large models have delivered remarkable performance across a wide range of perception and generation tasks, yet practical deployment is increasingly constrained by computational and memory budgets, as well as privacy requirements. Split execution alleviates these constraints by partitioning computation across devices, but it inevitably introduces intensive transmission and storage of intermediate features. Unlike conventional feature coding for CNNs that typically targets homogeneous spatial activation maps, modern large models generate heterogeneous features with varying statistical distributions and compression tolerances, e.g., multi-level/multi-modal representations and autoregressive context caches. These characteristics necessitate treating large model feature coding (LaMoFC) as a fundamental system component and call for a systematic evaluation framework. In this paper, we present a comprehensive benchmark and evaluation framework for LaMoFC. We first build the feature dataset LaMoFCBench, covering diverse task requirements across 4 categories and 16 scenarios while integrating widelyadopted architectures and various split-computing settings. We then specify representative split points according to practical application scenarios to extract intermediate features, establishing a unified pipeline for fair and reproducible comparisons. Finally, we benchmark mainstream universal feature codecs, exposing the profound misalignment between existing coding paradigms and the heterogeneous nature of large model features. These findings reveal that LaMoFC demands a fundamental departure from existing paradigms, and LaMoFCBench provides the shared empirical foundation to drive this transition. The data and code will be available at https://github.com/lartpang/LaMoFCBench.

The deployment of efficient long-context LLMs in applications like autonomous agents, long-chain reasoning, and creative writing is fundamentally bottlenecked by the linear growth of KV cache memory. Existing compression and eviction methods often struggle to balance accuracy, compression ratio, and hardware efficiency. We propose DeltaKV, a residual-based KV cache compression framework motivated by two empirical findings: long-range inter-token similarity and highly shared latent components in KV representations. Instead of discarding tokens, DeltaKV encodes semantic residuals relative to retrieved historical references, preserving fidelity while substantially reducing storage. To translate compression gains into real system speedups, we further introduce Sparse-vLLM, a high-performance inference engine with decoupled memory management and kernels optimized for sparse and irregular KV layouts. Experiments show that DeltaKV reduces KV cache memory to 29\% of the original while maintaining near-lossless accuracy on LongBench, SCBench, and AIME. When integrated with Sparse-vLLM, it achieves up to 2times throughput improvement over vLLM in long-context scenarios, demonstrating a practical path toward scalable long-context LLM deployment. Code, model checkpoints, and datasets are available at https://github.com/CURRENTF/Sparse-vLLM.

Long-context language models exhibit impressive performance but remain challenging to deploy due to high GPU memory demands during inference. We propose Memory-efficient Offloaded Mini-sequence Inference (MOM), a method that partitions critical layers into smaller "mini-sequences" and integrates seamlessly with KV cache offloading. Experiments on various Llama, Qwen, and Mistral models demonstrate that MOM reduces peak memory usage by over 50\% on average. On Meta-Llama-3.2-8B, MOM extends the maximum context length from 155k to 455k tokens on a single A100 80GB GPU, while keeping outputs identical and not compromising accuracy. MOM also maintains highly competitive throughput due to minimal computational overhead and efficient last-layer processing. Compared to traditional chunked prefill methods, MOM achieves a 35\% greater context length extension. More importantly, our method drastically reduces prefill memory consumption, eliminating it as the longstanding dominant memory bottleneck during inference. This breakthrough fundamentally changes research priorities, redirecting future efforts from prefill-stage optimizations to improving decode-stage residual KV cache efficiency.

Large language models (LLMs) often exhibit complementary strengths. Model routing harnesses these strengths by dynamically directing each query to the most suitable model, given a candidate model pool. However, routing performance relies on accurate model representations, and adding new models typically requires retraining, limiting scalability. To address these challenges, we propose a novel routing method using in-context vectors to represent model capabilities. The method proceeds in two stages. First, queries are embedded and projected into vectors, with a projector and LLM-based router trained to reconstruct the original queries, aligning vector representations with the router's semantic space. Second, each candidate model is profiled on a query set, and the router learns -- based on in-context vectors of query and model performance -- to predict whether each model can correctly answer new queries. Extensive experiments demonstrate that our method achieves state-of-the-art routing performance in both in-distribution and out-of-distribution tasks. Moreover, our method allows for seamless integration of new models without retraining the router. The code is available at https://github.com/lalalamdbf/ICL-Router.

Large Language Models (LLMs) are increasingly deployed in both latency-sensitive online services and cost-sensitive offline workloads. Co-locating these workloads on shared serving instances can improve resource utilization, but directly applying this approach to Prefill/Decode (P/D) disaggregated systems introduces severe load imbalance, as fluctuating request mixes alter the intrinsic P/D ratio. Existing dynamic adjustment techniques cannot keep up with the bursty traffic patterns of online services. We propose a latency-constraint disaggregated architecture, which separates cluster resources into latency-strict and latency-relaxed pools based on task latency requirements. This design enables flexible placement of offline decode tasks, mitigating P/D imbalance while preserving online performance. To fully exploit this flexibility, we propose (1) a bottleneck-based scheduler guided by a Roofline-based performance model for performance bottleneck based scheduling, and (2) a fast preemption mechanism that strictly enforces Service Level Objectives (SLOs) for online requests. Experiments on real-world traces show that compared to existing offline system approaches, our method improves offline throughput by up to 3x, while maintaining online request SLOs.

The proliferation of large language models (LLMs) has led to the adoption of Mixture-of-Experts (MoE) architectures that dynamically leverage specialized subnetworks for improved efficiency and performance. Despite their benefits, MoE models face significant challenges during inference, including inefficient memory management and suboptimal batching, due to misaligned design choices between the model architecture and the system policies. Furthermore, the conventional approach of training MoEs from scratch is increasingly prohibitive in terms of cost. In this paper, we propose a novel framework Read-ME that transforms pre-trained dense LLMs into smaller MoE models (in contrast to "upcycling" generalist MoEs), avoiding the high costs of ground-up training. Our approach employs activation sparsity to extract experts. To compose experts, we examine the widely-adopted layer-wise router design and show its redundancy, and thus we introduce the pre-gating router decoupled from the MoE backbone that facilitates system-friendly pre-computing and lookahead scheduling, enhancing expert-aware batching and caching. Our codesign therefore addresses critical gaps on both the algorithmic and system fronts, establishing a scalable and efficient alternative for LLM inference in resource-constrained settings. Read-ME outperforms other popular open-source dense models of similar scales, achieving improvements of up to 10.1% on MMLU, and improving mean end-to-end latency up to 6.1%. Codes are available at: https://github.com/VITA-Group/READ-ME.

Transformers have revolutionized the machine learning landscape, gradually making their way into everyday tasks and equipping our computers with "sparks of intelligence". However, their runtime requirements have prevented them from being broadly deployed on mobile. As personal devices become increasingly powerful and prompt privacy becomes an ever more pressing issue, we explore the current state of mobile execution of Large Language Models (LLMs). To achieve this, we have created our own automation infrastructure, MELT, which supports the headless execution and benchmarking of LLMs on device, supporting different models, devices and frameworks, including Android, iOS and Nvidia Jetson devices. We evaluate popular instruction fine-tuned LLMs and leverage different frameworks to measure their end-to-end and granular performance, tracing their memory and energy requirements along the way. Our analysis is the first systematic study of on-device LLM execution, quantifying performance, energy efficiency and accuracy across various state-of-the-art models and showcases the state of on-device intelligence in the era of hyperscale models. Results highlight the performance heterogeneity across targets and corroborates that LLM inference is largely memory-bound. Quantization drastically reduces memory requirements and renders execution viable, but at a non-negligible accuracy cost. Drawing from its energy footprint and thermal behavior, the continuous execution of LLMs remains elusive, as both factors negatively affect user experience. Last, our experience shows that the ecosystem is still in its infancy, and algorithmic as well as hardware breakthroughs can significantly shift the execution cost. We expect NPU acceleration, and framework-hardware co-design to be the biggest bet towards efficient standalone execution, with the alternative of offloading tailored towards edge deployments.

Scaling long-context ability is essential for Large Language Models (LLMs). To amortize the memory consumption across multiple devices in long-context training, inter-data partitioning (a.k.a. Data Parallelism) and intra-data partitioning (a.k.a. Context Parallelism) are commonly used. Current training frameworks predominantly treat the two techniques as orthogonal, and establish static communication groups to organize the devices as a static mesh (e.g., a 2D mesh). However, the sequences for LLM training typically vary in lengths, no matter for texts, multi-modalities or reinforcement learning. The mismatch between data heterogeneity and static mesh causes redundant communication and imbalanced computation, degrading the training efficiency. In this work, we introduce ByteScale, an efficient, flexible, and scalable LLM training framework for large-scale mixed training of long and short sequences. The core of ByteScale is a novel parallelism strategy, namely Hybrid Data Parallelism (HDP), which unifies the inter- and intra-data partitioning with a dynamic mesh design. In particular, we build a communication optimizer, which eliminates the redundant communication for short sequences by data-aware sharding and dynamic communication, and further compresses the communication cost for long sequences by selective offloading. Besides, we also develop a balance scheduler to mitigate the imbalanced computation by parallelism-aware data assignment. We evaluate ByteScale with the model sizes ranging from 7B to 141B, context lengths from 256K to 2048K, on a production cluster with more than 12,000 GPUs. Experiment results show that ByteScale outperforms the state-of-the-art training system by up to 7.89x.

The mixture of experts (MoE) model is a sparse variant of large language models (LLMs), designed to hold a better balance between intelligent capability and computational overhead. Despite its benefits, MoE is still too expensive to deploy on resource-constrained edge devices, especially with the demands of on-device inference services. Recent research efforts often apply model compression techniques, such as quantization, pruning and merging, to restrict MoE complexity. Unfortunately, due to their predefined static model optimization strategies, they cannot always achieve the desired quality-overhead trade-off when handling multiple requests, finally degrading the on-device quality of service. These limitations motivate us to propose the D^2MoE, an algorithm-system co-design framework that matches diverse task requirements by dynamically allocating the most proper bit-width to each expert. Specifically, inspired by the nested structure of matryoshka dolls, we propose the matryoshka weight quantization (MWQ) to progressively compress expert weights in a bit-nested manner and reduce the required runtime memory. On top of it, we further optimize the I/O-computation pipeline and design a heuristic scheduling algorithm following our hottest-expert-bit-first (HEBF) principle, which maximizes the expert parallelism between I/O and computation queue under constrained memory budgets, thus significantly reducing the idle temporal bubbles waiting for the experts to load. Evaluations on real edge devices show that D^2MoE improves the overall inference throughput by up to 1.39times and reduces the peak memory footprint by up to 53% over the latest on-device inference frameworks, while still preserving comparable serving accuracy as its INT8 counterparts.

Large language models (LLMs) have excelled in various applications, yet serving them at scale is challenging due to their substantial resource demands and high latency. Our real-world studies reveal that over 70% of user requests to LLMs have semantically similar counterparts, suggesting the potential for knowledge transfer among requests. However, naively caching and reusing past responses leads to a big quality drop. In this paper, we introduce IC-Cache, a caching system that enables live LLM capability augmentation to improve serving efficiency: by leveraging historical request-response pairs from larger models as in-context examples, IC-Cache empowers small LLMs to imitate and even exceed the compositional abilities (e.g., reasoning) of their larger counterparts, enabling selective offloading of requests to reduce cost and latency. Achieving this live augmentation at scale introduces intricate trade-offs between response quality, latency, and system throughput. For a new request, IC-Cache efficiently selects similar, high-utility examples to prepend them to the new request's input. At scale, it adaptively routes requests across LLMs of varying capabilities, accounting for response quality and serving loads. IC-Cache employs a cost-aware cache replay mechanism that refines example quality offline to maximize online cache utility and efficiency. Evaluations on millions of realistic requests demonstrate that IC-Cache improves LLM serving throughput by 1.4-5.9x and reduces latency by 28-71% without hurting response quality.

Large Language Model (LLM)-based multi-agent systems are increasingly powerful, but current agentic workflow optimization paradigms make an unsatisfying trade-off. Task-level methods spend substantial offline compute yet deploy only a single workflow, leaving complementary candidates unused, while query-level methods synthesize a new workflow per query at substantial inference cost. Our motivating analysis shows these paradigms are more complementary than competing: workflows discovered during offline search often solve different subsets of queries, and many queries handled by expensive query-level generation can already be solved by cheaper precomputed workflows. This suggests a different objective: rather than searching for one universally best workflow or regenerating one per instance, we should build a compact bank of reusable, complementary workflows and select among them adaptively at inference time. Doing so requires solving three coupled problems: generating complementary rather than redundant candidates, compressing them into a small deployable portfolio, and assigning each query to the right workflow under a performance-cost trade-off. To this end, we present FlowBank, a three-stage framework for portfolio-based agentic workflow optimization. Diversifying proposes DiverseFlow to steer search toward under-covered queries and produce a high-coverage candidate pool. Curating proposes CuraFlow to compress this pool into a compact portfolio with minimal redundancy. Matching casts deployment as edge-value prediction on a query-workflow bipartite graph and routes each incoming query to the portfolio member with the best predicted utility. Across five benchmarks, FlowBank achieves the highest average score among the evaluated methods while remaining cost-competitive, improving over the strongest automated and handcrafted baselines by 4.26% and 14.92% relative, respectively.

In this paper, we propose a novel learning paradigm, termed Chain-of-Model (CoM), which incorporates the causal relationship into the hidden states of each layer as a chain style, thereby introducing great scaling efficiency in model training and inference flexibility in deployment. We introduce the concept of Chain-of-Representation (CoR), which formulates the hidden states at each layer as a combination of multiple sub-representations (i.e., chains) at the hidden dimension level. In each layer, each chain from the output representations can only view all of its preceding chains in the input representations. Consequently, the model built upon CoM framework can progressively scale up the model size by increasing the chains based on the previous models (i.e., chains), and offer multiple sub-models at varying sizes for elastic inference by using different chain numbers. Based on this principle, we devise Chain-of-Language-Model (CoLM), which incorporates the idea of CoM into each layer of Transformer architecture. Based on CoLM, we further introduce CoLM-Air by introducing a KV sharing mechanism, that computes all keys and values within the first chain and then shares across all chains. This design demonstrates additional extensibility, such as enabling seamless LM switching, prefilling acceleration and so on. Experimental results demonstrate our CoLM family can achieve comparable performance to the standard Transformer, while simultaneously enabling greater flexiblity, such as progressive scaling to improve training efficiency and offer multiple varying model sizes for elastic inference, paving a a new way toward building language models. Our code will be released in the future at: https://github.com/microsoft/CoLM.

Mixture-of-Experts (MoE) models promise efficient scaling of large language models (LLMs) by activating only a small subset of experts per token, but their parallelized inference pipelines make elastic serving challenging. Existing strategies fall short: horizontal scaling provisions entire replicas of the current configuration, often tens to hundreds of accelerators, leading to coarse granularity, long provisioning delays, and costly overprovisioning. Vertical scaling offers finer adjustments but typically requires instance restarts, incurring downtime. These limitations make current approaches ill-suited for the bursty, short-lived traffic patterns common in cloud deployments. We present ElasticMoE, an elastic scaling framework for MoE LLMs that achieves fine-grained, low-latency, and zero-downtime scaling. ElasticMoE decouples inference execution from memory operations, enabling scaling steps to proceed concurrently with serving. An HBM Management Module (HMM) reuses weights and KV caches via zero-copy remapping, while high-bandwidth peer-to-peer transfers bring newly added accelerators online without interrupting service. A virtual memory based expert redistribution mechanism migrates MoE experts without costly buffer reallocations, reducing peak memory usage during expert parallelism reconfiguration. Our evaluation on Ascend NPUs with three popular MoE LLMs shows that ElasticMoE achieves up to 9x lower scale-up latency, up to 2x better throughput during scaling, and significantly improves SLO attainment compared to baselines. By enabling fine-grained, concurrent scaling with minimal disruption, ElasticMoE advances the practicality of deploying massive MoE LLMs in dynamic cloud environments.

The demand for deploying large language models(LLMs) on mobile devices continues to increase, driven by escalating data security concerns and cloud costs. However, network bandwidth and memory limitations pose challenges for deploying billion-level models on mobile devices. In this study, we investigate the outputs of different layers across various scales of LLMs and found that the outputs of most layers exhibit significant similarity. Moreover, this similarity becomes more pronounced as the model size increases, indicating substantial redundancy in the depth direction of the LLMs. Based on this observation, we propose an efficient model volume compression strategy, termed FoldGPT, which combines block removal and block parameter sharing.This strategy consists of three parts: (1) Based on the learnable gating parameters, we determine the block importance ranking while modeling the coupling effect between blocks. Then we delete some redundant layers based on the given removal rate. (2) For the retained blocks, we apply a specially designed group parameter sharing strategy, where blocks within the same group share identical weights, significantly compressing the number of parameters and slightly reducing latency overhead. (3) After sharing these Blocks, we "cure" the mismatch caused by sparsity with a minor amount of fine-tuning and introduce a tail-layer distillation strategy to improve the performance. Experiments demonstrate that FoldGPT outperforms previous state-of-the-art(SOTA) methods in efficient model compression, demonstrating the feasibility of achieving model lightweighting through straightforward block removal and parameter sharing.

The emergence of LLMs has ignited a fresh surge of breakthroughs in NLP applications, particularly in domains such as question-answering systems and text generation. As the need for longer context grows, a significant bottleneck in model deployment emerges due to the linear expansion of the Key-Value (KV) cache with the context length. Existing methods primarily rely on various hypotheses, such as sorting the KV cache based on attention scores for replacement or eviction, to compress the KV cache and improve model throughput. However, heuristics used by these strategies may wrongly evict essential KV cache, which can significantly degrade model performance. In this paper, we propose QAQ, a Quality Adaptive Quantization scheme for the KV cache. We theoretically demonstrate that key cache and value cache exhibit distinct sensitivities to quantization, leading to the formulation of separate quantization strategies for their non-uniform quantization. Through the integration of dedicated outlier handling, as well as an improved attention-aware approach, QAQ achieves up to 10x the compression ratio of the KV cache size with a neglectable impact on model performance. QAQ significantly reduces the practical hurdles of deploying LLMs, opening up new possibilities for longer-context applications. The code is available at github.com/ClubieDong/KVCacheQuantization.

Speculative decoding can significantly accelerate LLM serving, yet most deployments today disentangle speculator training from serving, treating speculator training as a standalone offline modeling problem. We show that this decoupled formulation introduces substantial deployment and adaptation lag: (1) high time-to-serve, since a speculator must be trained offline for a considerable period before deployment; (2) delayed utility feedback, since the true end-to-end decoding speedup is only known after training and cannot be inferred reliably from acceptance rate alone due to model-architecture and system-level overheads; and (3) domain-drift degradation, as the target model is repurposed to new domains and the speculator becomes stale and less effective. To address these issues, we present Aurora, a unified training-serving system that closes the loop by continuously learning a speculator directly from live inference traces. Aurora reframes online speculator learning as an asynchronous reinforcement-learning problem: accepted tokens provide positive feedback, while rejected speculator proposals provide implicit negative feedback that we exploit to improve sample efficiency. Our design integrates an SGLang-based inference server with an asynchronous training server, enabling hot-swapped speculator updates without service interruption. Crucially, Aurora supports day-0 deployment: a speculator can be served immediately and rapidly adapted to live traffic, improving system performance while providing immediate utility feedback. Across experiments, Aurora achieves a 1.5x day-0 speedup on recently released frontier models (e.g., MiniMax M2.1 229B and Qwen3-Coder-Next 80B). Aurora also adapts effectively to distribution shifts in user traffic, delivering an additional 1.25x speedup over a well-trained but static speculator on widely used models (e.g., Qwen3 and Llama3).

Large Language Models (LLMs) have demonstrated impressive quality when applied to predictive tasks such as relevance ranking and semantic search. However, deployment of such LLMs remains prohibitively expensive for industry applications with strict latency and throughput requirements. In this work, we present lessons and efficiency insights from developing a purely text-based decoder-only Small Language Model (SLM) for a semantic search application at LinkedIn. Particularly, we discuss model compression techniques such as pruning that allow us to reduce the model size by up to 40% while maintaining the accuracy. Additionally, we present context compression techniques that allow us to reduce the input context length by up to 10x with minimal loss of accuracy. Finally, we present practical lessons from optimizing the serving infrastructure for deploying such a system on GPUs at scale, serving millions of requests per second. Taken together, this allows us to increase our system's throughput by 10x in a real-world deployment, while meeting our quality bar.

Large Language Models (LLMs) have revolutionized natural language processing tasks with remarkable success. However, their formidable size and computational demands present significant challenges for practical deployment, especially in resource-constrained environments. As these challenges become increasingly pertinent, the field of model compression has emerged as a pivotal research area to alleviate these limitations. This paper presents a comprehensive survey that navigates the landscape of model compression techniques tailored specifically for LLMs. Addressing the imperative need for efficient deployment, we delve into various methodologies, encompassing quantization, pruning, knowledge distillation, and more. Within each of these techniques, we highlight recent advancements and innovative approaches that contribute to the evolving landscape of LLM research. Furthermore, we explore benchmarking strategies and evaluation metrics that are essential for assessing the effectiveness of compressed LLMs. By providing insights into the latest developments and practical implications, this survey serves as an invaluable resource for both researchers and practitioners. As LLMs continue to evolve, this survey aims to facilitate enhanced efficiency and real-world applicability, establishing a foundation for future advancements in the field.

Large multimodal models (LMMs) demonstrate impressive capabilities in understanding images, videos, and audio beyond text. However, efficiently serving LMMs in production environments poses significant challenges due to their complex architectures and heterogeneous characteristics across their multi-stage inference pipelines. We present the first comprehensive systems analysis of two prominent LMM architectures, decoder-only and cross-attention, across six representative open-source models, revealing key systems design implications. We also present an in-depth analysis of production LMM inference traces, uncovering unique workload characteristics, including variable, heavy-tailed request distributions and bursty traffic patterns. Based on these insights, we propose ModServe, a modular LMM serving system that decouples stages for independent optimization and adaptive scaling. ModServe dynamically reconfigures stages and handles bursty traffic with modality-aware scheduling and autoscaling to meet tail latency SLOs while minimizing costs. ModServe achieves 3.3-5.5x higher throughput (leading to 25-41.3% cost saving) while meeting SLOs on a 128-GPU cluster with production traces.

Large models training is plagued by the intense compute cost and limited hardware memory. A practical solution is low-precision representation but is troubled by loss in numerical accuracy and unstable training rendering the model less useful. We argue that low-precision floating points can perform well provided the error is properly compensated at the critical locations in the training process. We propose Collage which utilizes multi-component float representation in low-precision to accurately perform operations with numerical errors accounted. To understand the impact of imprecision to training, we propose a simple and novel metric which tracks the lost information during training as well as differentiates various precision strategies. Our method works with commonly used low-precision such as half-precision (16-bit floating points) and can be naturally extended to work with even lower precision such as 8-bit. Experimental results show that pre-training using Collage removes the requirement of using 32-bit floating-point copies of the model and attains similar/better training performance compared to (16, 32)-bit mixed-precision strategy, with up to 3.7times speedup and sim 15% to 23% less memory usage in practice.

With the increasing demand for deep learning models on mobile devices, splitting neural network computation between the device and a more powerful edge server has become an attractive solution. However, existing split computing approaches often underperform compared to a naive baseline of remote computation on compressed data. Recent studies propose learning compressed representations that contain more relevant information for supervised downstream tasks, showing improved tradeoffs between compressed data size and supervised performance. However, existing evaluation metrics only provide an incomplete picture of split computing. This study introduces supervised compression for split computing (SC2) and proposes new evaluation criteria: minimizing computation on the mobile device, minimizing transmitted data size, and maximizing model accuracy. We conduct a comprehensive benchmark study using 10 baseline methods, three computer vision tasks, and over 180 trained models, and discuss various aspects of SC2. We also release sc2bench, a Python package for future research on SC2. Our proposed metrics and package will help researchers better understand the tradeoffs of supervised compression in split computing.

Any-to-any multimodal models that jointly handle text, images, video, and audio represent a significant advance in multimodal AI. However, their complex architectures (typically combining multiple autoregressive LLMs, diffusion transformers, and other specialized components) pose substantial challenges for efficient model serving. Existing serving systems are mainly tailored to a single paradigm, such as autoregressive LLMs for text generation or diffusion transformers for visual generation. They lack support for any-to-any pipelines that involve multiple interconnected model components. As a result, developers must manually handle cross-stage interactions, leading to huge performance degradation. We present vLLM-Omni, a fully disaggregated serving system for any-to-any models. vLLM-Omni features a novel stage abstraction that enables users to decompose complex any-to-any architectures into interconnected stages represented as a graph, and a disaggregated stage execution backend that optimizes resource utilization and throughput across stages. Each stage is independently served by an LLM or diffusion engine with per-stage request batching, flexible GPU allocation, and unified inter-stage connectors for data routing. Experimental results demonstrate that vLLM-Omni reduces job completion time (JCT) by up to 91.4% compared to baseline methods. The code is public available at https://github.com/vllm-project/vllm-omni.

Multimodal foundation models serve numerous applications at the intersection of vision and language. Still, despite being pretrained on extensive data, they become outdated over time. To keep models updated, research into continual pretraining mainly explores scenarios with either (1) infrequent, indiscriminate updates on large-scale new data, or (2) frequent, sample-level updates. However, practical model deployment often operates in the gap between these two limit cases, as real-world applications often demand adaptation to specific subdomains, tasks or concepts -- spread over the entire, varying life cycle of a model. In this work, we complement current perspectives on continual pretraining through a research test bed as well as provide comprehensive guidance for effective continual model updates in such scenarios. We first introduce FoMo-in-Flux, a continual multimodal pretraining benchmark with realistic compute constraints and practical deployment requirements, constructed over 63 datasets with diverse visual and semantic coverage. Using FoMo-in-Flux, we explore the complex landscape of practical continual pretraining through multiple perspectives: (1) A data-centric investigation of data mixtures and stream orderings that emulate real-world deployment situations, (2) a method-centric investigation ranging from simple fine-tuning and traditional continual learning strategies to parameter-efficient updates and model merging, (3) meta learning rate schedules and mechanistic design choices, and (4) the influence of model and compute scaling. Together, our insights provide a practitioner's guide to continual multimodal pretraining for real-world deployment. Our benchmark and code is here: https://github.com/ExplainableML/fomo_in_flux.

Federated learning is a distributed learning paradigm in which multiple mobile clients train a global model while keeping data local. These mobile clients can have various available memory and network bandwidth. However, to achieve the best global model performance, how we can utilize available memory and network bandwidth to the maximum remains an open challenge. In this paper, we propose to assign each client a subset of the global model, having different layers and channels on each layer. To realize that, we design a constrained model search process with early stop to improve efficiency of finding the models from such a very large space; and a data-free knowledge distillation mechanism to improve the global model performance when aggregating models of such different structures. For fair and reproducible comparison between different solutions, we develop a new system, which can directly allocate different memory and bandwidth to each client according to memory and bandwidth logs collected on mobile devices. The evaluation shows that our solution can have accuracy increase ranging from 2.43\% to 15.81\% and provide 5\% to 40\% more memory and bandwidth utilization with negligible extra running time, comparing to existing state-of-the-art system-heterogeneous federated learning methods under different available memory and bandwidth, non-i.i.d.~datasets, image and text tasks.

Compared to traditional machine learning models, recent large language models (LLMs) can exhibit multi-task-solving capabilities through multiple dialogues and multi-modal data sources. These unique characteristics of LLMs, together with their large model size, make their deployment more challenging. Specifically, (i) deploying LLMs on local devices faces computational, memory, and energy resource issues, while (ii) deploying them in the cloud cannot guarantee real-time service and incurs communication/usage costs. In this paper, we design TMO, a local-cloud LLM inference system with Three-M Offloading: Multi-modal, Multi-task, and Multi-dialogue. TMO incorporates (i) a lightweight local LLM that can process simple tasks at high speed and (ii) a large-scale cloud LLM that can handle multi-modal data sources. We develop a resource-constrained reinforcement learning (RCRL) strategy for TMO that optimizes the inference location (i.e., local vs. cloud) and multi-modal data sources to use for each task/dialogue, aiming to maximize the long-term reward (response quality, latency, and usage cost) while adhering to resource constraints. We also contribute M4A1, a new dataset we curated that contains reward and cost metrics across multiple modality, task, dialogue, and LLM configurations, enabling evaluation of offloading decisions. We demonstrate the effectiveness of TMO compared to several exploration-decision and LLM-as-Agent baselines, showing significant improvements in latency, cost, and response quality.

Offline batch inference, which leverages the flexibility of request batching to achieve higher throughput and lower costs, is becoming more popular for latency-insensitive applications. Meanwhile, recent progress in model capability and modality makes requests more diverse in compute and memory demands, creating unique opportunities for throughput improvement by resource overlapping. However, a request schedule that maximizes resource overlapping can conflict with the schedule that maximizes prefix sharing, a widely-used performance optimization, causing sub-optimal inference throughput. We present BlendServe, a system that maximizes resource utilization of offline batch inference by combining the benefits of resource overlapping and prefix sharing using a resource-aware prefix tree. BlendServe exploits the relaxed latency requirements in offline batch inference to reorder and overlap requests with varied resource demands while ensuring high prefix sharing. We evaluate BlendServe on a variety of synthetic multi-modal workloads and show that it provides up to 1.44times throughput boost compared to widely-used industry standards, vLLM and SGLang.

KV-cache memory is a major bottleneck in real-world LLM serving, where systems must simultaneously support latency-sensitive small-batch requests and high-throughput concurrent workloads. Although many KV-cache compression methods improve offline accuracy or compression ratio, they often violate practical serving constraints such as paged memory layouts, regular memory access, and fused attention execution, limiting their effectiveness in deployment. In this work, we identify the minimal set of 4-bit KV-cache quantization methods that remain viable under these constraints. Our central finding is that a simple design--token-wise INT4 quantization with block-diagonal Hadamard rotation--consistently achieves the best accuracy-efficiency trade-off. Across multiple models and benchmarks, this approach recovers nearly all of the accuracy lost by naive INT4, while more complex methods such as vector quantization and Hessian-aware quantization provide only marginal additional gains once serving compatibility is taken into account. To make this practical, we implement a fused rotation-quantization kernel that integrates directly into paged KV-cache layouts and introduces zero measurable end-to-end overhead, matching plain INT4 throughput across concurrency levels. Our results show that effective KV-cache compression is fundamentally a systems co-design problem: under real serving constraints, lightweight block-diagonal Hadamard rotation is a viable method that delivers near-lossless accuracy without sacrificing serving efficiency.

Recent advancements in large language models (LLMs) and their multimodal variants have led to remarkable progress across various domains, demonstrating impressive capabilities and unprecedented potential. In the era of ubiquitous connectivity, leveraging communication networks to distribute intelligence is a transformative concept, envisioning AI-powered services accessible at the network edge. However, pushing large models from the cloud to resource-constrained environments faces critical challenges. Model inference on low-end devices leads to excessive latency and performance bottlenecks, while raw data transmission over limited bandwidth networks causes high communication overhead. This article presents AI Flow, a framework that streamlines the inference process by jointly leveraging the heterogeneous resources available across devices, edge nodes, and cloud servers, making intelligence flow across networks. To facilitate cooperation among multiple computational nodes, the proposed framework explores a paradigm shift in the design of communication network systems from transmitting information flow to intelligence flow, where the goal of communications is task-oriented and folded into the inference process. Experimental results demonstrate the effectiveness of the proposed framework through an image captioning use case, showcasing the ability to reduce response latency while maintaining high-quality captions. This article serves as a position paper for identifying the motivation, challenges, and principles of AI Flow.

Many NLP tasks benefit from using large language models (LLMs) that often have more than 100 billion parameters. With the release of BLOOM-176B and OPT-175B, everyone can download pretrained models of this scale. Still, using these models requires high-end hardware unavailable to many researchers. In some cases, LLMs can be used more affordably via RAM offloading or hosted APIs. However, these techniques have innate limitations: offloading is too slow for interactive inference, while APIs are not flexible enough for research that requires access to weights, attention or logits. In this work, we propose Petals - a system for inference and fine-tuning of large models collaboratively by joining the resources of multiple parties. We demonstrate that this strategy outperforms offloading for very large models, running inference of BLOOM-176B on consumer GPUs with approx 1 step per second, which is enough for many interactive LLM applications. Unlike most inference APIs, Petals also natively exposes hidden states of served models, allowing to train and share custom model extensions based on efficient fine-tuning methods.

We survey the large language model (LLM) serving area to understand the intricate dynamics between cost-efficiency and accuracy, which is magnified by the growing need for longer contextual understanding when deploying models at a massive scale. Our findings reveal that works in this space optimize along three distinct but conflicting goals: improving serving context length (C), improving serving accuracy (A), and improving serving performance (P). Drawing inspiration from the CAP theorem in databases, we propose a CAP principle for LLM serving, which suggests that any optimization can improve at most two of these three goals simultaneously. Our survey categorizes existing works within this framework. We find the definition and continuity of user-perceived measurement metrics are crucial in determining whether a goal has been met, akin to prior CAP databases in the wild. We recognize the CAP principle for LLM serving as a guiding principle, rather than a formal theorem, to inform designers of the inherent and dynamic trade-offs in serving models. As serving accuracy and performance have been extensively studied, this survey focuses on works that extend serving context length and address the resulting challenges.

Model compression is a critical technique to efficiently deploy neural network models on mobile devices which have limited computation resources and tight power budgets. Conventional model compression techniques rely on hand-crafted heuristics and rule-based policies that require domain experts to explore the large design space trading off among model size, speed, and accuracy, which is usually sub-optimal and time-consuming. In this paper, we propose AutoML for Model Compression (AMC) which leverage reinforcement learning to provide the model compression policy. This learning-based compression policy outperforms conventional rule-based compression policy by having higher compression ratio, better preserving the accuracy and freeing human labor. Under 4x FLOPs reduction, we achieved 2.7% better accuracy than the handcrafted model compression policy for VGG-16 on ImageNet. We applied this automated, push-the-button compression pipeline to MobileNet and achieved 1.81x speedup of measured inference latency on an Android phone and 1.43x speedup on the Titan XP GPU, with only 0.1% loss of ImageNet Top-1 accuracy.

Over the past year, the vLLM Semantic Router project has released a series of work spanning: (1) core routing mechanisms -- signal-driven routing, context-length pool routing, router performance engineering, policy conflict detection, low-latency embedding models, category-aware semantic caching, user-feedback-driven routing adaptation, hallucination detection, and hierarchical content-safety classification for privacy and jailbreak protection; (2) fleet optimization -- fleet provisioning and energy-efficiency analysis; (3) agentic and multimodal routing -- multimodal agent routing, tool selection, CUA security, and multi-turn context memory and safety; (4) governance and standards -- inference routing protocols and multi-provider API extensions. Each paper tackled a specific problem in LLM inference, but the problems are not independent; for example, fleet provisioning depends on the routing policy, which depends on the workload mix, shifting as organizations adopt agentic and multimodal workloads. This paper distills those results into the Workload-Router-Pool (WRP) architecture, a three-dimensional framework for LLM inference optimization. Workload characterizes what the fleet serves (chat vs. agent, single-turn vs. multi-turn, warm vs. cold, prefill-heavy vs. decode-heavy). Router determines how each request is dispatched (static semantic rules, online bandit adaptation, RL-based model selection, quality-aware cascading). Pool defines where inference runs (homogeneous vs. heterogeneous GPU, disaggregated prefill/decode, KV-cache topology). We map our prior work onto a 3x3 WRP interaction matrix, identify which cells we have covered and which remain open, and propose twenty-one concrete research directions at the intersections, each grounded in our prior measurements, tiered by maturity from engineering-ready to open research.

The adoption of Foundation Models in resource-constrained environments remains challenging due to their large size and inference costs. A promising way to overcome these limitations is post-training compression, which aims to balance reduced model size against performance degradation. This work presents Any Compression via Iterative Pruning (ACIP), a novel algorithmic approach to determine a compression-performance trade-off from a single stochastic gradient descent run. To ensure parameter efficiency, we use an SVD-reparametrization of linear layers and iteratively prune their singular values with a sparsity-inducing penalty. The resulting pruning order gives rise to a global parameter ranking that allows us to materialize models of any target size. Importantly, the compressed models exhibit strong predictive downstream performance without the need for costly fine-tuning. We evaluate ACIP on a large selection of open-weight LLMs and tasks, and demonstrate state-of-the-art results compared to existing factorisation-based compression methods. We also show that ACIP seamlessly complements common quantization-based compression techniques.

Large Language Model (LLM) inference is increasingly constrained by GPU memory capacity rather than compute throughput, driven by growing model sizes and the linear growth of the key-value (KV) cache during autoregressive decoding. Existing approaches mitigate memory pressure by offloading model state and KV tensors to host memory, but incur substantial latency due to limited PCIe bandwidth. We present Harvest, an opportunistic GPU cache management framework that exploits high-bandwidth peer-to-peer GPU interconnects to dynamically place model weights and KV cache in unused GPU memory. Harvest treats peer GPU memory as a transient cache tier, preserving correctness while reducing data movement overhead under dynamic memory availability. We demonstrate significant throughput speedup of more than 2 times by using Harvest to accelerate the retrieval of two widely-used inference components: expert layer weights and KV cache entries.

Serving many task-specialized LLM variants is often limited by the large size of fine-tuned checkpoints and the resulting cold-start latency. Since fine-tuned weights differ from their base model by relatively small structured residuals, a natural approach is to represent them as compressed deltas. We propose a simple 1-bit delta scheme that stores only the sign of the weight difference together with lightweight per-axis (row/column) FP16 scaling factors, learned from a small calibration set. This design preserves the compactness of 1-bit deltas while more accurately capturing variation across weight dimensions, leading to improved reconstruction quality over scalar alternatives. From a systems perspective, a streamlined loader that transfers packed deltas in a single operation per module reduces cold-start latency and storage overhead, with artifacts several times smaller than a full FP16 checkpoint. The method is drop-in, requires minimal calibration data, and maintains inference efficiency by avoiding dense reconstruction. Our experimental setup and source code are available at https://github.com/kuiumdjiev/Per-Axis-Weight-Deltas-for-Frequent-Model-Updates.

Expert-Specialized Fine-Tuning (ESFT) adapts Mixture-of-Experts (MoE) large language models to enhance their task-specific performance by selectively tuning the top-activated experts for the task. Serving these fine-tuned models at scale is challenging: deploying merged models in isolation is prohibitively resource-hungry, while existing multi-adapter serving systems with LoRA-style additive updates are incompatible with ESFT's expert-oriented paradigm. We present ExpertWeave, a system that serves multiple ESFT adapters concurrently over a single shared MoE base model, drastically reducing the memory footprint and improving resource utilization. To seamlessly integrate into existing inference pipelines for MoE models with non-intrusive modifications and minimal latency overhead, ExpertWeave introduces a virtual-memory-assisted expert weight manager that co-locates base-model and adapter experts without incurring memory overhead from fragmentation, and a fused kernel for batched rerouting to enable lightweight redirection of tokens to the appropriate experts at runtime. Our evaluations show that ExpertWeave can simultaneously serve multiple adapters of a 16B MoE model on a single accelerator where the baseline runs out of memory, or provides up to 94x more KV cache capacity and achieves up to 18% higher throughput while using comparable resources, all without compromising model accuracy. ExpertWeave maintains low overhead even when scaling to 20 adapters, with a 4-11% latency increase compared with serving the base model alone. Source code will be released soon.

Transfer learning is important for foundation models to adapt to downstream tasks. However, many foundation models are proprietary, so users must share their data with model owners to fine-tune the models, which is costly and raise privacy concerns. Moreover, fine-tuning large foundation models is computation-intensive and impractical for most downstream users. In this paper, we propose Offsite-Tuning, a privacy-preserving and efficient transfer learning framework that can adapt billion-parameter foundation models to downstream data without access to the full model. In offsite-tuning, the model owner sends a light-weight adapter and a lossy compressed emulator to the data owner, who then fine-tunes the adapter on the downstream data with the emulator's assistance. The fine-tuned adapter is then returned to the model owner, who plugs it into the full model to create an adapted foundation model. Offsite-tuning preserves both parties' privacy and is computationally more efficient than the existing fine-tuning methods that require access to the full model weights. We demonstrate the effectiveness of offsite-tuning on various large language and vision foundation models. Offsite-tuning can achieve comparable accuracy as full model fine-tuning while being privacy-preserving and efficient, achieving 6.5x speedup and 5.6x memory reduction. Code is available at https://github.com/mit-han-lab/offsite-tuning.

Transformer-based large language models (LLMs) exhibit impressive performance in generative tasks but introduce significant challenges in real-world serving due to inefficient use of the expensive, computation-optimized accelerators. This mismatch arises from the autoregressive nature of LLMs, where the generation phase comprises operators with varying resource demands. Specifically, the attention operator is memory-intensive, exhibiting a memory access pattern that clashes with the strengths of modern accelerators, especially as context length increases. To enhance the efficiency and cost-effectiveness of LLM serving, we introduce the concept of attention offloading. This approach leverages a collection of cheap, memory-optimized devices for the attention operator while still utilizing high-end accelerators for other parts of the model. This heterogeneous setup ensures that each component is tailored to its specific workload, maximizing overall performance and cost efficiency. Our comprehensive analysis and experiments confirm the viability of splitting the attention computation over multiple devices. Also, the communication bandwidth required between heterogeneous devices proves to be manageable with prevalent networking technologies. To further validate our theory, we develop Lamina, an LLM inference system that incorporates attention offloading. Experimental results indicate that Lamina can provide 1.48x-12.1x higher estimated throughput per dollar than homogeneous solutions.

Fine-tuning large pre-trained LLMs generally demands extensive GPU memory. Traditional first-order optimizers like SGD encounter substantial difficulties due to increased memory requirements from storing activations and gradients during both the forward and backward phases as the model size expands. Alternatively, zeroth-order (ZO) techniques can compute gradients using just forward operations, eliminating the need to store activations. Furthermore, by leveraging CPU capabilities, it's feasible to enhance both the memory and processing power available to a single GPU. We propose a novel framework, ZO2 (Zeroth-Order Offloading), for efficient zeroth-order fine-tuning of LLMs with only limited GPU memory. Our framework dynamically shifts model parameters between the CPU and GPU as required, optimizing computation flow and maximizing GPU usage by minimizing downtime. This integration of parameter adjustments with ZO's double forward operations reduces unnecessary data movement, enhancing the fine-tuning efficacy. Additionally, our framework supports an innovative low-bit precision approach in AMP mode to streamline data exchanges between the CPU and GPU. Employing this approach allows us to fine-tune extraordinarily large models, such as the OPT-175B with more than 175 billion parameters, on a mere 18GB GPU--achievements beyond the reach of traditional methods. Moreover, our framework achieves these results with almost no additional time overhead and absolutely no accuracy loss compared to standard zeroth-order methods. ZO2's code has been open-sourced in https://github.com/liangyuwang/zo2.

Recent progress in Language Models (LMs) has dramatically advanced the field of natural language processing (NLP), excelling at tasks like text generation, summarization, and question answering. However, their inference remains computationally expensive and energy intensive, especially in settings with limited hardware, power, or bandwidth. This makes it difficult to deploy LMs in mobile, edge, or cost sensitive environments. To address these challenges, recent approaches have introduced multi LLM intelligent model selection strategies that dynamically allocate computational resources based on query complexity -- using lightweight models for simpler queries and escalating to larger models only when necessary. This survey explores two complementary strategies for efficient LLM inference: (i) routing, which selects the most suitable model based on the query, and (ii) cascading or hierarchical inference (HI), which escalates queries through a sequence of models until a confident response is found. Both approaches aim to reduce computation by using lightweight models for simpler tasks while offloading only when needed. We provide a comparative analysis of these techniques across key performance metrics, discuss benchmarking efforts, and outline open challenges. Finally, we outline future research directions to enable faster response times, adaptive model selection based on task complexity, and scalable deployment across heterogeneous environments, making LLM based systems more efficient and accessible for real world applications.

The open-source model ecosystem now contains hundreds of thousands of pretrained models, yet picking the best model for a new dataset is increasingly infeasible: new models and unbenchmarked datasets emerge continuously, leaving practitioners with no prior records on either side. Existing approaches handle only fragments of this in-the-wild setting: AutoML and transferability estimation select models from small predefined pools or require expensive per-model forward passes on the target dataset, while model routing presupposes a given candidate pool. We introduce ModelLens, a unified framework for model recommendation in the wild. Our key insight is that public leaderboard interactions, though scattered and noisy, collectively trace out an implicit atlas of model capabilities across heterogeneous evaluation settings, a signal rich enough to learn from directly. By learning a performance-aware latent space over model--dataset--metric tuples, ModelLens ranks unseen models on unseen datasets without running candidates on the target dataset. On a new benchmark of 1.62M evaluation records spanning 47K models and 9.6K datasets, ModelLens surpasses baselines that either rely on metadata alone or require running each candidate on the target dataset. Its recommended Top-K pools further improve multiple representative routing methods by up to 81% across diverse QA benchmarks. Case studies on recently released benchmarks further confirm generalization to both text and vision-language tasks.

Recent large language model applications, such as Retrieval-Augmented Generation and chatbots, have led to an increased need to process longer input contexts. However, this requirement is hampered by inherent limitations. Architecturally, models are constrained by a context window defined during training. Additionally, processing extensive texts requires substantial GPU memory. We propose a novel approach, Finch, to compress the input context by leveraging the pre-trained model weights of the self-attention. Given a prompt and a long text, Finch iteratively identifies the most relevant Key (K) and Value (V) pairs over chunks of the text conditioned on the prompt. Only such pairs are stored in the KV cache, which, within the space constrained by the context window, ultimately contains a compressed version of the long text. Our proposal enables models to consume large inputs even with high compression (up to 93x) while preserving semantic integrity without the need for fine-tuning.

Production LLM deployments increasingly maintain heterogeneous model pools spanning order-of-magnitude cost differences. Existing routers make binary strong-vs-weak decisions and couple learned parameters to specific model identities, requiring retraining whenever the catalog changes. We present HyDRA (Hybrid Dynamic Routing Architecture), a framework that predicts fine-grained, multi-dimensional capability requirements per query and matches them against configuration-defined model profiles via shortfall matching. A ModernBERT encoder with K=4 independent sigmoid heads scores each query along reasoning, code generation, debugging, and tool use; a shortfall-matching algorithm then selects the cheapest model whose capabilities meet the predicted requirements. The deployed predictor runs at 86 ms median CPU inference latency in production, and is fully decoupled from the model catalog -- adding or removing models requires only a configuration change, with zero retraining. On SWE-Bench Verified (5-model pool: GPT-5.4-mini, Claude Haiku 4.5, GPT-5.3 Codex, Claude Sonnet 4.6, GPT-5.4), HyDRA's tunable shortfall threshold spans three regimes: peak-quality exceeds the always-strong Claude Sonnet 4.6 baseline (75.4% vs. 74.2% resolution) at 12.9% cost savings; iso-quality matches Sonnet at 54.1% cost savings, a 6x improvement over our prior in-house binary router at 9.1%; aggressive pushes savings to 72.5% for a 3.2-point quality trade. Results generalize across LiveCodeBench, BigCodeBench, and tau-bench. HyDRA is deployed to all users in GitHub Copilot's VS Code Chat auto-mode and -- to our knowledge for the first time in the LLM routing literature -- demonstrates language-invariant routing across CJK, European, and other script families.

Model pruning seeks to induce sparsity in a deep neural network's various connection matrices, thereby reducing the number of nonzero-valued parameters in the model. Recent reports (Han et al., 2015; Narang et al., 2017) prune deep networks at the cost of only a marginal loss in accuracy and achieve a sizable reduction in model size. This hints at the possibility that the baseline models in these experiments are perhaps severely over-parameterized at the outset and a viable alternative for model compression might be to simply reduce the number of hidden units while maintaining the model's dense connection structure, exposing a similar trade-off in model size and accuracy. We investigate these two distinct paths for model compression within the context of energy-efficient inference in resource-constrained environments and propose a new gradual pruning technique that is simple and straightforward to apply across a variety of models/datasets with minimal tuning and can be seamlessly incorporated within the training process. We compare the accuracy of large, but pruned models (large-sparse) and their smaller, but dense (small-dense) counterparts with identical memory footprint. Across a broad range of neural network architectures (deep CNNs, stacked LSTM, and seq2seq LSTM models), we find large-sparse models to consistently outperform small-dense models and achieve up to 10x reduction in number of non-zero parameters with minimal loss in accuracy.

Large Language Models (LLMs) with Mixture of Experts (MoE) architectures achieve remarkable performance improvements, but their random expert selection mechanism introduces significant data movement overhead that becomes the dominant bottleneck in multi-unit serving systems. To forecast the patterns underlying this data movement, we conduct comprehensive data-movement-centric profiling across three state-of-the-art large-scale MoE models (200B- 671B) using over 24,000 requests spanning diverse workloads. With the resulting 150GB+ trace files, we perform systematic analysis from both temporal and spatial perspectives and distill six key insights to guide the design of diverse future serving systems. Taking wafer-scale GPUs as a case study, we demonstrate that minor architectural modifications leveraging our insights achieve substantial performance gains, delivering 6.3X and 4.0X average speedups on DeepSeek V3 and Qwen3, respectively. Our work provides the first comprehensive data-centric analysis of MoE models at scale. Our profiling traces and analysis results are publicly available at {https://huggingface.co/datasets/core12345/MoE_expert_selection_trace. We will also release our simulation framework shortly to facilitate future research in this area.

Post-training model compression is essential for enhancing the portability of Large Language Models (LLMs) while preserving their performance. While several compression approaches have been proposed, less emphasis has been placed on selecting the most suitable set of data (the so-called calibration data) for finding the compressed model configuration. The choice of calibration data is a critical step in preserving model capabilities both intra- and inter-tasks. In this work, we address the challenge of identifying high-performance calibration sets for both pruning and quantization by analyzing intrinsic data properties rather than model-specific signals. We introduce \textbf{ZipCal}, a model-agnostic data curation strategy that maximizes lexical diversity based on Zipfian power laws. Experiments demonstrate that our method consistently outperforms standard uniform random sampling across various pruning benchmarks. Notably, it also performs on par, in terms of downstream performance, with a state-of-the-art method that relies on model perplexity. The latter becomes prohibitively expensive at large-scale models and datasets, while \textbf{ZipCal} is on average sim240times faster due to its tractable linear complexityWe make the code and the experiments available at https://anonymous.4open.science/r/zipcal-71CD/..

Fine-tuning large language models (LLMs) remains resource-intensive due to their sheer scale. While zeroth-order (ZO) optimization provides a memory-efficient alternative by eliminating backward passes, its application to multi-hundred-billion-parameter models is constrained by GPU memory and compute throughput. The ZO2 framework addresses the memory bottleneck by offloading model parameters to CPU memory and overlapping transformer block transfer with dual forward computation on a single GPU. However, ZO2 remains limited by its single-device execution and achieves modest throughput. In this work, we present DistZO2, a high-throughput, memory-efficient framework for distributed zeroth-order fine-tuning of LLMs. DistZO2 introduces three parallel strategies: (1) Perturbation Parallelism (PertP), which parallelizes the two perturbed forward passes across devices; (2) Distributed Data Parallelism (DDP), adapted to the scalar-gradient nature of ZO training; and (3) a unified 2D Parallelism design that combines PertP and DDP. To further mitigate communication bottlenecks introduced by parameter offloading, we propose a hardware-aware communication strategy that slices parameter blocks and redistributes them across GPUs via high-speed interconnects such as NVLink. DistZO2 scales zeroth-order fine-tuning to modern multi-GPU systems, preserving ZO2's memory efficiency while substantially improving training throughput. In our experiments on OPT-175B, DistZO2 achieves a 3x speedup over ZO2 with distributed computing. DistZO2's code has been open-sourced in https://github.com/liangyuwang/zo2.

Federated Learning (FL) facilitates collaborative training of a shared global model without exposing clients' private data. In practical FL systems, clients (e.g., edge servers, smartphones, and wearables) typically have disparate system resources. Conventional FL, however, adopts a one-size-fits-all solution, where a homogeneous large global model is transmitted to and trained on each client, resulting in an overwhelming workload for less capable clients and starvation for other clients. To address this issue, we propose FedConv, a client-friendly FL framework, which minimizes the computation and memory burden on resource-constrained clients by providing heterogeneous customized sub-models. FedConv features a novel learning-on-model paradigm that learns the parameters of the heterogeneous sub-models via convolutional compression. Unlike traditional compression methods, the compressed models in FedConv can be directly trained on clients without decompression. To aggregate the heterogeneous sub-models, we propose transposed convolutional dilation to convert them back to large models with a unified size while retaining personalized information from clients. The compression and dilation processes, transparent to clients, are optimized on the server leveraging a small public dataset. Extensive experiments on six datasets demonstrate that FedConv outperforms state-of-the-art FL systems in terms of model accuracy (by more than 35% on average), computation and communication overhead (with 33% and 25% reduction, respectively).

Two widely adopted techniques for LLM inference serving systems today are hybrid batching and disaggregated serving. A hybrid batch combines prefill and decode tokens of different requests in the same batch to improve resource utilization and throughput at the cost of increased latency per token. In contrast, disaggregated serving decouples compute-bound prefill and bandwidth-bound decode phases to optimize for service level objectives (SLOs) at the cost of resource under-utilization and KV-cache transfer overheads. To address the limitations of these techniques, we propose RAPID-Serve: a technique to concurrently execute prefill and decode on the same GPU(s) to meet latency SLOs while maintaining high throughput and efficient resource utilization. Furthermore, we propose Adaptive Resource Management for runtime compute resource allocation, optionally leveraging CU masking (a fine-grained Compute Unit partitioning feature on AMD InstinctTM GPUs). RAPID-Serve provides up to 4.1x (average 1.7x) unconstrained throughput improvement and 32x and higher (average 4.9x) throughput improvement under SLO constraints, showing it as an effective strategy compared to the state-of-the-art approaches, particularly in resource-constrained environments.

The rapid adoption of large language models (LLMs) has created significant challenges for efficient inference at scale. Unlike traditional workloads, LLM inference is constrained by both computation and the memory overhead of key-value (KV) caching, which accelerates decoding but quickly exhausts GPU memory. In this paper, we introduce the first queueing-theoretic framework that explicitly incorporates both computation and GPU memory constraints into the analysis of LLM inference. Based on this framework, we derive rigorous stability and instability conditions that determine whether an LLM inference service can sustain incoming demand without unbounded queue growth. This result offers a powerful tool for system deployment, potentially addressing the core challenge of GPU provisioning. By combining an estimated request arrival rate with our derived stable service rate, operators can calculate the necessary cluster size to avoid both costly over-purchasing and performance-violating under-provisioning. We further validate our theoretical predictions through extensive experiments in real GPU production environments. Our results show that the predicted stability conditions are highly accurate, with deviations typically within 10%.

Large language models have been widely adopted across different tasks, but their auto-regressive generation nature often leads to inefficient resource utilization during inference. While batching is commonly used to increase throughput, performance gains plateau beyond a certain batch size, especially with smaller models, a phenomenon that existing literature typically explains as a shift to the compute-bound regime. In this paper, through an in-depth GPU-level analysis, we reveal that large-batch inference remains memory-bound, with most GPU compute capabilities underutilized due to DRAM bandwidth saturation as the primary bottleneck. To address this, we propose a Batching Configuration Advisor (BCA) that optimizes memory allocation, reducing GPU memory requirements with minimal impact on throughput. The freed memory and underutilized GPU compute capabilities can then be leveraged by concurrent workloads. Specifically, we use model replication to improve serving throughput and GPU utilization. Our findings challenge conventional assumptions about LLM inference, offering new insights and practical strategies for improving resource utilization, particularly for smaller language models.

The performance of multi-turn, agentic LLM inference is increasingly dominated by KV-Cache storage I/O rather than computation. In prevalent disaggregated architectures, loading the massive KV-Cache from external storage creates a fundamental imbalance: storage NICs on prefill engines become bandwidth-saturated, while those on decoding engines remain idle. This asymmetry severely constrains overall system throughput. We present DualPath, an inference system that breaks this bottleneck by introducing dual-path KV-Cache loading. Beyond the traditional storage-to-prefill path, DualPath enables a novel storage-to-decode path, in which the KV-Cache is loaded into decoding engines and then efficiently transferred to prefill engines via RDMA over the compute network. DualPath combines this optimized data path -- which inherently avoids network congestion and avoids interference with latency-critical model execution communications -- with a global scheduler that dynamically balances load across prefill and decode engines. Our evaluation on three models with production agentic workloads demonstrates that DualPath improves offline inference throughput by up to 1.87times on our in-house inference system. It can also improve online serving throughput by an average factor of 1.96times without violating SLO.

The recent progress of AI can be largely attributed to large language models (LLMs). However, their escalating memory requirements introduce challenges for machine learning (ML) researchers and engineers. Addressing this requires developers to partition a large model to distribute it across multiple GPUs or TPUs. This necessitates considerable coding and intricate configuration efforts with existing model parallel tools, such as Megatron-LM, DeepSpeed, and Alpa. These tools require users' expertise in machine learning systems (MLSys), creating a bottleneck in LLM development, particularly for developers without MLSys background. In this work, we present Redco, a lightweight and user-friendly tool crafted to automate distributed training and inference for LLMs, as well as to simplify ML pipeline development. The design of Redco emphasizes two key aspects. Firstly, to automate model parallism, our study identifies two straightforward rules to generate tensor parallel strategies for any given LLM. Integrating these rules into Redco facilitates effortless distributed LLM training and inference, eliminating the need of additional coding or complex configurations. We demonstrate the effectiveness by applying Redco on a set of LLM architectures, such as GPT-J, LLaMA, T5, and OPT, up to the size of 66B. Secondly, we propose a mechanism that allows for the customization of diverse ML pipelines through the definition of merely three functions, eliminating redundant and formulaic code like multi-host related processing. This mechanism proves adaptable across a spectrum of ML algorithms, from foundational language modeling to complex algorithms like meta-learning and reinforcement learning. Consequently, Redco implementations exhibit much fewer code lines compared to their official counterparts.

Transformer based large language models have achieved tremendous success. However, the significant memory and computational costs incurred during the inference process make it challenging to deploy large models on resource-constrained devices. In this paper, we investigate compression and efficient inference methods for large language models from an algorithmic perspective. Regarding taxonomy, similar to smaller models, compression and acceleration algorithms for large language models can still be categorized into quantization, pruning, distillation, compact architecture design, dynamic networks. However, Large language models have two prominent characteristics compared to smaller models: (1) Most of compression algorithms require finetuning or even retraining the model after compression. The most notable aspect of large models is the very high cost associated with model finetuning or training. Therefore, many algorithms for large models, such as quantization and pruning, start to explore tuning-free algorithms. (2) Large models emphasize versatility and generalization rather than performance on a single task. Hence, many algorithms, such as knowledge distillation, focus on how to preserving their versatility and generalization after compression. Since these two characteristics were not very pronounced in early large models, we further distinguish large language models into medium models and ``real'' large models. Additionally, we also provide an introduction to some mature frameworks for efficient inference of large models, which can support basic compression or acceleration algorithms, greatly facilitating model deployment for users.

Deploying large language models (LLMs) on mobile devices is an emerging trend to enable data privacy and offline accessibility of LLM applications. Modern mobile neural processing units (NPUs) make such deployment increasingly feasible. However, existing mobile LLM inference frameworks suffer from high start-up latency due to their inevitable cold starts, i.e., launching LLM inferences when the model is not hosted in device memory. In this paper, we identify the key bottleneck of mobile LLM cold starts as the waste of flash bandwidth on unimportant model parameters. We design EdgeFlow, a mobile LLM inference framework that mitigates the cold start issue by adaptively adjusting the precisions of LLM parameters. Specifically, EdgeFlow leverages 1) an NPU-aware adaptive quantization algorithm that assigns different precisions to weights in a finer granularity according to their importance and NPU constraints, 2) an SIMD-friendly packing format that accelerates the transformation of various-precision weights into fixed-sized NPU-native data types, and 3) a synergistic granular pipeline that coordinates CPU and NPU computation in a fine-grained and dynamic manner. Experimental results show that EdgeFlow reduces cold-start latency by up to 4.07x compared with three state-of-the-art mobile LLM inference frameworks, i.e., llama.cpp, MNN, and llm.npu, under comparable model accuracy.

Optimizing the trade-off among predictive performance and computational cost is a central focus in the deployment of Large Language Models (LLMs). Current routing methods primarily rely on direct mapping from queries to models based on surface-level features, making them susceptible to the memorization trap and leading to poor generalizability on out-of-distribution (OOD) data. In this paper, we propose DecoR, a novel routing framework that recasts the routing task as a matching process of sifting similar queries from historical logs, effectively mitigating the memorization trap. To enhance matching accuracy, we introduce a query capability deconstruction method that decouples linguistic surface forms from task-intrinsic requirements, directing matching toward capability dimensions to ground decisions in essential task attributes. Furthermore, we develop CodaSet, a comprehensive benchmark for assessing routing generalization, where experimental results demonstrate that DecoR maintains superior accuracy while substantially lowering inference costs across both in-distribution and OOD settings. All the codes and data are available at https://github.com/lvbotenbest/DecoR.

Efficient inference for on-device Large Language Models (LLMs) remains challenging due to limited hardware resources and the high cost of the prefill stage, which processes the full input context to construct Key-Value (KV) caches. We present SparKV, an adaptive KV loading framework that combines cloud-based KV streaming with on-device computation. SparKV models the cost of individual KV chunks and decides whether each chunk should be streamed or computed locally, while overlapping the two execution paths to reduce latency. To handle fluctuations in wireless connectivity and edge resource availability, SparKV further refines offline-generated schedules at runtime to rebalance communication and computation costs. Experiments across diverse datasets, LLMs, and edge devices show that SparKV reduces Time-to-First-Token by 1.3$x-5.1x with negligible impact on response quality, while lowering per-request energy consumption by 1.5x to 3.3x, demonstrating its robustness and practicality for real-world on-device deployment.

The scaling of Large Language Models (LLMs) drives interest in matrix-based optimizers (e.g., Shampoo, Muon, SOAP) for their convergence efficiency; yet their requirement for holistic updates conflicts with the tensor fragmentation in distributed frameworks like Megatron. Existing solutions are suboptimal: synchronous approaches suffer from computational redundancy, while layer-wise partitioning fails to reconcile this conflict without violating the geometric constraints of efficient communication primitives. To bridge this gap, we propose Canzona, a Unified, Asynchronous, and Load-Balanced framework that decouples logical optimizer assignment from physical parameter distribution. For Data Parallelism, we introduce an alpha-Balanced Static Partitioning strategy that respects atomicity while neutralizing the load imbalance. For Tensor Parallelism, we design an Asynchronous Compute pipeline utilizing Micro-Group Scheduling to batch fragmented updates and hide reconstruction overhead. Extensive evaluations on the Qwen3 model family (up to 32B parameters) on 256 GPUs demonstrate that our approach preserves the efficiency of established parallel architectures, achieving a 1.57x speedup in end-to-end iteration time and reducing optimizer step latency by 5.8x compared to the baseline.

We present ROCKET, a training-free model compression method that achieves state-of-the-art performance in comparison with factorization, structured-sparsification and dynamic compression baselines. Operating under a global compression budget, ROCKET comprises two key innovations: First, it formulates layer-wise compression allocation as a multi-choice knapsack problem, selecting the optimal compression level for each layer to minimize total reconstruction error while adhering to a target model size. Second, it introduces a single-step sparse matrix factorization inspired by dictionary learning: using only a small calibration set, it sparsifies weight coefficients based on activation-weights sensitivity and then updates the dictionary in closed form via least squares bypassing iterative optimization, sparse coding, or backpropagation entirely. ROCKET consistently outperforms existing compression approaches across different model architectures at 20-50\% compression rates. Notably, it retains over 90\% of the original model's performance at 30\% compression without any fine-tuning. Moreover, when applying a light fine-tuning phase, recovery is substantially enhanced: for instance, compressing Qwen3-14B to an 8B-parameter model and healing it with just 30 million tokens yields performance nearly on par with the original Qwen3-8B. The code for ROCKET is at github.com/mts-ai/ROCKET/tree/main.

Large-scale machine learning models deliver strong performance across a wide range of tasks but come with significant computational and resource constraints. To mitigate these challenges, local smaller models are often deployed alongside larger models, relying on routing and deferral mechanisms to offload complex tasks. However, existing approaches inadequately balance the capabilities of these models, often resulting in unnecessary deferrals or sub-optimal resource usage. In this work we introduce a novel loss function called Gatekeeper for calibrating smaller models in cascade setups. Our approach fine-tunes the smaller model to confidently handle tasks it can perform correctly while deferring complex tasks to the larger model. Moreover, it incorporates a mechanism for managing the trade-off between model performance and deferral accuracy, and is broadly applicable across various tasks and domains without any architectural changes. We evaluate our method on encoder-only, decoder-only, and encoder-decoder architectures. Experiments across image classification, language modeling, and vision-language tasks show that our approach substantially improves deferral performance.

As LLM agents are deployed in long-horizon sessions, context accumulation drives up inference costs. Existing approaches utilize text pruning or dynamic memory eviction to minimize token footprints; however, their unconstrained sequence mutations alter layouts, introducing prefix mismatches and cache invalidation. This reveals a critical trade-off between text sparsity and prompt cache continuity. To address this, we present TokenPilot, a dual-granularity context management framework. Globally, Ingestion-Aware Compaction acts as a framework harness to stabilize prompt prefixes and eliminate open-world environmental noise at the ingestion gate. Locally, Lifecycle-Aware Eviction monitors the ongoing residual utility of context segments, enforcing a conservative batch-turn schedule to offload content segments only when task relevance expires. Experiments on PinchBench and Claw-Eval under both isolated and continuous modes demonstrate that TokenPilot reduces costs by 61% and 56% in isolated mode, and 61% and 87% in continuous mode, while maintaining competitive performance compared to prior systems. TokenPilot has been integrated into LightMem2 at https://github.com/zjunlp/LightMem2.

Many deep learning applications benefit from using large models with billions of parameters. Training these models is notoriously expensive due to the need for specialized HPC clusters. In this work, we consider alternative setups for training large models: using cheap "preemptible" instances or pooling existing resources from multiple regions. We analyze the performance of existing model-parallel algorithms in these conditions and find configurations where training larger models becomes less communication-intensive. Based on these findings, we propose SWARM parallelism, a model-parallel training algorithm designed for poorly connected, heterogeneous and unreliable devices. SWARM creates temporary randomized pipelines between nodes that are rebalanced in case of failure. We empirically validate our findings and compare SWARM parallelism with existing large-scale training approaches. Finally, we combine our insights with compression strategies to train a large Transformer language model with 1B shared parameters (approximately 13B before sharing) on preemptible T4 GPUs with less than 200Mb/s network.

Disaggregated inference architectures physically separate prefill and decode phases onto distinct GPU pools, creating competing "agents" that share a fixed hardware budget. We provide, to our knowledge, the first formal game-theoretic analysis of this architecture, using NVIDIA Dynamo as a concrete case study. We model disaggregated serving as three coupled games: a two-player resource game between prefill and decode pools, a selfish caching game over the hierarchical KV cache, and a congestion game with positive externalities for request routing. We empirically validate the latter two; the P/D resource game is treated analytically (Section 9.2). We characterize how GPU saturation induces regime transitions that shift the game's payoff structure: below saturation, selfish behavior has bounded Price of Anarchy (PoA); at saturation, superlinear latency and cache externalities drive our empirical estimator PoA-hat (defined in Section 6.4) upward. Based on this analysis, we design an adaptive controller that detects saturation transitions in real time and adjusts routing parameters accordingly, shifting from cache-affinity exploitation to load-balanced congestion avoidance. We instantiate our framework on a 3-node NVIDIA B200 cluster running Dynamo with two models, Nemotron-4-340B (TP=8, full-node workers with cross-InfiniBand KV transfers) and Llama-3.1-70B (TP=4), and find the same three-regime PoA-hat structure with the same first post-knee grid point (C=128) on both models. Adaptive routing shifts each model to a better operating point. Our strongest result is on the 70B 1P/5D topology, where PoA-hat drops 3.1x (66.4 to 21.5) in the saturated phase at a 13% throughput cost. On the 70B 1P/2D, PoA-hat drops 2.2x and TTFT P99 drops 7.6x (see Section 8.5).

Adapting pre-trained large language models (LLMs) is crucial but challenging due to their enormous size. Parameter-efficient fine-tuning (PEFT) techniques typically employ additive adapters applied to frozen model weights. To further reduce memory usage, model weights can be compressed through quantization. However, existing PEFT methods often yield suboptimal model quality due to restrictive assumptions, such as imposing low-rank constraints on adapters to reduce trainable parameters. We find that sketching, a popular data compression technique, can serve as an efficient adaptation strategy for LLMs while avoiding low-rank assumptions. We introduce SketchTune, a compressive adaptation strategy that compresses LLM weights into compact fine-tunable sketches, integrating compression and adaptation into a unified framework. This integration eliminates the need for complex two-path computation common in existing PEFT techniques, enabling faster and more memory-efficient training and inference. SketchTune is supported by mathematical insights into matrix classes that are better approximated using sketching rather than low-rank methods. Our rigorous evaluations with Llama-1/2/3 models demonstrate that SketchTune outperforms leading PEFT methods across diverse tasks including math problem-solving, common sense reasoning, and instruction following, while using substantially smaller base models and comparable trainable parameters. As a highlight, SketchTune outperforms LoRA, DoRA, and S2FT on commonsense and math benchmarks using 2.6-3.5times smaller base models and exceeds LoftQ in accuracy by 14.48% on GSM8K with 7.3times fewer trainable parameters.

Fine-grained Mixture-of-Experts (MoE) models sparsely activate only a subset of experts per token, reducing activated computation while maintaining high model capacity. However, in memory-constrained inference scenarios, only a small set of experts can be cached. Experts not in the cache must be fetched from slow external storage (e.g., UFS), leading to frequent evictions and substantial I/O overhead. We propose ReMoE, a router fine-tuning framework designed to boost token-wise expert reuse. ReMoE biases the router toward recently selected experts, producing temporally stable routing that better matches cache locality constraints. By increasing short-horizon expert reuse, ReMoE reduces expert fetches from storage without adding inference-time computation. Experiments on DeepSeek and Qwen models show that ReMoE improves expert reuse by 26% while maintaining downstream task performance. Real-system evaluations further confirm these benefits, improving output throughput by 8.4% under vLLM GPU-CPU expert offloading and reducing TPOT by 43.6-49.8% under llama.cpp on Jetson Orin NX, corresponding to a 1.77-1.99times decode speedup across diverse workloads. Checkpoints and usage instructions are available at https://github.com/BUAA-OSCAR/ReMoE.

The Mixture of Experts (MoE) architecture has demonstrated significant advantages as it enables to increase the model capacity without a proportional increase in computation. However, the large MoE model size still introduces substantial memory demands, which usually requires expert offloading on resource-constrained platforms and incurs significant overhead. Hybrid CPU-GPU inference has been proposed to leverage CPU computation to reduce expert loading overhead but faces major challenges: on one hand, the expert activation patterns of MoE models are highly unstable, rendering the fixed mapping strategies in existing works inefficient; on the other hand, the hybrid CPU-GPU schedule for MoE is inherently complex due to the diverse expert sizes, structures, uneven workload distribution, etc. To address these challenges, in this paper, we propose HybriMoE, a hybrid CPU-GPU inference framework that improves resource utilization through a novel CPU-GPU scheduling and cache management system. HybriMoE introduces (i) a dynamic intra-layer scheduling strategy to balance workloads across CPU and GPU, (ii) an impact-driven inter-layer prefetching algorithm, and (iii) a score-based caching algorithm to mitigate expert activation instability. We implement HybriMoE on top of the kTransformers framework and evaluate it on three widely used MoE-based LLMs. Experimental results demonstrate that HybriMoE achieves an average speedup of 1.33times in the prefill stage and 1.70times in the decode stage compared to state-of-the-art hybrid MoE inference framework. Our code is available at: https://github.com/PKU-SEC-Lab/HybriMoE.

Efficient long-context LLM deployment is stalled by a dichotomy between amortized compression, which struggles with out-of-distribution generalization, and Test-Time Training, which incurs prohibitive synthetic data costs and requires modifying model weights, creating stateful parameters that complicate concurrent serving. We propose Latent Context Compilation, a framework that fundamentally shifts context processing from adaptation to compilation. By utilizing a disposable LoRA module as a compiler, we distill long contexts into compact buffer tokens -- stateless, portable memory artifacts that are plug-and-play compatible with frozen base models. Crucially, we introduce a self-aligned optimization strategy that eliminates the need for synthetic context-relevant QA pairs. By regularizing context reconstruction task with context-agnostic random queries, we force compressed tokens to reside within the model's existing instruction-following manifold. Experiments with Llama-3.1-8B demonstrate that Latent Context Compilation preserves fine-grained details and reasoning capabilities where prior methods falter, effectively decoupling memory density from model parameters even at a 16x compression ratio.

Mixture-of-Experts (MoE) architectures offer a general solution to the high inference costs of large language models (LLMs) via sparse routing, bringing faster and more accurate models, at the cost of massive parameter counts. For example, the SwitchTransformer-c2048 model has 1.6 trillion parameters, requiring 3.2TB of accelerator memory to run efficiently, which makes practical deployment challenging and expensive. In this paper, we present a solution to this memory problem, in form of a new compression and execution framework called QMoE. Specifically, QMoE consists of a scalable algorithm which accurately compresses trillion-parameter MoEs to less than 1 bit per parameter, in a custom format co-designed with bespoke GPU decoding kernels to facilitate efficient end-to-end compressed inference, with minor runtime overheads relative to uncompressed execution. Concretely, QMoE can compress the 1.6 trillion parameter SwitchTransformer-c2048 model to less than 160GB (20x compression, 0.8 bits per parameter) at only minor accuracy loss, in less than a day on a single GPU. This enables, for the first time, the execution of a trillion-parameter model on affordable commodity hardware, like a single server with 4x NVIDIA A6000 or 8x NVIDIA 3090 GPUs, at less than 5% runtime overhead relative to ideal uncompressed inference. The source code and compressed models are available at github.com/IST-DASLab/qmoe.

Transformer-based long context generative models power emerging AI applications like hour-long video understanding and project-level coding agent. Deploying long context transformers (e.g., 100K to 10M tokens) is prohibitively expensive compared to short context (e.g., 4K tokens) model variants. Reducing the cost of long-context transformers is becoming a pressing research and engineering challenge starting from the year of 2024. This work describes a concurrent programming framework for quantitatively analyzing the efficiency challenges in serving multiple long-context requests under limited size of GPU high-bandwidth memory (HBM) regime. We give a detailed analysis of how all additional computational costs, compared to 4K context, trace back to one single source: the large size of the KV cache. We use a 34B GPT-3.5 level model of 50K context on A100 NVLink as a running example, and describe how its large KV cache causes four types of deployment challenges: (1) prefilling long inputs takes much longer compute time and GPU memory than short inputs; (2) after prefilling, the large KV cache residing on the GPU HBM substantially restricts the number of concurrent users being served; (3) during decoding, repeatedly reading the KV cache from HBM to SM largely increases latency; (4) when KV cache memory overflows, swapping it from HBM to DDR causes significant context switching latency. We use this framework to analyze existing works and identify possibilities of combining them to build end-to-end systems. Overall, this work offers a foundational framework for analyzing long context transformer deployment and identifies directions towards reducing the inference cost of 1M context to be as cheap as 4K.

Large Language Models increasingly rely on distributed techniques for their training and inference. These techniques require communication across devices which can reduce scaling efficiency as the number of devices increases. While some distributed techniques can overlap, and thus, hide this communication with independent computations, techniques such as Tensor Parallelism (TP) inherently serialize communication with model execution. One approach to hide this serialized communication is to interleave it with the producer operation (of the communicated data) in a fine-grained manner. However, this fine-grained interleaving of communication and computation in software can be difficult. Furthermore, as with any concurrent execution, it requires compute and memory resources to be shared between computation and communication, causing resource contention that reduces overlapping efficacy. To overcome these challenges, we propose T3 which applies hardware-software co-design to transparently overlap serialized communication while minimizing resource contention with compute. T3 transparently fuses producer operations with the subsequent communication via a simple configuration of the producer's output address space and requires minor software changes. At the hardware level, T3 adds a lightweight track and trigger mechanism to orchestrate the producer's compute, and communication. It further uses compute-enhanced memories for communication's attendant compute. As a result, T3 reduces resource contention, and efficiently overlaps serialized communication with computation. For important Transformer models like T-NLG, T3 speeds up communication-heavy sublayers by 30% geomean (max 47%) and reduces data movement by 22% geomean (max 36%). Furthermore, T3's benefits persist as models scale: geomean 29% for sublayers in sim500-billion parameter models, PALM and MT-NLG.

The rapid growth of large language models (LLMs) has outpaced the evolution of single-GPU hardware, making model scale increasingly constrained by memory capacity rather than computation. While modern training systems extend GPU memory through distributed parallelism and offloading across CPU and storage tiers, they fundamentally retain a GPU-centric execution paradigm in which GPUs host persistent model replicas and full autograd graphs. As a result, scaling large models remains tightly coupled to multi-GPU clusters, complex distributed runtimes, and unpredictable host memory consumption, creating substantial barriers for node-scale post-training workloads such as instruction tuning, alignment, and domain adaptation. We present Horizon-LM, a memory-centric training system that redefines the roles of CPU and GPU for large-model optimization. Horizon-LM treats host memory as the authoritative parameter store and uses GPUs solely as transient compute engines through a CPU-master, GPU-template execution model. By eliminating persistent GPU-resident modules and autograd graphs, employing explicit recomputation with manual gradient propagation, and introducing a pipelined double-buffered execution engine, Horizon-LM decouples model scale from GPU count and bounds memory usage to the theoretical parameter footprint. On a single H200 GPU with 1.5\,TB host RAM, Horizon-LM reliably trains models up to 120B parameters. On a standard single A100 machine, Horizon-LM achieves up to 12.2times higher training throughput than DeepSpeed ZeRO-3 with CPU offloading while preserving numerical correctness. Across platforms and scales, Horizon-LM sustains high device utilization and predictable memory growth, demonstrating that host memory, not GPU memory, defines the true feasibility boundary for node-scale large-model training.

Model compression is increasingly essential for deploying large language models (LLMs), yet existing evaluations are limited in method coverage and focus primarily on knowledge-centric benchmarks. Thus, we introduce UniComp, a unified evaluation framework for comparing pruning, quantization, and knowledge distillation. UniComp evaluates compressed models along three dimensions: performance, reliability, and efficiency, using a diverse set of capability- and safety-oriented benchmarks together with a hardware-aware efficiency analysis. Through extensive evaluation of six compression techniques on modern LLMs across more than 40 datasets, we find that (i) compression exhibits a consistent knowledge bias, where knowledge-intensive tasks are relatively preserved while reasoning, multilingual, and instruction-following capabilities degrade substantially; (ii) quantization provides the best overall trade-off between retained performance and efficiency, whereas distillation yields strong runtime acceleration gains at high computational cost; and (iii) task-specific calibration can significantly improve the reasoning ability of pruned models by up to 50%.

Collocating deep learning training tasks improves GPU utilization but causes drastic slowdowns due to resource contention and risks Out-of-Memory (OOM) failures. Accurate memory estimation is essential for robust collocation, while GPU utilization -- a key proxy for resource contention -- enables interference-aware scheduling to reduce slowdowns and improve throughput. Existing GPU memory estimators span three paradigms -- analytical models, CPU-side libraries, and ML-based estimators -- each with distinct limitations: dependence on detailed model specifications, intrusive integration, poor generalization, and varying latency overhead. GPU heterogeneity further complicates estimation, as identical tasks can exhibit markedly different memory footprints across hardware generations. GPU utilization remains comparatively understudied, further complicated by the non-additive nature of utilization metrics and hardware sensitivity. We conduct a systematic analysis of representative estimators from each paradigm -- Horus, PyTorch FakeTensor, and our lightweight ML-based estimator -- evaluating accuracy, generalizability, and practical overhead. We construct a synthetic dataset spanning MLPs, CNNs, and Transformers with controlled architectural variations, and train MLP- and Transformer-based estimators for memory prediction. We further experiment with utilization estimation on the same dataset. Our evaluation reveals key tradeoffs and validates estimators against real-world unseen models. Significant challenges remain: analytical models are hardware-dependent, CPU-side libraries impose intrusive integration costs, and ML-based estimators struggle with cross-architecture generalization. We release all datasets, tools, and artifacts to support further research.

Pre-training large language models (LLMs) increasingly requires distributed compute, yet bandwidth constraints make it difficult to scale beyond well-provisioned datacenters-especially when model parallelism forces frequent, large inter-device communications. We study whether SparseLoCo, a low-communication data parallel method based on infrequent synchronization and sparse pseudo-gradient exchange, can be combined with low-bandwidth pipeline model parallelism via activation and activation-gradient compression. We introduce a heterogeneous distributed training framework where some participants host full replicas on high-bandwidth interconnects, while resource-limited participants are grouped to jointly instantiate a replica using pipeline parallelism with subspace-projected inter-stage communication. To make the recently introduced subspace pipeline compression compatible with SparseLoCo, we study a number of adaptations. Across large-scale language modeling experiments (178M-1B parameters) on standard pretraining corpora, we find that activation compression composes with SparseLoCo at modest cost, while selective (heterogeneous) compression consistently improves the loss-communication tradeoff relative to compressing all replicas-especially at aggressive compression ratios. These results suggest a practical path to incorporating low-bandwidth model parallelism and heterogeneous participants into LLM pre-training.

Model routing allocates queries to the suitable model, improving system performance while reducing costs. However, existing routing methods face practical limitations that hinder scalability in large-scale applications and struggle to keep up with the rapid growth of the large language model (LLM) ecosystem. To tackle these challenges, we propose TagRouter, a training-free model routing method designed to optimize the synergy among multiple LLMs for open-domain text generation tasks. Experimental results demonstrate that TagRouter outperforms 13 baseline methods, increasing the accept rate of system by 6.15% and reducing costs by 17.20%, achieving optimal cost-efficiency. Our findings provides the LLM community with an efficient and scalable solution for model ensembling, offering users an evolvable "super model."

As the input length of large language model (LLM) serving continues to grow, the KV cache has become a dominant bottleneck in AI infrastructure. It limits GPU memory capacity, serving concurrency, cache reuse, and distributed scalability. Multiple important problems, including position-independent KV cache, prefix KV cache compression, hot/cold KV cache separation, and distributed KV cache management, all depend on how the KV cache is represented and managed. However, existing serving systems largely rely on a monolithic KV cache abstraction, where the KV cache is treated as a homogeneous sequence of token-level memory blocks and managed with similar policies across attention heads and serving scenarios. We observe that KV cache utility is highly structured across KV heads: different heads exhibit different functional roles, attention distances, and runtime importance. Therefore, a full KV cache is not always necessary for every head, token range, or serving scenario. We present RedKnot, a head-aware KV cache management system for LLM serving. RedKnot breaks the conventional monolithic KV cache abstraction by decomposing the KV cache along KV heads, whose importance and effective attention ranges vary significantly across serving scenarios. This head-level decomposition turns the KV cache from a monolithic tensor abstraction into a structured memory object, enabling RedKnot to uniformly support position-independent KV reuse, prefix KV compression, hot/cold KV separation, and distributed KV placement while preserving output fidelity and improving resource efficiency, without requiring model retraining or fine-tuning. RedKnot establishes a new foundation for AI infrastructure by transforming the KV cache from a monolithic, passive runtime artifact into a dynamic, model-aware runtime substrate for scalable LLM serving.

Large language model (LLM) routing assigns each query to the most suitable model from an ensemble. We introduce LLMRouterBench, a large-scale benchmark and unified framework for LLM routing. It comprises over 400K instances from 21 datasets and 33 models. Moreover, it provides comprehensive metrics for both performance-oriented routing and performance-cost trade-off routing, and integrates 10 representative routing baselines. Using LLMRouterBench, we systematically re-evaluate the field. While confirming strong model complementarity-the central premise of LLM routing-we find that many routing methods exhibit similar performance under unified evaluation, and several recent approaches, including commercial routers, fail to reliably outperform a simple baseline. Meanwhile, a substantial gap remains to the Oracle, driven primarily by persistent model-recall failures. We further show that backbone embedding models have limited impact, that larger ensembles exhibit diminishing returns compared to careful model curation, and that the benchmark also enables latency-aware analysis. All code and data are available at https://github.com/ynulihao/LLMRouterBench.

Serving disaggregated large language models (LLMs) over tens of thousands of xPU devices (GPUs or NPUs) with reliable performance faces multiple challenges. 1) Ignoring the diversity (various prefixes and tidal requests), treating all the prompts in a mixed pool is inadequate. To facilitate the similarity per scenario and minimize the inner mismatch on P/D (prefill and decoding) processing, fine-grained organization is required, dynamically adjusting P/D ratios for better performance. 2) Due to inaccurate estimation on workload (queue status or maintained connections), the global scheduler easily incurs unnecessary timeouts in prefill. 3) Block-fixed device-to-device (D2D) KVCache transfer over cluster-level RDMA (remote direct memory access) fails to achieve desired D2D utilization as expected. To overcome previous problems, this paper proposes an end-to-end system P/D-Serve, complying with the paradigm of MLOps (machine learning operations), which models end-to-end (E2E) P/D performance and enables: 1) fine-grained P/D organization, mapping the service with RoCE (RDMA over converged ethernet) as needed, to facilitate similar processing and dynamic adjustments on P/D ratios; 2) on-demand forwarding upon rejections for idle prefill, decoupling the scheduler from regular inaccurate reports and local queues, to avoid timeouts in prefill; and 3) efficient KVCache transfer via optimized D2D access. P/D-Serve is implemented upon Ascend and MindSpore, has been deployed over tens of thousands of NPUs for more than eight months in commercial use, and further achieves 60\%, 42\% and 46\% improvements on E2E throughput, time-to-first-token (TTFT) SLO (service level objective) and D2D transfer time. As the E2E system with optimizations, P/D-Serve achieves 6.7x increase on throughput, compared with aggregated LLMs.

Large language models (LLMs) have revolutionized numerous applications, yet their deployment remains challenged by memory constraints on local devices. While scaling laws have enhanced LLM capabilities, the primary bottleneck has shifted from capability to availability, emphasizing the need for efficient memory management. Traditional compression methods, such as quantization, often require predefined compression ratios and separate compression processes for each setting, complicating deployment in variable memory environments. In this paper, we introduce BitStack, a novel, training-free weight compression approach that enables megabyte-level trade-offs between memory usage and model performance. By leveraging weight decomposition, BitStack can dynamically adjust the model size with minimal transmission between running memory and storage devices. Our approach iteratively decomposes weight matrices while considering the significance of each parameter, resulting in an approximately 1-bit per parameter residual block in each decomposition iteration. These blocks are sorted and stacked in storage as basic transmission units, with different quantities loaded based on current memory availability. Extensive experiments across a wide range of tasks demonstrate that, despite offering fine-grained size control, BitStack consistently matches or surpasses strong quantization baselines, particularly at extreme compression ratios. To the best of our knowledge, this is the first decomposition-based method that effectively bridges the gap to practical compression techniques like quantization. Code is available at https://github.com/xinghaow99/BitStack.