Daily Papers

Speculative decoding is a technique that uses multiple language models to accelerate infer- ence. Previous works have used an experi- mental approach to optimize the throughput of the inference pipeline, which involves LLM training and can be costly. This study of spec- ulative decoding proposes a theory that ana- lytically connects the key hyperparameters of pre-trained LLMs to the throughput efficiency of a downstream SD-based inference system. The theory allows the prediction of throughput- optimal hyperparameters for the components of an inference system before their pre-training.

KV cache compression promises increased throughput and efficiency with negligible loss in performance. While the gains in throughput are indisputable and recent literature has indeed shown minimal degradation on particular benchmarks, in general the consequences of compression in realistic scenarios such as multi-instruction prompting have been insufficiently studied. In this paper, we identify several pitfalls practitioners should be aware of when deploying KV cache compressed LLMs. Importantly, we show that certain instructions degrade much more rapidly with compression, effectively causing them to be completely ignored by the LLM. As a practical example of that, we highlight system prompt leakage as a case study, empirically showing the impact of compression on leakage and general instruction following. We show several factors that play a role in prompt leakage: compression method, instruction order, and KV eviction bias. We then propose simple changes to KV cache eviction policies that can reduce the impact of these factors and improve the overall performance in multi-instruction tasks.

AlphaFold predicts protein structures from the amino acid sequence at or near experimental resolution, solving the 50-year-old protein folding challenge, leading to progress by transforming large-scale genomics data into protein structures. AlphaFold will also greatly change the scientific research model from low-throughput to high-throughput manner. The AlphaFold framework is a mixture of two types of workloads: MSA construction based on CPUs and model inference on GPUs. The first CPU stage dominates the overall runtime, taking hours for a single protein due to the large database sizes and I/O bottlenecks. However, GPUs in this CPU stage remain idle, resulting in low GPU utilization and restricting the capacity of large-scale structure predictions. Therefore, we proposed ParaFold, an open-source parallel version of AlphaFold for high throughput protein structure predictions. ParaFold separates the CPU and GPU parts to enable large-scale structure predictions. ParaFold also effectively reduces the CPU and GPU runtime with two optimizations without compromising the quality of prediction results: using multi-threaded parallelism on CPUs and using optimized JAX compilation on GPUs. We evaluated ParaFold with three datasets of different size and protein lengths. We evaluated the accuracy and efficiency of optimizations on CPUs and GPUs, and showed the large-scale prediction capability by running ParaFold inferences of 19,704 small proteins in five hours on one NVIDIA DGX-2. Using the JAX compile optimization, ParaFold attained a 13.8X average speedup over AlphaFold. ParaFold offers a rapid and effective approach for high-throughput structure predictions, leveraging the predictive power by running on supercomputers, with shorter time, and at a lower cost. The development of ParaFold will greatly speed up high-throughput studies and render the protein "structure-omics" feasible.

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.

Generating long-form content, such as minute-long videos and extended texts, is increasingly important for modern generative models. Block diffusion improves inference efficiency via KV caching and block-wise causal inference and has been widely adopted in diffusion language models and video generation. However, in long-context settings, block diffusion still incurs substantial overhead from repeatedly computing attention over a growing KV cache. We identify an underexplored property of block diffusion: cross-step redundancy of attention within a block. Our analysis shows that attention outputs from tokens outside the current block remain largely stable across diffusion steps, while block-internal attention varies significantly. Based on this observation, we propose FlashBlock, a cached block-external attention mechanism that reuses stable attention output, reducing attention computation and KV cache access without modifying the diffusion process. Moreover, FlashBlock is orthogonal to sparse attention and can be combined as a complementary residual reuse strategy, substantially improving model accuracy under aggressive sparsification. Experiments on diffusion language models and video generation demonstrate up to 1.44times higher token throughput and up to 1.6times reduction in attention time, with negligible impact on generation quality. Project page: https://caesarhhh.github.io/FlashBlock/.

Automatic software system optimization can improve software speed, reduce operating costs, and save energy. Traditional approaches to optimization rely on manual tuning and compiler heuristics, limiting their ability to generalize across diverse codebases and system contexts. Recent methods using Large Language Models (LLMs) offer automation to address these limitations, but often fail to scale to the complexity of real-world software systems and applications. We present SysLLMatic, a system that integrates LLMs with profiling-guided feedback and system performance insights to automatically optimize software code. We evaluate it on three benchmark suites: HumanEval_CPP (competitive programming in C++), SciMark2 (scientific kernels in Java), and DaCapoBench (large-scale software systems in Java). Results show that SysLLMatic can improve system performance, including latency, throughput, energy efficiency, memory usage, and CPU utilization. It consistently outperforms state-of-the-art LLM baselines on microbenchmarks. On large-scale application codes, it surpasses traditional compiler optimizations, achieving average relative improvements of 1.85x in latency and 2.24x in throughput. Our findings demonstrate that LLMs, guided by principled systems thinking and appropriate performance diagnostics, can serve as viable software system optimizers. We further identify limitations of our approach and the challenges involved in handling complex applications. This work provides a foundation for generating optimized code across various languages, benchmarks, and program sizes in a principled manner.

Training multi-billion to trillion-parameter language models efficiently on GPU clusters requires leveraging multiple parallelism strategies. We present Galvatron, a novel open-source framework (dubbed 'Optimus-Megatron' in the implementation) that dynamically combines data parallelism, tensor model parallelism, and pipeline parallelism to optimize training throughput. Built atop PyTorch and integrating NVIDIA's Megatron-LM and Microsoft's DeepSpeed, Galvatron automatically selects and adjusts parallelism strategies in real time based on model architecture, hardware, and training dynamics. This paper details Galvatron's key features -- automatic hybrid parallelism selection, layer-wise and phase-wise strategy optimization, and runtime adaptation -- and contrasts them with existing static frameworks. We describe the system's technical stack, including its use of DeepSpeed's ZeRO and NCCL communication, and provide an in-depth implementation overview of its core modules (profilers, strategy selector, parallelism manager). We then illustrate how Galvatron can be seamlessly integrated into existing training pipelines with minimal code modifications, providing companies a plug-and-play solution for efficient large-model training. Finally, we situate Galvatron in context with related efforts (NVIDIA Megatron-LM, Microsoft DeepSpeed, Google GShard, Meta FairScale, etc.), highlighting how it advances the state of the art in distributed deep learning. References to the GitHub repository and relevant literature are provided throughout.

Monocular Depth Estimation (MDE) has emerged as a pivotal task in computer vision, supporting numerous real-world applications. However, deploying accurate depth estimation models on resource-limited edge devices, especially Application-Specific Integrated Circuits (ASICs), is challenging due to the high computational and memory demands. Recent advancements in foundational depth estimation deliver impressive results but further amplify the difficulty of deployment on ASICs. To address this, we propose QuartDepth which adopts post-training quantization to quantize MDE models with hardware accelerations for ASICs. Our approach involves quantizing both weights and activations to 4-bit precision, reducing the model size and computation cost. To mitigate the performance degradation, we introduce activation polishing and compensation algorithm applied before and after activation quantization, as well as a weight reconstruction method for minimizing errors in weight quantization. Furthermore, we design a flexible and programmable hardware accelerator by supporting kernel fusion and customized instruction programmability, enhancing throughput and efficiency. Experimental results demonstrate that our framework achieves competitive accuracy while enabling fast inference and higher energy efficiency on ASICs, bridging the gap between high-performance depth estimation and practical edge-device applicability. Code: https://github.com/shawnricecake/quart-depth

Recurrent Neural Network (RNN) inference exhibits low hardware utilization due to the strict data dependencies across time-steps. Batching multiple requests can increase throughput. However, RNN batching requires a large amount of padding since the batched input sequences may largely differ in length. Schemes that dynamically update the batch every few time-steps avoid padding. However, they require executing different RNN layers in a short timespan, decreasing energy efficiency. Hence, we propose E-BATCH, a low-latency and energy-efficient batching scheme tailored to RNN accelerators. It consists of a runtime system and effective hardware support. The runtime concatenates multiple sequences to create large batches, resulting in substantial energy savings. Furthermore, the accelerator notifies it when the evaluation of a sequence is done, so that a new sequence can be immediately added to a batch, thus largely reducing the amount of padding. E-BATCH dynamically controls the number of time-steps evaluated per batch to achieve the best trade-off between latency and energy efficiency for the given hardware platform. We evaluate E-BATCH on top of E-PUR and TPU. In E-PUR, E-BATCH improves throughput by 1.8x and energy-efficiency by 3.6x, whereas in TPU, it improves throughput by 2.1x and energy-efficiency by 1.6x, over the state-of-the-art.

This paper introduces a parallel and asynchronous Transformer framework designed for efficient and accurate multilingual lip synchronization in real-time video conferencing systems. The proposed architecture integrates translation, speech processing, and lip-synchronization modules within a pipeline-parallel design that enables concurrent module execution through message-queue-based decoupling, reducing end-to-end latency by up to 3.1 times compared to sequential approaches. To enhance computational efficiency and throughput, the inference workflow of each module is optimized through low-level graph compilation, mixed-precision quantization, and hardware-accelerated kernel fusion. These optimizations provide substantial gains in efficiency while preserving model accuracy and visual quality. In addition, a context-adaptive silence-detection component segments the input speech stream at semantically coherent boundaries, improving translation consistency and temporal alignment across languages. Experimental results demonstrate that the proposed parallel architecture outperforms conventional sequential pipelines in processing speed, synchronization stability, and resource utilization. The modular, message-oriented design makes this work applicable to resource-constrained IoT communication scenarios including telemedicine, multilingual kiosks, and remote assistance systems. Overall, this work advances the development of low-latency, resource-efficient multimodal communication frameworks for next-generation AIoT systems.

We present a single-GPU (H100, bf16) evaluation of GPT-OSS-20B (Mixture-of-Experts; 20.9B total, approx. 3.61B active) against dense baselines Qwen3-32B and Yi-34B across multiple dimensions. We measure true time-to-first-token (TTFT), full-decode throughput (TPOT), end-to-end latency percentiles, peak VRAM with past key values (PKV) held, and energy via a consistent nvidia-smi-based sampler. At a 2048-token context with 64-token decode, GPT-OSS-20B delivers higher decode throughput and tokens per Joule than dense baselines Qwen3-32B and Yi-34B, while substantially reducing peak VRAM and energy per 1000 generated tokens; its TTFT is higher due to MoE routing overhead. With only 17.3% of parameters active (3.61B of 20.9B), GPT-OSS-20B provides about 31.8% higher decode throughput and 25.8% lower energy per 1000 generated tokens than Qwen3-32B at 2048/64, while using 31.7% less peak VRAM. Normalized by active parameters, GPT-OSS-20B shows markedly stronger per-active-parameter efficiency (APE), underscoring MoE's deployment advantages. We do not evaluate accuracy; this is a deployment-focused study. We release code and consolidated results to enable replication and extension.

Recently, various distributed strategies for large language model training have been proposed. However, these methods provided limited solutions for the trade-off between memory consumption and communication cost. In this paper, we rethink the impact of memory consumption and communication costs on the training speed of large language models, and propose a memory-communication balanced strategy set Partial Redundancy Optimizer (PaRO). PaRO provides comprehensive options which reduces the amount and frequency of inter-group communication with minor memory redundancy by fine-grained sharding strategy, thereby improving the training efficiency in various training scenarios. Additionally, we propose a Hierarchical Overlapping Ring (HO-Ring) communication topology to enhance communication efficiency between nodes or across switches in large language model training. Our experiments demonstrate that PaRO significantly improves training throughput by 1.19x-2.50x compared to the SOTA method and achieves a near-linear scalability. The HO-Ring algorithm improves communication efficiency by 36.5% compared to the traditional Ring algorithm.

Similarity search finds application in specialized database systems handling complex data such as images or videos, which are typically represented by high-dimensional features and require specific indexing structures. This paper tackles the problem of better utilizing GPUs for this task. While GPUs excel at data-parallel tasks, prior approaches are bottlenecked by algorithms that expose less parallelism, such as k-min selection, or make poor use of the memory hierarchy. We propose a design for k-selection that operates at up to 55% of theoretical peak performance, enabling a nearest neighbor implementation that is 8.5x faster than prior GPU state of the art. We apply it in different similarity search scenarios, by proposing optimized design for brute-force, approximate and compressed-domain search based on product quantization. In all these setups, we outperform the state of the art by large margins. Our implementation enables the construction of a high accuracy k-NN graph on 95 million images from the Yfcc100M dataset in 35 minutes, and of a graph connecting 1 billion vectors in less than 12 hours on 4 Maxwell Titan X GPUs. We have open-sourced our approach for the sake of comparison and reproducibility.

Recent advances in machine learning (ML) have spotlighted the pressing need for computing architectures that bridge the gap between memory bandwidth and processing power. The advent of deep neural networks has pushed traditional Von Neumann architectures to their limits due to the high latency and energy consumption costs associated with data movement between the processor and memory for these workloads. One of the solutions to overcome this bottleneck is to perform computation within the main memory through processing-in-memory (PIM), thereby limiting data movement and the costs associated with it. However, DRAM-based PIM struggles to achieve high throughput and energy efficiency due to internal data movement bottlenecks and the need for frequent refresh operations. In this work, we introduce OPIMA, a PIM-based ML accelerator, architected within an optical main memory. OPIMA has been designed to leverage the inherent massive parallelism within main memory while performing high-speed, low-energy optical computation to accelerate ML models based on convolutional neural networks. We present a comprehensive analysis of OPIMA to guide design choices and operational mechanisms. Additionally, we evaluate the performance and energy consumption of OPIMA, comparing it with conventional electronic computing systems and emerging photonic PIM architectures. The experimental results show that OPIMA can achieve 2.98x higher throughput and 137x better energy efficiency than the best-known prior work.

The rapid evolution of large language models (LLMs), driven by growing parameter scales, adoption of mixture-of-experts (MoE) architectures, and expanding context lengths, imposes unprecedented demands on AI infrastructure. Traditional AI clusters face limitations in compute intensity, memory bandwidth, inter-chip communication, and latency, compounded by variable workloads and strict service-level objectives. Addressing these issues requires fundamentally redesigned hardware-software integration. This paper introduces Huawei CloudMatrix, a next-generation AI datacenter architecture, realized in the production-grade CloudMatrix384 supernode. It integrates 384 Ascend 910C NPUs and 192 Kunpeng CPUs interconnected via an ultra-high-bandwidth Unified Bus (UB) network, enabling direct all-to-all communication and dynamic pooling of resources. These features optimize performance for communication-intensive operations, such as large-scale MoE expert parallelism and distributed key-value cache access. To fully leverage CloudMatrix384, we propose CloudMatrix-Infer, an advanced LLM serving solution incorporating three core innovations: a peer-to-peer serving architecture that independently scales prefill, decode, and caching; a large-scale expert parallelism strategy supporting EP320 via efficient UB-based token dispatch; and hardware-aware optimizations including specialized operators, microbatch-based pipelining, and INT8 quantization. Evaluation with the DeepSeek-R1 model shows CloudMatrix-Infer achieves state-of-the-art efficiency: prefill throughput of 6,688 tokens/s per NPU and decode throughput of 1,943 tokens/s per NPU (<50 ms TPOT). It effectively balances throughput and latency, sustaining 538 tokens/s even under stringent 15 ms latency constraints, while INT8 quantization maintains model accuracy across benchmarks.

Extreme edge platforms, such as in-vehicle smart devices, require efficient deployment of quantized deep neural networks (DNNs) to enable intelligent applications with limited amounts of energy, memory, and computing resources. However, many edge devices struggle to boost inference throughput of various quantized DNNs due to the varying quantization levels, and these devices lack floating-point (FP) support for on-device learning, which prevents them from improving model accuracy while ensuring data privacy. To tackle the challenges above, we propose a precision-scalable RISC-V DNN processor with on-device learning capability. It facilitates diverse precision levels of fixed-point DNN inference, spanning from 2-bit to 16-bit, and enhances on-device learning through improved support with FP16 operations. Moreover, we employ multiple methods such as FP16 multiplier reuse and multi-precision integer multiplier reuse, along with balanced mapping of FPGA resources, to significantly improve hardware resource utilization. Experimental results on the Xilinx ZCU102 FPGA show that our processor significantly improves inference throughput by 1.6sim14.6times and energy efficiency by 1.1sim14.6times across various DNNs, compared to the prior art, XpulpNN. Additionally, our processor achieves a 16.5times higher FP throughput for on-device learning.

Compared to traditional Vision Transformers (ViT), Mixture-of-Experts Vision Transformers (MoE-ViT) are introduced to scale model size without a proportional increase in computational complexity, making them a new research focus. Given the high performance and reconfigurability, FPGA-based accelerators for MoE-ViT emerge, delivering substantial gains over general-purpose processors. However, existing accelerators often fall short of fully exploring the design space, leading to suboptimal trade-offs between resource utilization and performance. To overcome this problem, we introduce UbiMoE, a novel end-to-end FPGA accelerator tailored for MoE-ViT. Leveraging the unique computational and memory access patterns of MoE-ViTs, we develop a latency-optimized streaming attention kernel and a resource-efficient reusable linear kernel, effectively balancing performance and resource consumption. To further enhance design efficiency, we propose a two-stage heuristic search algorithm that optimally tunes hardware parameters for various FPGA resource constraints. Compared to state-of-the-art (SOTA) FPGA designs, UbiMoE achieves 1.34x and 3.35x throughput improvements for MoE-ViT on Xilinx ZCU102 and Alveo U280 platforms, respectively, while enhancing energy efficiency by 1.75x and 1.54x. Our implementation is available at https://github.com/DJ000011/UbiMoE.

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/.

Large Language Model (LLM) inference is typically memory-intensive, especially when processing large batch sizes and long sequences, due to the large size of key-value (KV) cache. Vector Quantization (VQ) is recently adopted to alleviate this issue, but we find that the existing approach is susceptible to distribution shift due to its reliance on calibration datasets. To address this limitation, we introduce NSNQuant, a calibration-free Vector Quantization (VQ) technique designed for low-bit compression of the KV cache. By applying a three-step transformation-1) a token-wise normalization (Normalize), 2) a channel-wise centering (Shift), and 3) a second token-wise normalization (Normalize)-with Hadamard transform, NSNQuant effectively aligns the token distribution with the standard normal distribution. This alignment enables robust, calibration-free vector quantization using a single reusable codebook. Extensive experiments show that NSNQuant consistently outperforms prior methods in both 1-bit and 2-bit settings, offering strong generalization and up to 3times throughput gain over full-precision baselines.

The introduction of 8-bit floating-point (FP8) computation units in modern AI accelerators has generated significant interest in FP8-based large language model (LLM) inference. Unlike 16-bit floating-point formats, FP8 in deep learning requires a shared scaling factor. Additionally, while E4M3 and E5M2 are well-defined at the individual value level, their scaling and accumulation methods remain unspecified and vary across hardware and software implementations. As a result, FP8 behaves more like a quantization format than a standard numeric representation. In this work, we provide the first comprehensive analysis of FP8 computation and acceleration on two AI accelerators: the NVIDIA H100 and Intel Gaudi 2. Our findings highlight that the Gaudi 2, by leveraging FP8, achieves higher throughput-to-power efficiency during LLM inference, offering valuable insights into the practical implications of FP8 adoption for datacenter-scale LLM serving.

Decoder-only Transformer models such as GPT have demonstrated superior performance in text generation, by autoregressively predicting the next token. However, the performance of GPT is bounded by low compute-to-memory-ratio and high memory access. Throughput-oriented architectures such as GPUs target parallel processing rather than sequential token generation, and are not efficient for GPT acceleration, particularly on-device inference applications. Process-in-memory (PIM) architectures can significantly reduce data movement and provide high computation parallelism, and are promising candidates to accelerate GPT inference. In this work, we propose PIM-GPT that aims to achieve high throughput, high energy efficiency and end-to-end acceleration of GPT inference. PIM-GPT leverages DRAM-based PIM solutions to perform multiply-accumulate (MAC) operations on the DRAM chips, greatly reducing data movement. A compact application-specific integrated chip (ASIC) is designed and synthesized to initiate instructions to PIM chips and support data communication along with necessary arithmetic computations. At the software level, the mapping scheme is designed to maximize data locality and computation parallelism by partitioning a matrix among DRAM channels and banks to utilize all in-bank computation resources concurrently. We develop an event-driven clock-cycle accurate simulator to validate the efficacy of the proposed PIM-GPT architecture. Overall, PIM-GPT achieves 41-137times, 631-1074times speedup and 339-1085times, 890-1632times energy efficiency over GPU and CPU baseline, respectively, on 8 GPT models with up to 1.4 billion parameters.

Large Language Models (LLMs), constrained by their auto-regressive nature, suffer from slow decoding. Speculative decoding methods have emerged as a promising solution to accelerate LLM decoding, attracting attention from both systems and AI research communities. Recently, the pursuit of better draft quality has driven a trend toward parametrically larger draft models, which inevitably introduces substantial computational overhead. While existing work attempts to balance the trade-off between prediction accuracy and compute latency, we address this fundamental dilemma through architectural innovation. We propose PRISM, which disaggregates the computation of each predictive step across different parameter sets, refactoring the computational pathways of draft models to successfully decouple model capacity from inference cost. Through extensive experiments, we demonstrate that PRISM outperforms all existing draft architectures, achieving exceptional acceptance lengths while maintaining minimal draft latency for superior end-to-end speedup. We also re-examine scaling laws with PRISM, revealing that PRISM scales more effectively with expanding data volumes than other draft architectures. Through rigorous and fair comparison, we show that PRISM boosts the decoding throughput of an already highly optimized inference engine by more than 2.6x.

Analog In-memory Computing (IMC) has demonstrated energy-efficient and low latency implementation of convolution and fully-connected layers in deep neural networks (DNN) by using physics for computing in parallel resistive memory arrays. However, recurrent neural networks (RNN) that are widely used for speech-recognition and natural language processing have tasted limited success with this approach. This can be attributed to the significant time and energy penalties incurred in implementing nonlinear activation functions that are abundant in such models. In this work, we experimentally demonstrate the implementation of a non-linear activation function integrated with a ramp analog-to-digital conversion (ADC) at the periphery of the memory to improve in-memory implementation of RNNs. Our approach uses an extra column of memristors to produce an appropriately pre-distorted ramp voltage such that the comparator output directly approximates the desired nonlinear function. We experimentally demonstrate programming different nonlinear functions using a memristive array and simulate its incorporation in RNNs to solve keyword spotting and language modelling tasks. Compared to other approaches, we demonstrate manifold increase in area-efficiency, energy-efficiency and throughput due to the in-memory, programmable ramp generator that removes digital processing overhead.

Online inference is becoming a key service product for many businesses, deployed in cloud platforms to meet customer demands. Despite their revenue-generation capability, these services need to operate under tight Quality-of-Service (QoS) and cost budget constraints. This paper introduces KAIROS, a novel runtime framework that maximizes the query throughput while meeting QoS target and a cost budget. KAIROS designs and implements novel techniques to build a pool of heterogeneous compute hardware without online exploration overhead, and distribute inference queries optimally at runtime. Our evaluation using industry-grade deep learning (DL) models shows that KAIROS yields up to 2X the throughput of an optimal homogeneous solution, and outperforms state-of-the-art schemes by up to 70%, despite advantageous implementations of the competing schemes to ignore their exploration overhead.

The rapid advancement of large language models (LLMs) has been paralleled by unprecedented increases in computational demands, with training costs for state-of-the-art models doubling every few months. Training models directly in low-precision arithmetic offers a solution, by improving both computational throughput and energy efficiency. Specifically, NVIDIA's recent Blackwell architecture facilitates extremely low-precision operations, specifically FP4 variants, promising substantial efficiency gains. Yet, current algorithms for training LLMs in FP4 precision face significant accuracy degradation and often rely on mixed-precision fallbacks. In this paper, we systematically investigate hardware-supported FP4 training and introduce Quartet, a new approach enabling accurate, end-to-end FP4 training with all the major computations (in e.g. linear layers) being performed in low precision. Through extensive evaluations on Llama-type models, we reveal a new low-precision scaling law that quantifies performance trade-offs across varying bit-widths and allows us to identify a "near-optimal" low-precision training technique in terms of accuracy-vs-computation, called Quartet. We implement Quartet using optimized CUDA kernels tailored for NVIDIA Blackwell GPUs, and show that it can achieve state-of-the-art accuracy for FP4 precision, successfully training billion-scale models. Our method demonstrates that fully FP4-based training is a competitive alternative to standard-precision and FP8 training. Our code is available at https://github.com/IST-DASLab/Quartet.

Diffusion Language Models (DLMs) have recently emerged as a strong alternative to autoregressive language models (LMs). DLMs offer comparable accuracy with faster inference speed via parallel decoding. However, standard DLM decoding strategies relying on high-confidence tokens encounter an inherent information-theoretic bottleneck that restricts decoding progress and ultimately slows generation. We demonstrate both theoretically and empirically that prioritizing high-confidence tokens is inherently inefficient. High-probability tokens carry negligible information and strictly relying on them limits the effective progress made in each decoding round. We prove that the number of decoding rounds must grow linearly with the sample's total information (negative log-likelihood) and inversely with the per-round information budget, establishing a bits-to-rounds principle. We also propose Explore-Then-Exploit (ETE), a training-free decoding strategy that maximizes information throughput and decoding efficiency. ETE combines cross-block decoding with targeted exploration of high-uncertainty tokens to reshape the conditional distribution and trigger cascades of confident predictions. Experiments verify our theoretical bounds and demonstrate that ETE consistently reduces the required number of decoding rounds compared to confidence-only baselines without compromising generation quality.

There has been significant progress in developing neural network architectures that both achieve high predictive performance and that also achieve high application-level inference throughput (e.g., frames per second). Another metric of increasing importance is GPU utilization during inference: the measurement of how well a deployed neural network uses the computational capabilities of the GPU on which it runs. Achieving high GPU utilization is critical to increasing application-level throughput and ensuring a good return on investment for deploying GPUs. This paper analyzes the GPU utilization of convolutional neural network (CNN) inference. We first survey the GPU utilization of CNNs to show that there is room to improve the GPU utilization of many of these CNNs. We then investigate the GPU utilization of networks within a neural architecture search (NAS) search space, and explore how using GPU utilization as a metric could potentially be used to accelerate NAS itself. Our study makes the case that there is room to improve the inference-time GPU utilization of CNNs and that knowledge of GPU utilization has the potential to benefit even applications that do not target utilization itself. We hope that the results of this study will spur future innovation in designing GPU-efficient neural networks.

We introduce Motif-2-12.7B, a new open-weight foundation model that pushes the efficiency frontier of large language models by combining architectural innovation with system-level optimization. Designed for scalable language understanding and robust instruction generalization under constrained compute budgets, Motif-2-12.7B builds upon Motif-2.6B with the integration of Grouped Differential Attention (GDA), which improves representational efficiency by disentangling signal and noise-control attention pathways. The model is pre-trained on 5.5 trillion tokens spanning diverse linguistic, mathematical, scientific, and programming domains using a curriculum-driven data scheduler that gradually changes the data composition ratio. The training system leverages the MuonClip optimizer alongside custom high-performance kernels, including fused PolyNorm activations and the Parallel Muon algorithm, yielding significant throughput and memory efficiency gains in large-scale distributed environments. Post-training employs a three-stage supervised fine-tuning pipeline that successively enhances general instruction adherence, compositional understanding, and linguistic precision. Motif-2-12.7B demonstrates competitive performance across diverse benchmarks, showing that thoughtful architectural scaling and optimized training design can rival the capabilities of much larger models.

Large language models (LLMs) face low hardware efficiency during decoding, especially for long-context reasoning tasks. This paper introduces Step-3, a 321B-parameter VLM with hardware-aware model-system co-design optimized for minimizing decoding costs. Step-3 innovates in two key dimensions: (1) A novel Multi-Matrix Factorization Attention (MFA) mechanism that significantly reduces both KV cache size and computation while maintaining high attention expressiveness, and (2) Attention-FFN Disaggregation (AFD), a distributed inference system that decouples attention and Feed-Forward Network (FFN) layers into specialized subsystems. This co-design achieves unprecedented cost efficiency: Step-3 significantly reduces theoretical decoding costs compared with models like DeepSeek-V3 and Qwen3 MoE 235B, with the gains widening at longer context. Step-3 achieves low cost while activating 38B parameters per token (more than DeepSeek-V3 and Qwen3 MoE 235B), demonstrating that hardware-aligned attention arithmetic intensity, MoE sparsity, and AFD are critical to cost-effectiveness. We perform a head-to-head comparison with DeepSeek-V3 in its favorable scenarios. Our implementation on Hopper GPUs achieves a decoding throughput of up to 4,039 tokens per second per GPU under 50ms TPOT SLA (4K context, FP8, no MTP). It is higher than DeepSeek-V3's 2,324 in the same setup and sets a new Pareto frontier for LLM decoding.

We introduce LongCat-Flash, a 560-billion-parameter Mixture-of-Experts (MoE) language model designed for both computational efficiency and advanced agentic capabilities. Stemming from the need for scalable efficiency, LongCat-Flash adopts two novel designs: (a) Zero-computation Experts, which enables dynamic computational budget allocation and activates 18.6B-31.3B (27B on average) per token depending on contextual demands, optimizing resource usage. (b) Shortcut-connected MoE, which enlarges the computation-communication overlap window, demonstrating notable gains in inference efficiency and throughput compared to models of a comparable scale. We develop a comprehensive scaling framework for large models that combines hyperparameter transfer, model-growth initialization, a multi-pronged stability suite, and deterministic computation to achieve stable and reproducible training. Notably, leveraging the synergy among scalable architectural design and infrastructure efforts, we complete model training on more than 20 trillion tokens within 30 days, while achieving over 100 tokens per second (TPS) for inference at a cost of \$0.70 per million output tokens. To cultivate LongCat-Flash towards agentic intelligence, we conduct a large-scale pre-training on optimized mixtures, followed by targeted mid- and post-training on reasoning, code, and instructions, with further augmentation from synthetic data and tool use tasks. Comprehensive evaluations demonstrate that, as a non-thinking foundation model, LongCat-Flash delivers highly competitive performance among other leading models, with exceptional strengths in agentic tasks. The model checkpoint of LongCat-Flash is open-sourced to foster community research. LongCat Chat: https://longcat.ai Hugging Face: https://huggingface.co/meituan-longcat GitHub: https://github.com/meituan-longcat

The escalating demand for efficient decoding in large language models (LLMs) is particularly critical for reasoning-intensive architectures like OpenAI-o3 and DeepSeek-R1, which depend on extended chain-of-thought reasoning. This study investigates speculative decoding techniques through dense LLM architectures to establish foundational insights for accelerating reasoning tasks. While speculative decoding methods leveraging parallel draft-verification cycles have emerged as promising acceleration techniques, the scaling laws governing decoding efficiency remain under-explored compared to conventional backbone LLMs developed through Pretraining->SFT->RLHF training paradigms. In this work, we discover Log-linear Scaling Laws (Theorem 1.1, 1.2 and 1.3) governing draft model acceptance rate (or decoding speed) across three dimensions: pretraining token volume, draft model capacity, and decoding batch size. Building on these laws, we achieve Scylla, which coordinates multi-dimensional scaling for popular LLMs (Llama2/3, Qwen2.5). Empirical validation shows Scylla achieves 1.5-2.2 higher acceptance rate than EAGLE2 and 0.3 higher than EAGLE3 at temperature T = 0, with peak performance gains on summarization and QA tasks (Figure 2). Industrial inference engine deployments demonstrate 2X decoding throughput improvements over EAGLE2 (Table 5), validating the transformative potential of systematic scaling for efficient LLM inference. Code will be released later.

State space models (SSMs) like Mamba have recently attracted much attention. Compared to Transformer-based large language models (LLMs), Mamba achieves linear computation complexity with the sequence length and demonstrates superior performance. However, Mamba is hard to accelerate due to the scattered activation outliers and the complex computation dependency, rendering existing LLM accelerators inefficient. In this paper, we propose LightMamba that co-designs the quantization algorithm and FPGA accelerator architecture for efficient Mamba inference. We first propose an FPGA-friendly post-training quantization algorithm that features rotation-assisted quantization and power-of-two SSM quantization to reduce the majority of computation to 4-bit. We further design an FPGA accelerator that partially unrolls the Mamba computation to balance the efficiency and hardware costs. Through computation reordering as well as fine-grained tiling and fusion, the hardware utilization and memory efficiency of the accelerator get drastically improved. We implement LightMamba on Xilinx Versal VCK190 FPGA and achieve 4.65x to 6.06x higher energy efficiency over the GPU baseline. When evaluated on Alveo U280 FPGA, LightMamba reaches 93 tokens/s, which is 1.43x that of the GPU baseline.

We introduce the Llama-Nemotron series of models, an open family of heterogeneous reasoning models that deliver exceptional reasoning capabilities, inference efficiency, and an open license for enterprise use. The family comes in three sizes -- Nano (8B), Super (49B), and Ultra (253B) -- and performs competitively with state-of-the-art reasoning models such as DeepSeek-R1 while offering superior inference throughput and memory efficiency. In this report, we discuss the training procedure for these models, which entails using neural architecture search from Llama 3 models for accelerated inference, knowledge distillation, and continued pretraining, followed by a reasoning-focused post-training stage consisting of two main parts: supervised fine-tuning and large scale reinforcement learning. Llama-Nemotron models are the first open-source models to support a dynamic reasoning toggle, allowing users to switch between standard chat and reasoning modes during inference. To further support open research and facilitate model development, we provide the following resources: 1. We release the Llama-Nemotron reasoning models -- LN-Nano, LN-Super, and LN-Ultra -- under the commercially permissive NVIDIA Open Model License Agreement. 2. We release the complete post-training dataset: Llama-Nemotron-Post-Training-Dataset. 3. We also release our training codebases: NeMo, NeMo-Aligner, and Megatron-LM.

A simple design recipe for deep Transformers is to compose identical building blocks. But standard transformer blocks are far from simple, interweaving attention and MLP sub-blocks with skip connections & normalisation layers in precise arrangements. This complexity leads to brittle architectures, where seemingly minor changes can significantly reduce training speed, or render models untrainable. In this work, we ask to what extent the standard transformer block can be simplified? Combining signal propagation theory and empirical observations, we motivate modifications that allow many block components to be removed with no loss of training speed, including skip connections, projection or value parameters, sequential sub-blocks and normalisation layers. In experiments on both autoregressive decoder-only and BERT encoder-only models, our simplified transformers emulate the per-update training speed and performance of standard transformers, while enjoying 15% faster training throughput, and using 15% fewer parameters.

We introduce xLLM, an intelligent and efficient Large Language Model (LLM) inference framework designed for high-performance, large-scale enterprise-grade serving, with deep optimizations for diverse AI accelerators. To address these challenges, xLLM builds a novel decoupled service-engine architecture. At the service layer, xLLM-Service features an intelligent scheduling module that efficiently processes multimodal requests and co-locates online and offline tasks through unified elastic scheduling to maximize cluster utilization. This module also relies on a workload-adaptive dynamic Prefill-Decode (PD) disaggregation policy and a novel Encode-Prefill-Decode (EPD) disaggregation policy designed for multimodal inputs. Furthermore, it incorporates a distributed architecture to provide global KV Cache management and robust fault-tolerant capabilities for high availability. At the engine layer, xLLM-Engine co-optimizes system and algorithm designs to fully saturate computing resources. This is achieved through comprehensive multi-layer execution pipeline optimizations, an adaptive graph mode and an xTensor memory management. xLLM-Engine also further integrates algorithmic enhancements such as optimized speculative decoding and dynamic EPLB, collectively serving to substantially boost throughput and inference efficiency. Extensive evaluations demonstrate that xLLM delivers significantly superior performance and resource efficiency. Under identical TPOT constraints, xLLM achieves throughput up to 1.7x that of MindIE and 2.2x that of vLLM-Ascend with Qwen-series models, while maintaining an average throughput of 1.7x that of MindIE with Deepseek-series models. xLLM framework is publicly available at https://github.com/jd-opensource/xllm and https://github.com/jd-opensource/xllm-service.

Transformer-based large language models (LLMs) are now deployed to hundreds of millions of users. LLM inference is commonly performed on batches of sequences that share a prefix, such as few-shot examples or a chatbot system prompt. Decoding in this large-batch setting can be bottlenecked by the attention operation, which reads large key-value (KV) caches from memory and computes inefficient matrix-vector products for every sequence in the batch. In this work, we introduce Hydragen, a hardware-aware exact implementation of attention with shared prefixes. Hydragen computes attention over the shared prefix and unique suffixes separately. This decomposition enables efficient prefix attention by batching queries together across sequences, reducing redundant memory reads and enabling the use of hardware-friendly matrix multiplications. Our method can improve end-to-end LLM throughput by up to 32x against competitive baselines, with speedup growing with the batch size and shared prefix length. Hydragen also enables the use of very long shared contexts: with a high batch size, increasing the prefix length from 1K to 16K tokens decreases Hydragen throughput by less than 15%, while the throughput of baselines drops by over 90%. Hydragen generalizes beyond simple prefix-suffix decomposition and can be applied to tree-based prompt sharing patterns, allowing us to further reduce inference time on competitive programming problems by 55%.

Fine-tuning large language models (LLMs) with parameter-efficient techniques such as LoRA and QLoRA has enabled adaptation of foundation models on modest hardware. Yet the efficiency of such training on consumer-grade GPUs, especially under strict 8 GB VRAM limits, remains underexplored. We present a controlled profiling study of LoRA/QLoRA fine-tuning using the Qwen2.5-1.5B-Instruct model on a single NVIDIA RTX 4060. Across three representative configurations, we systematically vary batch size, sequence length, optimizer choice (AdamW vs. PagedAdamW), and precision (fp16 vs. bf16). We report throughput (tokens/s), time per 10k tokens, and VRAM footprint, alongside energy estimates derived from GPU board power limits. Our results show that paged optimizers improve throughput by up to 25% (628 tok/s vs. 500 tok/s baseline), while bf16 degrades efficiency relative to fp16. Despite 8 GB constraints, sequence lengths up to 2048 tokens were feasible using parameter-efficient strategies. To our knowledge, this is the first systematic case study of LLM fine- tuning efficiency on consumer GPUs, providing reproducible benchmarks and practical guidelines for resource-constrained researchers and practitioners.

To improve the efficiency of distributed large language model (LLM) inference, various parallelization strategies, such as tensor and pipeline parallelism, have been proposed. However, the distinct computational characteristics inherent in the two stages of LLM inference-prefilling and decoding-render a single static parallelization strategy insufficient for the effective optimization of both stages. In this work, we present Seesaw, an LLM inference engine optimized for throughput-oriented tasks. The key idea behind Seesaw is dynamic model re-sharding, a technique that facilitates the dynamic reconfiguration of parallelization strategies across stages, thereby maximizing throughput at both phases. To mitigate re-sharding overhead and optimize computational efficiency, we employ tiered KV cache buffering and transition-minimizing scheduling. These approaches work synergistically to reduce the overhead caused by frequent stage transitions while ensuring maximum batching efficiency. Our evaluation demonstrates that Seesaw achieves a throughput increase of up to 1.78x (1.36x on average) compared to vLLM, the most widely used state-of-the-art LLM inference engine.

Diffusion language models (DLMs) have emerged as a promising alternative to autoregressive (AR) generation, yet their reliance on Transformer backbones limits inference efficiency due to quadratic attention or KV-cache overhead. We introduce DiffuMamba, a masked diffusion language model built on a bidirectional Mamba backbone that combines the diffusion objective with linear-time sequence modeling, and DiffuMamba-H, a hybrid variant with interleaved attention. Across scales up to 1.3B parameters, our models match Transformer-based diffusion in downstream performance while achieving up to 8.2x and 4.3x higher inference throughput, respectively, on long sequences. We further present a systematic analysis of inference efficiency across modern DLM variants combining asymptotic complexity with empirical measurements. Notably, cache-efficient block diffusion with Mamba mixers emerges as the only strategy that scales linearly with sequence length and achieves the strongest performance across all baselines, suggesting a promising direction for future diffusion-based generation systems.

This work proposes XR-NPE, a high-throughput Mixed-precision SIMD Neural Processing Engine, designed for extended reality (XR) perception workloads like visual inertial odometry (VIO), object classification, and eye gaze extraction. XR-NPE is first to support FP4, Posit (4,1), Posit (8,0), and Posit (16,1) formats, with layer adaptive hybrid-algorithmic implementation supporting ultra-low bit precision to significantly reduce memory bandwidth requirements, and accompanied by quantization-aware training for minimal accuracy loss. The proposed Reconfigurable Mantissa Multiplication and Exponent processing Circuitry (RMMEC) reduces dark silicon in the SIMD MAC compute engine, assisted by selective power gating to reduce energy consumption, providing 2.85x improved arithmetic intensity. XR-NPE achieves a maximum operating frequency of 1.72 GHz, area 0.016 mm2 , and arithmetic intensity 14 pJ at CMOS 28nm, reducing 42% area, 38% power compared to the best of state-of-the-art MAC approaches. The proposed XR-NPE based AXI-enabled Matrix-multiplication co-processor consumes 1.4x fewer LUTs, 1.77x fewer FFs, and provides 1.2x better energy efficiency compared to SoTA accelerators on VCU129. The proposed co-processor provides 23% better energy efficiency and 4% better compute density for VIO workloads. XR-NPE establishes itself as a scalable, precision-adaptive compute engine for future resource-constrained XR devices. The complete set for codes for results reproducibility are released publicly, enabling designers and researchers to readily adopt and build upon them. https://github.com/mukullokhande99/XR-NPE.

Large Language Models (LLMs) have shown remarkable comprehension abilities but face challenges in GPU memory usage during inference, hindering their scalability for real-time applications like chatbots. To accelerate inference, we store computed keys and values (KV cache) in the GPU memory. Existing methods study the KV cache compression to reduce memory by pruning the pre-computed KV cache. However, they neglect the inter-layer dependency between layers and huge memory consumption in pre-computation. To explore these deficiencies, we find that the number of crucial keys and values that influence future generations decreases layer by layer and we can extract them by the consistency in attention weights. Based on the findings, we propose PyramidInfer, a method that compresses the KV cache by layer-wise retaining crucial context. PyramidInfer saves significant memory by computing fewer keys and values without sacrificing performance. Experimental results show PyramidInfer improves 2.2x throughput compared to Accelerate with over 54% GPU memory reduction in KV cache.

Large Language Models (LLMs) have driven significant progress, yet their growing parameter counts and context windows incur prohibitive compute, energy, and monetary costs. We introduce EfficientLLM, a novel benchmark and the first comprehensive empirical study evaluating efficiency techniques for LLMs at scale. Conducted on a production-class cluster (48xGH200, 8xH200 GPUs), our study systematically explores three key axes: (1) architecture pretraining (efficient attention variants: MQA, GQA, MLA, NSA; sparse Mixture-of-Experts (MoE)), (2) fine-tuning (parameter-efficient methods: LoRA, RSLoRA, DoRA), and (3) inference (quantization methods: int4, float16). We define six fine-grained metrics (Memory Utilization, Compute Utilization, Latency, Throughput, Energy Consumption, Compression Rate) to capture hardware saturation, latency-throughput balance, and carbon cost. Evaluating over 100 model-technique pairs (0.5B-72B parameters), we derive three core insights: (i) Efficiency involves quantifiable trade-offs: no single method is universally optimal; e.g., MoE reduces FLOPs and improves accuracy but increases VRAM by 40%, while int4 quantization cuts memory/energy by up to 3.9x at a 3-5% accuracy drop. (ii) Optima are task- and scale-dependent: MQA offers optimal memory-latency trade-offs for constrained devices, MLA achieves lowest perplexity for quality-critical tasks, and RSLoRA surpasses LoRA efficiency only beyond 14B parameters. (iii) Techniques generalize across modalities: we extend evaluations to Large Vision Models (Stable Diffusion 3.5, Wan 2.1) and Vision-Language Models (Qwen2.5-VL), confirming effective transferability. By open-sourcing datasets, evaluation pipelines, and leaderboards, EfficientLLM provides essential guidance for researchers and engineers navigating the efficiency-performance landscape of next-generation foundation models.

Serving numerous users and requests concurrently requires good fairness in Large Language Models (LLMs) serving system. This ensures that, at the same cost, the system can meet the Service Level Objectives (SLOs) of more users , such as time to first token (TTFT) and time between tokens (TBT), rather than allowing a few users to experience performance far exceeding the SLOs. To achieve better fairness, the preemption-based scheduling policy dynamically adjusts the priority of each request to maintain balance during runtime. However, existing systems tend to overly prioritize throughput, overlooking the overhead caused by preemption-induced context switching, which is crucial for maintaining fairness through priority adjustments. In this work, we identify three main challenges that result in this overhead. 1) Inadequate I/O utilization. 2) GPU idleness. 3) Unnecessary I/O transmission during multi-turn conversations. Our key insight is that the block-based KV cache memory policy in existing systems, while achieving near-zero memory waste, leads to discontinuity and insufficient granularity in the KV cache memory. To respond, we introduce FastSwitch, a fairness-aware serving system that not only aligns with existing KV cache memory allocation policy but also mitigates context switching overhead. Our evaluation shows that FastSwitch outperforms the state-of-the-art LLM serving system vLLM with speedups of 1.4-11.2x across different tail TTFT and TBT.

Recent work has shown that attention-based language models excel at recall, the ability to ground generations in tokens previously seen in context. However, the efficiency of attention-based models is bottle-necked during inference by the KV-cache's aggressive memory consumption. In this work, we explore whether we can improve language model efficiency (e.g. by reducing memory consumption) without compromising on recall. By applying experiments and theory to a broad set of architectures, we identify a key tradeoff between a model's state size and recall ability. We show that efficient alternatives to attention (e.g. H3, Mamba, RWKV) maintain a fixed-size recurrent state, but struggle at recall. We propose BASED a simple architecture combining linear and sliding window attention. By varying BASED window size and linear attention feature dimension, we can dial the state size and traverse the pareto frontier of the recall-memory tradeoff curve, recovering the full quality of attention on one end and the small state size of attention-alternatives on the other. We train language models up to 1.3b parameters and show that BASED matches the strongest sub-quadratic models (e.g. Mamba) in perplexity and outperforms them on real-world recall-intensive tasks by 6.22 accuracy points. Implementations of linear attention are often less efficient than optimized standard attention implementations. To make BASED competitive, we develop IO-aware algorithms that enable 24x higher throughput on language generation than FlashAttention-2, when generating 1024 tokens using 1.3b parameter models. Code for this work is provided at: https://github.com/HazyResearch/based.

Large Language Models (LLMs) have achieved state-of-the-art performance across various language tasks but pose challenges for practical deployment due to their substantial memory requirements. Furthermore, the latest generative models suffer from high inference costs caused by the memory bandwidth bottleneck in the auto-regressive decoding process. To address these issues, we propose an efficient weight-only quantization method that reduces memory consumption and accelerates inference for LLMs. To ensure minimal quality degradation, we introduce a simple and effective heuristic approach that utilizes only the model weights of a pre-trained model. This approach is applicable to both Mixture-of-Experts (MoE) and dense models without requiring additional fine-tuning. To demonstrate the effectiveness of our proposed method, we first analyze the challenges and issues associated with LLM quantization. Subsequently, we present our heuristic approach, which adaptively finds the granularity of quantization, effectively addressing these problems. Furthermore, we implement highly efficient GPU GEMMs that perform on-the-fly matrix multiplication and dequantization, supporting the multiplication of fp16 or bf16 activations with int8 or int4 weights. We evaluate our approach on large-scale open source models such as OPT-175B and internal MoE models, showcasing minimal accuracy loss while achieving up to 3.65 times higher throughput on the same number of GPUs.

Batch inference for agentic workloads stresses the GPU key-value (KV) cache in a sustained and cumulative manner, often causing severe throughput degradation well before memory capacity is exhausted. We identify this phenomenon as middle-phase thrashing, a previously under-characterized pathology in which cache efficiency collapses as long-lived agents accumulate state over time. We argue that mitigating this pathology requires moving beyond reactive, request-level cache management to proactive, agent-level admission control. Drawing inspiration from congestion control in distributed systems, we view the KV cache as a shared resource whose efficient utilization depends on feedback-driven regulation. Based on this insight, we present CONCUR, a lightweight control layer that regulates agent admission to bound aggregate cache pressure while preserving execution continuity. CONCUR adapts a cache-aware control algorithm to dynamically adjust the number of active agents using runtime cache signals. Across large models and real-world agent workloads, CONCUR prevents middle-phase thrashing and improves batch inference throughput by up to 4.09x on Qwen3-32B and 1.9x on DeepSeek-V3, while remaining compatible with existing LLM serving systems.

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.

Many LLM tasks are performed in large batches or even offline, and the performance indictor for which is throughput. These tasks usually show the characteristic of prefix sharing, where different prompt input can partially show the common prefix. However, the existing LLM inference engines tend to optimize the streaming requests and show limitations of supporting the large batched tasks with the prefix sharing characteristic. The existing solutions use the LRU-based cache to reuse the KV context of common prefix. The KV context that is about to be reused may prematurely be evicted with the implicit cache management. Even if not evicted, the lifetime of the shared KV context is extended since requests sharing the same context are not scheduled together, resulting in larger memory usage. These streaming oriented systems schedule the requests in the first-come-first-serve or similar order. As a result, the requests with larger ratio of decoding steps may be scheduled too late to be able to mix with the prefill chunks to increase the hardware utilization. Besides, the token and request number based batching can limit the size of token-batch, which keeps the GPU from saturating for the iterations dominated by decoding tokens. We propose BatchLLM to address the above problems. BatchLLM explicitly identifies the common prefixes globally. The requests sharing the same prefix will be scheduled together to reuse the KV context the best, which also shrinks the lifetime of common KV memory. BatchLLM reorders the requests and schedules the requests with larger ratio of decoding first to better mix the decoding tokens with the latter prefill chunks and applies memory-centric token batching to enlarge the token-batch sizes, which helps to increase the GPU utilization. Extensive evaluation shows that BatchLLM outperforms vLLM by 1.1x to 2x on a set of microbenchmarks and two typical industry workloads.

With the ubiquitous use of modern large language models (LLMs) across industries, the inference serving for these models is ever expanding. Given the high compute and memory requirements of modern LLMs, more and more top-of-the-line GPUs are being deployed to serve these models. Energy availability has come to the forefront as the biggest challenge for data center expansion to serve these models. In this paper, we present the trade-offs brought up by making energy efficiency the primary goal of LLM serving under performance SLOs. We show that depending on the inputs, the model, and the service-level agreements, there are several knobs available to the LLM inference provider to use for being energy efficient. We characterize the impact of these knobs on the latency, throughput, as well as the energy. By exploring these trade-offs, we offer valuable insights into optimizing energy usage without compromising on performance, thereby paving the way for sustainable and cost-effective LLM deployment in data center environments.

Two critical limitations of organic-inorganic lead halide perovskite materials for solar cells are their poor stability in humid environments and inclusion of toxic lead. In this study, high-throughput density functional theory (DFT) methods are used to computationally model and screen 1845 halide perovskites in search of new materials without these limitations that are promising for solar cell applications. This study focuses on finding materials that are comprised of nontoxic elements, stable in a humid operating environment, and have an optimal bandgap for one of single junction, tandem Si-perovskite, or quantum dot-based solar cells. Single junction materials are also screened on predicted single junction photovoltaic (PV) efficiencies exceeding 22.7%, which is the current highest reported PV efficiency for halide perovskites. Generally, these methods qualitatively reproduce the properties of known promising nontoxic halide perovskites that have either been experimentally evaluated or predicted from theory. From a set of 1845 materials, 15 materials pass all screening criteria for single junction cell applications, 13 of which have not been previously investigated, such as (CH3NH3)0.75Cs0.25SnI3, ((NH2)2CH)Ag0.5Sb0.5Br3, CsMn0.875Fe0.125I3, ((CH3)2NH2)Ag0.5Bi0.5I3, and ((NH2)2CH)0.5Rb0.5SnI3. These materials, together with others predicted in this study, may be promising candidate materials for stable, highly efficient, and non-toxic perovskite-based solar cells.

Large Language Model (LLM)-based search agents have shown remarkable capabilities in solving complex tasks by dynamically decomposing problems and addressing them through interleaved reasoning and retrieval. However, this interleaved paradigm introduces substantial efficiency bottlenecks. First, we observe that both highly accurate and overly approximate retrieval methods degrade system efficiency: exact search incurs significant retrieval overhead, while coarse retrieval requires additional reasoning steps during generation. Second, we identify inefficiencies in system design, including improper scheduling and frequent retrieval stalls, which lead to cascading latency -- where even minor delays in retrieval amplify end-to-end inference time. To address these challenges, we introduce SearchAgent-X, a high-efficiency inference framework for LLM-based search agents. SearchAgent-X leverages high-recall approximate retrieval and incorporates two key techniques: priority-aware scheduling and non-stall retrieval. Extensive experiments demonstrate that SearchAgent-X consistently outperforms state-of-the-art systems such as vLLM and HNSW-based retrieval across diverse tasks, achieving up to 3.4times higher throughput and 5times lower latency, without compromising generation quality. SearchAgent-X is available at https://github.com/tiannuo-yang/SearchAgent-X.

Large language models (LLMs) have surged in popularity and are extensively used in commercial applications, where the efficiency of model serving is crucial for the user experience. Most current research focuses on optimizing individual sub-procedures, e.g. local inference and communication, however, there is no comprehensive framework that provides a holistic system view for optimizing LLM serving in an end-to-end manner. In this work, we conduct a detailed analysis to identify major bottlenecks that impact end-to-end latency in LLM serving systems. Our analysis reveals that a comprehensive LLM serving endpoint must address a series of efficiency bottlenecks that extend beyond LLM inference. We then propose ScaleLLM, an optimized system for resource-efficient LLM serving. Our extensive experiments reveal that with 64 concurrent requests, ScaleLLM achieves a 4.3x speed up over vLLM and outperforms state-of-the-arts with 1.5x higher throughput.

The dominance of large decoder-only language models has overshadowed encoder-decoder architectures, despite their fundamental efficiency advantages in sequence processing. For small language models (SLMs) - those with 1 billion parameters or fewer - our systematic analysis across GPU, CPU, and NPU platforms reveals that encoder-decoder architectures achieve 47% lower first-token latency and 4.7x higher throughput compared to decoder-only models on edge devices. These gains may be attributed to encoder-decoder's one-time input processing and efficient separation of understanding and generation phases. We introduce a novel knowledge distillation framework that enables encoder-decoder models to leverage capabilities from large scalable decoder-only teachers while preserving their architectural advantages, achieving up to 6 average performance points improvement across diverse tasks, with significant gains in asymmetric sequence tasks where input and output distributions can benefit from different processing approaches. When combined with modern advances like Rotary Positional Embeddings (RoPE) and Vision encoders, our systematic investigation demonstrates that encoder-decoder architectures provide a more practical path toward deploying capable language models in resource-constrained environments. Our findings challenge the prevailing trend toward decoder-only scaling, showing that architectural choices become increasingly crucial as parameter budgets decrease, particularly for on-device and edge deployments where computational efficiency is paramount.

Scaling distributed training of Large Language Models (LLMs) requires not only algorithmic advances but also efficient utilization of heterogeneous hardware resources. While existing methods such as DiLoCo have demonstrated promising results, they often fail to fully exploit computational clusters under dynamic workloads. To address this limitation, we propose a three-stage method that combines Multi-Instance Training (MIT), Adaptive Batched DiLoCo, and switch mode mechanism. MIT allows individual nodes to run multiple lightweight training streams with different model instances in parallel and merge them to combine knowledge, increasing throughput and reducing idle time. Adaptive Batched DiLoCo dynamically adjusts local batch sizes to balance computation and communication, substantially lowering synchronization delays. Switch mode further stabilizes training by seamlessly introducing gradient accumulation once adaptive batch sizes grow beyond hardware-friendly limits. Together, these innovations improve both convergence speed and system efficiency. We also provide a theoretical estimate of the number of communications required for the full convergence of a model trained using our method.

State-of-the-art LLMs often rely on scale with high computational costs, which has sparked a research agenda to reduce parameter counts and costs without significantly impacting performance. Our study focuses on Transformer-based LLMs, specifically applying low-rank parametrization to the computationally intensive feedforward networks (FFNs), which are less studied than attention blocks. In contrast to previous works, (i) we explore low-rank parametrization at scale, up to 1.3B parameters; (ii) within Transformer language models rather than convolutional architectures; and (iii) starting from training from scratch. Experiments on the large RefinedWeb dataset show that low-rank parametrization is both efficient (e.g., 2.6times FFN speed-up with 32\% parameters) and effective during training. Interestingly, these structured FFNs exhibit steeper scaling curves than the original models. Motivated by this finding, we develop the wide and structured networks surpassing the current medium-sized and large-sized Transformer in perplexity and throughput performance. Our code is available at https://github.com/CLAIRE-Labo/StructuredFFN/tree/main.

Large-scale language models have become increasingly challenging and expensive to train. Among various methods addressing this issue, Pipeline Parallelism has been widely employed to accommodate massive model weights within limited GPU memory. This paper introduces Hanayo, a wave-like pipeline parallelism strategy that boasts a concise structure and practical applicability, alongside a high-performance pipeline execution runtime to tackle the challenges of pipeline strategy implementation. Hanayo mitigates the issues of pipeline bubbles and excessive memory consumption prevalent in existing schemes, without resorting to model duplicates as in Chimera. Our evaluation, conducted on four distinct computing clusters and involving both GPT-like and BERT-like architectures with up to 32 GPUs, demonstrates up to a 30.4 \% increase in throughput compared to the state-of-the-art approach.

The widespread adoption of large language models such as ChatGPT and Bard has led to unprecedented demand for these technologies. The burgeoning cost of inference for ever-increasing model sizes coupled with hardware shortages has limited affordable access and poses a pressing need for efficiency approaches geared towards high throughput and performance. Multi-input multi-output (MIMO) algorithms such as data multiplexing, offer a promising solution with a many-fold increase in throughput by performing inference for multiple inputs at the cost of a single input. Yet these approaches are not currently performant enough to be deployed in modern systems. We change that by developing MUX-PLMs, a class of high throughput pre-trained language models (PLMs) trained with data multiplexing, that can be fine-tuned for any downstream task to yield high-throughput high-performance. Our novel multiplexing and demultiplexing modules proficiently entangle and disentangle inputs, and enable high-performance high throughput that are competitive with vanilla PLMs while achieving 2x/5x inference speedup with only a 1-4% drop on a broad suite of tasks.

During inference for transformer-based large language models (LLM), prefilling is the computation of the key-value (KV) cache for input tokens in the prompt prior to autoregressive generation. For longer input prompt lengths, prefilling will incur a significant overhead on decoding time. In this work, we highlight the following pitfall of prefilling: for batches containing high-varying prompt lengths, significant computation is wasted by the standard practice of padding sequences to the maximum length. As LLMs increasingly support longer context lengths, potentially up to 10 million tokens, variations in prompt lengths within a batch become more pronounced. To address this, we propose Prepacking, a simple yet effective method to optimize prefilling computation. To avoid redundant computation on pad tokens, prepacking combines prompts of varying lengths into a sequence and packs multiple sequences into a compact batch using a bin-packing algorithm. It then modifies the attention mask and positional encoding to compute multiple prefilled KV-caches for multiple prompts within a single sequence. On standard curated dataset containing prompts with varying lengths, we obtain a significant speed and memory efficiency improvements as compared to the default padding-based prefilling computation within Huggingface across a range of base model configurations and inference serving scenarios.

Large language models (LLMs) are widely applied in chatbots, code generators, and search engines. Workloads such as chain-of-thought, complex reasoning, and agent services significantly increase the inference cost by invoking the model repeatedly. Optimization methods such as parallelism, compression, and caching have been adopted to reduce costs, but the diverse service requirements make it hard to select the right method. Recently, specialized LLM inference engines have emerged as a key component for integrating the optimization methods into service-oriented infrastructures. However, a systematic study on inference engines is still lacking. This paper provides a comprehensive evaluation of 25 open-source and commercial inference engines. We examine each inference engine in terms of ease-of-use, ease-of-deployment, general-purpose support, scalability, and suitability for throughput- and latency-aware computation. Furthermore, we explore the design goals of each inference engine by investigating the optimization techniques it supports. In addition, we assess the ecosystem maturity of open source inference engines and handle the performance and cost policy of commercial solutions. We outline future research directions that include support for complex LLM-based services, support of various hardware, and enhanced security, offering practical guidance to researchers and developers in selecting and designing optimized LLM inference engines. We also provide a public repository to continually track developments in this fast-evolving field: https://github.com/sihyeong/Awesome-LLM-Inference-Engine

This paper presents a comprehensive analysis of quantization techniques for optimizing Large Language Models (LLMs), specifically focusing on Post-Training Quantization (PTQ) and Quantization-Aware Training (QAT). Through empirical evaluation across models ranging from 10M to 1B parameters, we demonstrate that quantization can achieve up to 68% reduction in model size while maintaining performance within 6% of full-precision baselines when utilizing our proposed scaling factor {\gamma}. Our experiments show that INT8 quantization delivers a 40% reduction in computational cost and power consumption, while INT4 quantization further improves these metrics by 60%. We introduce a novel theoretical framework for mixed-precision quantization, deriving optimal bit allocation strategies based on layer sensitivity and weight variance. Hardware efficiency evaluations on edge devices reveal that our quantization approach enables up to 2.4x throughput improvement for INT8 and 3x for INT4, with 60% power reduction compared to full-precision models.

Charts are important for presenting and explaining complex data relationships. Recently, multimodal large language models (MLLMs) have shown remarkable capabilities in various chart understanding tasks. However, the sheer size of these models in terms of parameters and computational requirements limits their use in resource-constrained environments. In this paper, we present TinyChart, an efficient MLLM for chart understanding with only 3B parameters. TinyChart overcomes two key challenges in efficient chart understanding: (1) reduce the burden of learning numerical computations through a Program-of-Thoughts (PoT) learning strategy, which trains the model to generate Python programs for numerical calculations, and (2) reduce lengthy vision feature sequences produced by the vision transformer for high-resolution images through a Vision Token Merging module, which gradually merges most similar vision tokens. Extensive experiments demonstrate that our 3B TinyChart achieves SOTA performance on a variety of chart understanding benchmarks including ChartQA, Chart-to-Text, Chart-to-Table, OpenCQA, and ChartX. It outperforms several chart understanding MLLM with up to 13B parameters such as ChartLlama and ChartAst, and close-sourced general-purpose MLLM GPT-4V on ChartQA. It also demonstrates its superior efficiency with higher throughput during inference due to a smaller model scale and more efficient vision encoding. Our code and model are available at https://github.com/X-PLUG/mPLUG-DocOwl/tree/main/TinyChart.

AlphaFold has transformed protein structure prediction, but emerging applications such as virtual ligand screening, proteome-wide folding, and de novo binder design demand predictions at a massive scale, where runtime and memory costs become prohibitive. A major bottleneck lies in the Pairformer backbone of AlphaFold3-style models, which relies on computationally expensive triangular primitives-especially triangle attention-for pairwise reasoning. We introduce Pairmixer, a streamlined alternative that eliminates triangle attention while preserving higher-order geometric reasoning capabilities that are critical for structure prediction. Pairmixer substantially improves computational efficiency, matching state-of-the-art structure predictors across folding and docking benchmarks, delivering up to 4x faster inference on long sequences while reducing training cost by 34%. Its efficiency alleviates the computational burden of downstream applications such as modeling large protein complexes, high-throughput ligand and binder screening, and hallucination-based design. Within BoltzDesign, for example, Pairmixer delivers over 2x faster sampling and scales to sequences ~30% longer than the memory limits of Pairformer.

We introduce LongCat-Flash-Omni, a state-of-the-art open-source omni-modal model with 560 billion parameters, excelling at real-time audio-visual interaction. By adopting a curriculum-inspired progressive training strategy that transitions from simpler to increasingly complex modality sequence modeling tasks, LongCat-Flash-Omni attains comprehensive multimodal capabilities while maintaining strong unimodal capability. Building upon LongCat-Flash, which adopts a high-performance Shortcut-connected Mixture-of-Experts (MoE) architecture with zero-computation experts, LongCat-Flash-Omni integrates efficient multimodal perception and speech reconstruction modules. Despite its immense size of 560B parameters (with 27B activated), LongCat-Flash-Omni achieves low-latency real-time audio-visual interaction. For training infrastructure, we developed a modality-decoupled parallelism scheme specifically designed to manage the data and model heterogeneity inherent in large-scale multimodal training. This innovative approach demonstrates exceptional efficiency by sustaining over 90% of the throughput achieved by text-only training. Extensive evaluations show that LongCat-Flash-Omni achieves state-of-the-art performance on omni-modal benchmarks among open-source models. Furthermore, it delivers highly competitive results across a wide range of modality-specific tasks, including text, image, and video understanding, as well as audio understanding and generation. We provide a comprehensive overview of the model architecture design, training procedures, and data strategies, and open-source the model to foster future research and development in the community.

Large language model (LLM) applications are evolving beyond simple chatbots into dynamic, general-purpose agentic programs, which scale LLM calls and output tokens to help AI agents reason, explore, and solve complex tasks. However, existing LLM serving systems ignore dependencies between programs and calls, missing significant opportunities for optimization. Our analysis reveals that programs submitted to LLM serving engines experience long cumulative wait times, primarily due to head-of-line blocking at both the individual LLM request and the program. To address this, we introduce Autellix, an LLM serving system that treats programs as first-class citizens to minimize their end-to-end latencies. Autellix intercepts LLM calls submitted by programs, enriching schedulers with program-level context. We propose two scheduling algorithms-for single-threaded and distributed programs-that preempt and prioritize LLM calls based on their programs' previously completed calls. Our evaluation demonstrates that across diverse LLMs and agentic workloads, Autellix improves throughput of programs by 4-15x at the same latency compared to state-of-the-art systems, such as vLLM.

Diffusion Large Language Models (dLLMs) enable breakthroughs in reasoning and parallel decoding but suffer from prohibitive quadratic computational complexity and memory overhead during inference. Current caching techniques accelerate decoding by storing full-layer states, yet impose substantial memory usage that limit long-context applications. Our analysis of attention patterns in dLLMs reveals persistent cross-layer sparsity, with pivotal tokens remaining salient across decoding steps and low-relevance tokens staying unimportant, motivating selective cache eviction. We propose Sparse-dLLM, the first training-free framework integrating dynamic cache eviction with sparse attention via delayed bidirectional sparse caching. By leveraging the stability of token saliency over steps, it retains critical tokens and dynamically evicts unimportant prefix/suffix entries using an attention-guided strategy. Extensive experiments on LLaDA and Dream series demonstrate Sparse-dLLM achieves up to 10times higher throughput than vanilla dLLMs, with comparable performance and similar peak memory costs, outperforming previous methods in efficiency and effectiveness.

High-quality main content extraction from web pages is a critical prerequisite for constructing large-scale training corpora. While traditional heuristic extractors are efficient, they lack the semantic reasoning required to handle the structural heterogeneity of the modern web. Conversely, well-pretrained generative Large Language Models (LLMs) offer superior document comprehension but are prohibited by excessive computational costs, limited context windows, and hallucination risks when applied at web scale. We present Dripper, a lightweight framework that resolves these bottlenecks through four contributions: (1) We reformulate extraction as a constrained sequence labeling task using SLMs (Small Language Models). This paradigm eliminates generative hallucinations and achieves exceptional efficiency, reaching a throughput of 3.08 pages per second on a single A100 GPU. (2) We construct WebMainBench, a rigorous benchmark of 7,809 human-annotated pages covering 5,434 unique domains and multiple languages. Evaluations show our Dripper-0.6B model outperforms heuristics like Trafilatura and rivals massive models like DeepSeek-V3.2(685B), GPT-5 and Gemini-2.5-Pro, offering an optimal efficiency-accuracy trade-off. (3) We demonstrate infrastructural value by pre-training a 1B model on a Dripper-curated corpus (63B tokens). This model significantly outperforms baselines in downstream tasks, proving the critical role of extraction quality and the effectiveness of our framework. (4) We open-source the Dripper-0.6B weights and codebase to facilitate the construction of high-quality datasets.

We introduce AIBrix, a cloud-native, open-source framework designed to optimize and simplify large-scale LLM deployment in cloud environments. Unlike traditional cloud-native stacks, AIBrix follows a co-design philosophy, ensuring every layer of the infrastructure is purpose-built for seamless integration with inference engines like vLLM. AIBrix introduces several key innovations to reduce inference costs and enhance performance including high-density LoRA management for dynamic adapter scheduling, LLM-specific autoscalers, and prefix-aware, load-aware routing. To further improve efficiency, AIBrix incorporates a distributed KV cache, boosting token reuse across nodes, leading to a 50% increase in throughput and a 70% reduction in inference latency. AIBrix also supports unified AI runtime which streamlines model management while maintaining vendor-agnostic engine compatibility. For large-scale multi-node inference, AIBrix employs hybrid orchestration -- leveraging Kubernetes for coarse-grained scheduling and Ray for fine-grained execution -- to balance efficiency and flexibility. Additionally, an SLO-driven GPU optimizer dynamically adjusts resource allocations, optimizing heterogeneous serving to maximize cost efficiency while maintaining service guarantees. Finally, AIBrix enhances system reliability with AI accelerator diagnostic tools, enabling automated failure detection and mock-up testing to improve fault resilience. AIBrix is available at https://github.com/vllm-project/aibrix.

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.

Quantization can accelerate large language model (LLM) inference. Going beyond INT8 quantization, the research community is actively exploring even lower precision, such as INT4. Nonetheless, state-of-the-art INT4 quantization techniques only accelerate low-batch, edge LLM inference, failing to deliver performance gains in large-batch, cloud-based LLM serving. We uncover a critical issue: existing INT4 quantization methods suffer from significant runtime overhead (20-90%) when dequantizing either weights or partial sums on GPUs. To address this challenge, we introduce QoQ, a W4A8KV4 quantization algorithm with 4-bit weight, 8-bit activation, and 4-bit KV cache. QoQ stands for quattuor-octo-quattuor, which represents 4-8-4 in Latin. QoQ is implemented by the QServe inference library that achieves measured speedup. The key insight driving QServe is that the efficiency of LLM serving on GPUs is critically influenced by operations on low-throughput CUDA cores. Building upon this insight, in QoQ algorithm, we introduce progressive quantization that can allow low dequantization overhead in W4A8 GEMM. Additionally, we develop SmoothAttention to effectively mitigate the accuracy degradation incurred by 4-bit KV quantization. In the QServe system, we perform compute-aware weight reordering and take advantage of register-level parallelism to reduce dequantization latency. We also make fused attention memory-bound, harnessing the performance gain brought by KV4 quantization. As a result, QServe improves the maximum achievable serving throughput of Llama-3-8B by 1.2x on A100, 1.4x on L40S; and Qwen1.5-72B by 2.4x on A100, 3.5x on L40S, compared to TensorRT-LLM. Remarkably, QServe on L40S GPU can achieve even higher throughput than TensorRT-LLM on A100. Thus, QServe effectively reduces the dollar cost of LLM serving by 3x. Code is available at https://github.com/mit-han-lab/qserve.

Practical large language model (LLM) services may involve a long system prompt, which specifies the instructions, examples, and knowledge documents of the task and is reused across numerous requests. However, the long system prompt causes throughput/latency bottlenecks as the cost of generating the next token grows w.r.t. the sequence length. This paper aims to improve the efficiency of LLM services that involve long system prompts. Our key observation is that handling these system prompts requires heavily redundant memory accesses in existing causal attention computation algorithms. Specifically, for batched requests, the cached hidden states (i.e., key-value pairs) of system prompts are transferred from off-chip DRAM to on-chip SRAM multiple times, each corresponding to an individual request. To eliminate such a redundancy, we propose RelayAttention, an attention algorithm that allows reading these hidden states from DRAM exactly once for a batch of input tokens. RelayAttention is a free lunch: it maintains the generation quality while requiring no model retraining, as it is based on a mathematical reformulation of causal attention.

Transformer-based language models have become the standard approach to solving natural language processing tasks. However, industry adoption usually requires the maximum throughput to comply with certain latency constraints that prevents Transformer models from being used in production. To address this gap, model compression techniques such as quantization and pruning may be used to improve inference efficiency. However, these compression techniques require specialized software to apply and deploy at scale. In this work, we propose a new pipeline for creating and running Fast Transformer models on CPUs, utilizing hardware-aware pruning, knowledge distillation, quantization, and our own Transformer inference runtime engine with optimized kernels for sparse and quantized operators. We demonstrate the efficiency of our pipeline by creating a Fast DistilBERT model showing minimal accuracy loss on the question-answering SQuADv1.1 benchmark, and throughput results under typical production constraints and environments. Our results outperform existing state-of-the-art Neural Magic's DeepSparse runtime performance by up to 50% and up to 4.1x performance speedup over ONNX Runtime. Source code is publicly available at https://github.com/intel/intel-extension-for-transformers.

Mixture of Experts (MoE) models have emerged as the de facto architecture for scaling up language models without significantly increasing the computational cost. Recent MoE models demonstrate a clear trend towards high expert granularity (smaller expert intermediate dimension) and higher sparsity (constant number of activated experts with higher number of total experts), which improve model quality per FLOP. However, fine-grained MoEs suffer from increased activation memory footprint and reduced hardware efficiency due to higher IO costs, while sparser MoEs suffer from wasted computations due to padding in Grouped GEMM kernels. In response, we propose a memory-efficient algorithm to compute the forward and backward passes of MoEs with minimal activation caching for the backward pass. We also design GPU kernels that overlap memory IO with computation benefiting all MoE architectures. Finally, we propose a novel "token rounding" method that minimizes the wasted compute due to padding in Grouped GEMM kernels. As a result, our method SonicMoE reduces activation memory by 45% and achieves a 1.86x compute throughput improvement on Hopper GPUs compared to ScatterMoE's BF16 MoE kernel for a fine-grained 7B MoE. Concretely, SonicMoE on 64 H100s achieves a training throughput of 213 billion tokens per day comparable to ScatterMoE's 225 billion tokens per day on 96 H100s for a 7B MoE model training with FSDP-2 using the lm-engine codebase. Under high MoE sparsity settings, our tile-aware token rounding algorithm yields an additional 1.16x speedup on kernel execution time compared to vanilla top-K routing while maintaining similar downstream performance. We open-source all our kernels to enable faster MoE model training.

Multimodal Large Language Models (MLLMs) have achieved remarkable success across various applications, yet their computational overhead during deployment remains a critical challenge. While Key-Value (KV) caching improves inference efficiency by trading memory for computation, the growing memory footprint from storing extensive KV caches reduces throughput and limits long-term execution on devices with constrained GPU memory. Existing approaches primarily focus on dropping unimportant tokens to reduce the KV cache size, mitigating memory constraints at the cost of potential information loss. In contrast, we propose a simple yet effective visual quantization strategy that preserves all visual tokens while significantly reducing memory consumption. To achieve an extreme quantization ratio, i.e., 1-bit quantization, we propose group-specific quantization and quantile-based quantization approaches, motivated by the inherent patterns of the KV cache. Our method is plug-and-play, enabling seamless integration into various MLLMs to improve memory efficiency without architectural modifications. Extensive experiments demonstrate that our approach effectively reduces memory overhead while maintaining computational efficiency and preserving multimodal performance.

Large language models (LLMs) can solve challenging tasks. However, their inference computation on modern GPUs is highly inefficient due to the increasing number of tokens they must attend to as they generate new ones. To address this inefficiency, we capitalize on LLMs' problem-solving capabilities to optimize their own inference-time efficiency. We demonstrate with two specific tasks: (a) evaluating complex arithmetic expressions and (b) summarizing news articles. For both tasks, we create custom datasets to fine-tune an LLM. The goal of fine-tuning is twofold: first, to make the LLM learn to solve the evaluation or summarization task, and second, to train it to identify the minimal attention spans required for each step of the task. As a result, the fine-tuned model is able to convert these self-identified minimal attention spans into sparse attention masks on-the-fly during inference. We develop a custom CUDA kernel to take advantage of the reduced context to attend to. We demonstrate that using this custom CUDA kernel improves the throughput of LLM inference by 28%. Our work presents an end-to-end demonstration showing that training LLMs to self-select their attention spans speeds up autoregressive inference in solving real-world tasks.

Dynamic adaptation has become an essential technique in accelerating distributed machine learning (ML) training. Recent studies have shown that dynamically adjusting model structure (e.g., lottery ticket hypothesis) or hyperparameters (e.g., batch size) can significantly accelerate training without sacrificing accuracy. However, existing ML cluster schedulers are not designed to handle dynamic adaptation. We show that existing schemes fail to provide fairness and degrade system efficiency when the training throughput changes over time under dynamic adaptation. We design Shockwave, a scheduler with future planning that builds on two key ideas. First, Shockwave extends classic market theory from static settings to dynamic settings to co-optimize efficiency and fairness. Second, Shockwave utilizes stochastic dynamic programming to handle dynamic changes. We build a system for Shockwave and validate its performance with both trace-driven simulation and cluster experiments. Results show that for traces of ML jobs with dynamic adaptation, Shockwave improves makespan by 1.3X and fairness by 2X when compared with existing fair scheduling schemes.

Recent advances in quantum information science enabled the development of quantum communication network prototypes and created an opportunity to study full-stack quantum network architectures. This work develops SeQUeNCe, a comprehensive, customizable quantum network simulator. Our simulator consists of five modules: Hardware models, Entanglement Management protocols, Resource Management, Network Management, and Application. This framework is suitable for simulation of quantum network prototypes that capture the breadth of current and future hardware technologies and protocols. We implement a comprehensive suite of network protocols and demonstrate the use of SeQUeNCe by simulating a photonic quantum network with nine routers equipped with quantum memories. The simulation capabilities are illustrated in three use cases. We show the dependence of quantum network throughput on several key hardware parameters and study the impact of classical control message latency. We also investigate quantum memory usage efficiency in routers and demonstrate that redistributing memory according to anticipated load increases network capacity by 69.1% and throughput by 6.8%. We design SeQUeNCe to enable comparisons of alternative quantum network technologies, experiment planning, and validation and to aid with new protocol design. We are releasing SeQUeNCe as an open source tool and aim to generate community interest in extending it.

The surgence of Mixture of Experts (MoE) in Large Language Models promises a small price of execution cost for a much larger model parameter count and learning capacity, because only a small fraction of parameters are activated for each input token. However, it is commonly observed that some experts are activated far more often than others, leading to system inefficiency when running the experts on different devices in parallel. Therefore, we introduce Mixture of Grouped Experts (MoGE), which groups the experts during selection and balances the expert workload better than MoE in nature. It constrains tokens to activate an equal number of experts within each predefined expert group. When a model execution is distributed on multiple devices, this architectural design ensures a balanced computational load across devices, significantly enhancing throughput, particularly for the inference phase. Further, we build Pangu Pro MoE on Ascend NPUs, a sparse model based on MoGE with 72 billion total parameters, 16 billion of which are activated for each token. The configuration of Pangu Pro MoE is optimized for Ascend 300I Duo and 800I A2 through extensive system simulation studies. Our experiments indicate that MoGE indeed leads to better expert load balancing and more efficient execution for both model training and inference on Ascend NPUs. The inference performance of Pangu Pro MoE achieves 1148 tokens/s per card and can be further improved to 1528 tokens/s per card by speculative acceleration, outperforming comparable 32B and 72B Dense models. Furthermore, we achieve an excellent cost-to-performance ratio for model inference on Ascend 300I Duo. Our studies show that Ascend NPUs are capable of training Pangu Pro MoE with massive parallelization to make it a leading model within the sub-100B total parameter class, outperforming prominent open-source models like GLM-Z1-32B and Qwen3-32B.

Large scale deep learning model, such as modern language models and diffusion architectures, have revolutionized applications ranging from natural language processing to computer vision. However, their deployment in distributed or decentralized environments raises significant privacy concerns, as sensitive data may be exposed during inference. Traditional techniques like secure multi-party computation, homomorphic encryption, and differential privacy offer partial remedies but often incur substantial computational overhead, latency penalties, or limited compatibility with non-linear network operations. In this work, we introduce Equivariant Encryption (EE), a novel paradigm designed to enable secure, "blind" inference on encrypted data with near zero performance overhead. Unlike fully homomorphic approaches that encrypt the entire computational graph, EE selectively obfuscates critical internal representations within neural network layers while preserving the exact functionality of both linear and a prescribed set of non-linear operations. This targeted encryption ensures that raw inputs, intermediate activations, and outputs remain confidential, even when processed on untrusted infrastructure. We detail the theoretical foundations of EE, compare its performance and integration complexity against conventional privacy preserving techniques, and demonstrate its applicability across a range of architectures, from convolutional networks to large language models. Furthermore, our work provides a comprehensive threat analysis, outlining potential attack vectors and baseline strategies, and benchmarks EE against standard inference pipelines in decentralized settings. The results confirm that EE maintains high fidelity and throughput, effectively bridging the gap between robust data confidentiality and the stringent efficiency requirements of modern, large scale model inference.

Both biological and artificial neural networks inherently balance their performance with their operational cost, which balances their computational abilities. Typically, an efficient neuromorphic neural network is one that learns representations that reduce the redundancies and dimensionality of its input. This is for instance achieved in sparse coding, and sparse representations derived from natural images yield representations that are heterogeneous, both in their sampling of input features and in the variance of those features. Here, we investigated the connection between natural images' structure, particularly oriented features, and their corresponding sparse codes. We showed that representations of input features scattered across multiple levels of variance substantially improve the sparseness and resilience of sparse codes, at the cost of reconstruction performance. This echoes the structure of the model's input, allowing to account for the heterogeneously aleatoric structures of natural images. We demonstrate that learning kernel from natural images produces heterogeneity by balancing between approximate and dense representations, which improves all reconstruction metrics. Using a parametrized control of the kernels' heterogeneity used by a convolutional sparse coding algorithm, we show that heterogeneity emphasizes sparseness, while homogeneity improves representation granularity. In a broader context, these encoding strategy can serve as inputs to deep convolutional neural networks. We prove that such variance-encoded sparse image datasets enhance computational efficiency, emphasizing the benefits of kernel heterogeneity to leverage naturalistic and variant input structures and possible applications to improve the throughput of neuromorphic hardware.

In the field of antibody engineering, an essential task is to design a novel antibody whose paratopes bind to a specific antigen with correct epitopes. Understanding antibody structure and its paratope can facilitate a mechanistic understanding of its function. Therefore, antibody structure prediction from its sequence alone has always been a highly valuable problem for de novo antibody design. AlphaFold2, a breakthrough in the field of structural biology, provides a solution to predict protein structure based on protein sequences and computationally expensive coevolutionary multiple sequence alignments (MSAs). However, the computational efficiency and undesirable prediction accuracy of antibodies, especially on the complementarity-determining regions (CDRs) of antibodies limit their applications in the industrially high-throughput drug design. To learn an informative representation of antibodies, we employed a deep antibody language model (ALM) on curated sequences from the observed antibody space database via a transformer model. We also developed a novel model named xTrimoABFold to predict antibody structure from antibody sequence based on the pretrained ALM as well as efficient evoformers and structural modules. The model was trained end-to-end on the antibody structures in PDB by minimizing the ensemble loss of domain-specific focal loss on CDR and the frame-aligned point loss. xTrimoABFold outperforms AlphaFold2 and other protein language model based SOTAs, e.g., OmegaFold, HelixFold-Single, and IgFold with a large significant margin (30+\% improvement on RMSD) while performing 151 times faster than AlphaFold2. To the best of our knowledge, xTrimoABFold achieved state-of-the-art antibody structure prediction. Its improvement in both accuracy and efficiency makes it a valuable tool for de novo antibody design and could make further improvements in immuno-theory.

Many deep learning models, developed in recent years, reach higher ImageNet accuracy than ResNet50, with fewer or comparable FLOPS count. While FLOPs are often seen as a proxy for network efficiency, when measuring actual GPU training and inference throughput, vanilla ResNet50 is usually significantly faster than its recent competitors, offering better throughput-accuracy trade-off. In this work, we introduce a series of architecture modifications that aim to boost neural networks' accuracy, while retaining their GPU training and inference efficiency. We first demonstrate and discuss the bottlenecks induced by FLOPs-optimizations. We then suggest alternative designs that better utilize GPU structure and assets. Finally, we introduce a new family of GPU-dedicated models, called TResNet, which achieve better accuracy and efficiency than previous ConvNets. Using a TResNet model, with similar GPU throughput to ResNet50, we reach 80.8 top-1 accuracy on ImageNet. Our TResNet models also transfer well and achieve state-of-the-art accuracy on competitive single-label classification datasets such as Stanford cars (96.0%), CIFAR-10 (99.0%), CIFAR-100 (91.5%) and Oxford-Flowers (99.1%). They also perform well on multi-label classification and object detection tasks. Implementation is available at: https://github.com/mrT23/TResNet.

Matrix multiplication (MatMul) typically dominates the overall computational cost of large language models (LLMs). This cost only grows as LLMs scale to larger embedding dimensions and context lengths. In this work, we show that MatMul operations can be completely eliminated from LLMs while maintaining strong performance at billion-parameter scales. Our experiments show that our proposed MatMul-free models achieve performance on-par with state-of-the-art Transformers that require far more memory during inference at a scale up to at least 2.7B parameters. We investigate the scaling laws and find that the performance gap between our MatMul-free models and full precision Transformers narrows as the model size increases. We also provide a GPU-efficient implementation of this model which reduces memory usage by up to 61% over an unoptimized baseline during training. By utilizing an optimized kernel during inference, our model's memory consumption can be reduced by more than 10x compared to unoptimized models. To properly quantify the efficiency of our architecture, we build a custom hardware solution on an FPGA which exploits lightweight operations beyond what GPUs are capable of. We processed billion-parameter scale models at 13W beyond human readable throughput, moving LLMs closer to brain-like efficiency. This work not only shows how far LLMs can be stripped back while still performing effectively, but also points at the types of operations future accelerators should be optimized for in processing the next generation of lightweight LLMs. Our code implementation is available at https://github.com/ridgerchu/matmulfreellm.

The full-size MLPs and the projection layers in attention introduce tremendous model sizes of large language models (LLMs), consuming extensive computational resources in pre-training. We empirically observe that the activations of pre-trained LLMs exhibit low-rank property. Motivated by such observations, we propose CoLA and its memory-efficient implementation, CoLA-M, to replace these full-size layers with compute-efficient auto-encoders that naturally enforce low-rank activations throughout training. This fundamental architectural change eliminates the activation redundancy and significantly boosts model capacity and training efficiency. Experiments on LLaMA models with 60 million to 7 billion parameters show that CoLA reduces the computing cost by bf 2times and improves training throughput by bf 1.86times while maintaining full-rank level performance. CoLA-M further squeezes memory cost without sacrificing throughput, offering a pre-training approach with collectively superior parameter, computing, and memory efficiency. The LLMs produced are also bf 2times smaller, enabling faster inference with lower memory cost on resource-constrained platforms.

Pre-Layer Normalization (Pre-LN) is the de facto choice for large language models (LLMs) and is crucial for stable pretraining and effective transfer learning. However, Pre-LN is inefficient due to repeated statistical calculations and suffers from the curse of depth. As layers grow, the magnitude and variance of the hidden state escalate, destabilizing training. Efficiency-oriented normalization-free methods such as Dynamic Tanh (DyT) improve speed but remain fragile at depth. To jointly address stability and efficiency, we propose Bounded Hyperbolic Tanh (BHyT), a drop-in replacement for Pre-LN. BHyT couples a tanh nonlinearity with explicit, data-driven input bounding to keep activations within a non-saturating range. It prevents depth-wise growth in activation magnitude and variance and comes with a theoretical stability guarantee. For efficiency, BHyT computes exact statistics once per block and replaces a second normalization with a lightweight variance approximation, enhancing efficiency. Empirically, BHyT demonstrates improved stability and efficiency during pretraining, achieving an average of 15.8% faster training and an average of 4.2% higher token generation throughput compared to RMSNorm., while matching or surpassing its inference performance and robustness across language understanding and reasoning benchmarks. Our code is available at: https://anonymous.4open.science/r/BHyT

We show that the majority of the inference computations for large generative models such as LLaMA and OPT can be performed with both weights and activations being cast to 4 bits, in a way that leads to practical speedups while at the same time maintaining good accuracy. We achieve this via a hybrid quantization strategy called QUIK, which compresses most of the weights and activations to 4-bit, while keeping some outlier weights and activations in higher-precision. Crucially, our scheme is designed with computational efficiency in mind: we provide GPU kernels with highly-efficient layer-wise runtimes, which lead to practical end-to-end throughput improvements of up to 3.1x relative to FP16 execution. Code and models are provided at https://github.com/IST-DASLab/QUIK.

Sparse training is one of the promising techniques to reduce the computational cost of DNNs while retaining high accuracy. In particular, N:M fine-grained structured sparsity, where only N out of consecutive M elements can be nonzero, has attracted attention due to its hardware-friendly pattern and capability of achieving a high sparse ratio. However, the potential to accelerate N:M sparse DNN training has not been fully exploited, and there is a lack of efficient hardware supporting N:M sparse training. To tackle these challenges, this paper presents a computation-efficient training scheme for N:M sparse DNNs using algorithm, architecture, and dataflow co-design. At the algorithm level, a bidirectional weight pruning method, dubbed BDWP, is proposed to leverage the N:M sparsity of weights during both forward and backward passes of DNN training, which can significantly reduce the computational cost while maintaining model accuracy. At the architecture level, a sparse accelerator for DNN training, namely SAT, is developed to neatly support both the regular dense operations and the computation-efficient N:M sparse operations. At the dataflow level, multiple optimization methods ranging from interleave mapping, pre-generation of N:M sparse weights, and offline scheduling, are proposed to boost the computational efficiency of SAT. Finally, the effectiveness of our training scheme is evaluated on a Xilinx VCU1525 FPGA card using various DNN models and datasets. Experimental results show the SAT accelerator with the BDWP sparse training method under 2:8 sparse ratio achieves an average speedup of 1.75x over that with the dense training, accompanied by a negligible accuracy loss of 0.56% on average. Furthermore, our proposed training scheme significantly improves the training throughput by 2.97~25.22x and the energy efficiency by 1.36~3.58x over prior FPGA-based accelerators.

Large Language Models (LLMs) have achieved remarkable results. But existing models are expensive to train and deploy, and it is also difficult to expand their knowledge beyond pre-training data without forgetting previous knowledge. This paper proposes a new neural network architecture, ModuleFormer, that leverages modularity to improve the efficiency and flexibility of large language models. ModuleFormer is based on the Sparse Mixture of Experts (SMoE). Unlike the previous SMoE-based modular language model [Gururangan et al., 2021], which requires domain-labeled data to learn domain-specific experts, ModuleFormer can induce modularity from uncurated data with its new load balancing and load concentration losses. ModuleFormer is a modular architecture that includes two different types of modules, new stick-breaking attention heads, and feedforward experts. Different modules are sparsely activated conditions on the input token during training and inference. In our experiment, we found that the modular architecture enables three important abilities for large pre-trained language models: 1) Efficiency, since ModuleFormer only activates a subset of its modules for each input token, thus it could achieve the same performance as dense LLMs with more than two times throughput; 2) Extendability, ModuleFormer is more immune to catastrophic forgetting than dense LLMs and can be easily extended with new modules to learn new knowledge that is not included in the training data; 3) Specialisation, finetuning ModuleFormer could specialize a subset of modules to the finetuning task, and the task-unrelated modules could be easily pruned for a lightweight deployment.

As computational demands continue to rise, assessing the environmental footprint of AI requires moving beyond energy and water consumption to include the material demands of specialized hardware. This study quantifies the material footprint of AI training by linking computational workloads to physical hardware needs. The elemental composition of the Nvidia A100 SXM 40 GB graphics processing unit (GPU) was analyzed using inductively coupled plasma optical emission spectroscopy, which identified 32 elements. The results show that AI hardware consists of about 90% heavy metals and only trace amounts of precious metals. The elements copper, iron, tin, silicon, and nickel dominate the GPU composition by mass. In a multi-step methodology, we integrate these measurements with computational throughput per GPU across varying lifespans, accounting for the computational requirements of training specific AI models at different training efficiency regimes. Scenario-based analyses reveal that, depending on Model FLOPs Utilization (MFU) and hardware lifespan, training GPT-4 requires between 1,174 and 8,800 A100 GPUs, corresponding to the extraction and eventual disposal of up to 7 tons of toxic elements. Combined software and hardware optimization strategies can reduce material demands: increasing MFU from 20% to 60% lowers GPU requirements by 67%, while extending lifespan from 1 to 3 years yields comparable savings; implementing both measures together reduces GPU needs by up to 93%. Our findings highlight that incremental performance gains, such as those observed between GPT-3.5 and GPT-4, come at disproportionately high material costs. The study underscores the necessity of incorporating material resource considerations into discussions of AI scalability, emphasizing that future progress in AI must align with principles of resource efficiency and environmental responsibility.

The growing demand for large language models (LLMs) requires serving systems to handle many concurrent requests with diverse service level objectives (SLOs). This exacerbates head-of-line (HoL) blocking during the compute-intensive prefill phase, where long-running requests monopolize resources and delay higher-priority ones, leading to widespread time-to-first-token (TTFT) SLO violations. While chunked prefill enables interruptibility, it introduces an inherent trade-off between responsiveness and throughput: reducing chunk size improves response latency but degrades computational efficiency, whereas increasing chunk size maximizes throughput but exacerbates blocking. This necessitates an adaptive preemption mechanism. However, dynamically balancing execution granularity against scheduling overheads remains a key challenge. In this paper, we propose FlowPrefill, a TTFT-goodput-optimized serving system that resolves this conflict by decoupling preemption granularity from scheduling frequency. To achieve adaptive prefill scheduling, FlowPrefill introduces two key innovations: 1) Operator-Level Preemption, which leverages operator boundaries to enable fine-grained execution interruption without the efficiency loss associated with fixed small chunking; and 2) Event-Driven Scheduling, which triggers scheduling decisions only upon request arrival or completion events, thereby supporting efficient preemption responsiveness while minimizing control-plane overhead. Evaluation on real-world production traces shows that FlowPrefill improves maximum goodput by up to 5.6times compared to state-of-the-art systems while satisfying heterogeneous SLOs.

With the emergence of wearable devices and other embedded systems, deploying large language models (LLMs) on edge platforms has become an urgent need. However, this is challenging because of their high computational and memory demands. Although recent low-bit quantization methods (e.g., BitNet, DeepSeek) compress weights to as low as 1.58~bits with minimal accuracy loss, edge deployment is still constrained by limited on-chip resources, power budgets, and the often-neglected long latency of the prefill stage. We present TeLLMe, the first table-lookup-based ternary LLM accelerator for low-power edge FPGAs that fully supports both prefill and autoregressive decoding using 1.58-bit weights and 8-bit activations. TeLLMe incorporates several novel techniques, including (1) a table-lookup-based ternary matrix multiplication (TLMM) engine utilizing grouped activations and online precomputation for low resource utilization and high throughput; (2) a fine-grained analytic URAM-based weight buffer management scheme for efficient loading and compute engine access; (3) a streaming dataflow architecture that fuses floating-point element-wise operations with linear computations to hide latency; (4) a reversed-reordered prefill stage attention with fused attention operations for high memory efficiency; and (5) a resource-efficient specialized decoding stage attention. Under a 5~W power budget, TeLLMe delivers up to 25~tokens/s decoding throughput and 0.45--0.96~s time-to-first-token (TTFT) for 64--128 token prompts, marking a significant energy-efficiency advancement in LLM inference on edge FPGAs.

Recently, large language models (LLMs) have achieved huge success in the natural language processing (NLP) field, driving a growing demand to extend their deployment from the cloud to edge devices. However, deploying LLMs on resource-constrained edge devices poses significant challenges, including (1) intensive computations and huge model sizes, (2) great memory and bandwidth demands introduced by the autoregressive generation process, and (3) limited scalability for handling long sequences. To address these challenges, we propose AccLLM, a comprehensive acceleration framework that enables efficient and fast long-context LLM inference through algorithm and hardware co-design. At the algorithmic level, we integrate (1) pruning, (2) {\Lambda}-shaped attention, and (3) an innovative W2A8KV4 (2-bit weights, 8-bit activations, and 4-bit KV cache) quantization scheme, thus effectively reducing memory and bandwidth requirements while facilitating LLMs' long-sequence generation. At the hardware level, we design a dedicated FPGA-based accelerator with a reconfigurable computing engine to effectively and flexibly accommodate diverse operations arising from our compression algorithm, thereby fully translating the algorithmic innovations into tangible hardware efficiency. We validate AccLLM on the Xilinx Alveo U280 FPGA, demonstrating a 4.07x energy efficiency and a 2.98x throughput compared to the state-of-the-art work FlightLLM.

Efficient distributed training serves as a powerful catalyst and an essential foundation for the development of large-scale neural networks. In distributed training scenarios, various pipeline parallelism methods are cleverly designed and widely employed. In this paper, we propose ZeroPP, a highly efficient and flexible pipeline parallelism method that trades off pipeline bubbles, memory usage, and communication through adaptive scheduling units. ZeroPP achieves minimal pipeline bubbles by carefully staggering the computation tasks of forward, input gradient, and weight gradient within a scheduling unit. Additionally, ZeroPP optimizes the combination of pipeline parallelism and fully sharded data parallelism using a blockwise schedule. We conduct experiments with popular GPT-style models and observe up to a 30% increase in throughput compared to the state-of-the-art breath-first pipeline parallelism. Besides, our evaluation also demonstrates up to a 68% increase in throughput and a 10% reduction in memory consumption compared to the memory-efficient 1F1B method.

Diffusion language models hold the promise of fast parallel generation, while autoregressive (AR) models typically excel in quality due to their causal structure aligning naturally with language modeling. This raises a fundamental question: can we achieve a synergy with high throughput, higher GPU utilization, and AR level quality? Existing methods fail to effectively balance these two aspects, either prioritizing AR using a weaker model for sequential drafting (speculative decoding), leading to lower drafting efficiency, or using some form of left-to-right (AR-like) decoding logic for diffusion, which still suffers from quality degradation and forfeits its potential parallelizability. We introduce TiDAR, a sequence-level hybrid architecture that drafts tokens (Thinking) in Diffusion and samples final outputs (Talking) AutoRegressively - all within a single forward pass using specially designed structured attention masks. This design exploits the free GPU compute density, achieving a strong balance between drafting and verification capacity. Moreover, TiDAR is designed to be serving-friendly (low overhead) as a standalone model. We extensively evaluate TiDAR against AR models, speculative decoding, and diffusion variants across generative and likelihood tasks at 1.5B and 8B scales. Thanks to the parallel drafting and sampling as well as exact KV cache support, TiDAR outperforms speculative decoding in measured throughput and surpasses diffusion models like Dream and Llada in both efficiency and quality. Most notably, TiDAR is the first architecture to close the quality gap with AR models while delivering 4.71x to 5.91x more tokens per second.

Recent advances in large language models (LLMs) have driven impressive progress in omni-modal understanding and generation. However, training omni-modal LLMs remains a significant challenge due to the heterogeneous model architectures required to process diverse modalities, necessitating sophisticated system design for efficient large-scale training. Existing frameworks typically entangle model definition with parallel logic, incurring limited scalability and substantial engineering overhead for end-to-end omni-modal training. % We present \veomni, a modular and efficient training framework to accelerate the development of omni-modal LLMs. \veomni introduces model-centric distributed recipes that decouples communication from computation, enabling efficient 3D parallelism on omni-modal LLMs. \veomni also features a flexible configuration interface supporting seamless integration of new modalities with minimal code change. % Using \veomni, a omni-modal mixture-of-experts (MoE) model with 30B parameters can be trained with over 2,800 tokens/sec/GPU throughput and scale to 160K context lengths via 3D parallelism on 128 GPUs, showcasing its superior efficiency and scalability for training large omni-modal LLMs.

Diffusion Transformers (DiTs) have become a leading backbone for video generation, yet their quadratic attention cost remains a major bottleneck. Sparse attention reduces this cost by computing only a subset of attention blocks. However, prior methods often either drop the remaining blocks, which incurs information loss, or rely on learned predictors to approximate them, introducing training overhead and potential output distribution shifting. In this paper, we show that the missing contributions can be recovered without training: after semantic clustering, keys and values within each block exhibit strong similarity and can be well summarized by a small set of cluster centroids. Based on this observation, we introduce SVG-EAR, a parameter-free linear compensation branch that uses the centroid to approximate skipped blocks and recover their contributions. While centroid compensation is accurate for most blocks, it can fail on a small subset. Standard sparsification typically selects blocks by attention scores, which indicate where the model places its attention mass, but not where the approximation error would be largest. SVG-EAR therefore performs error-aware routing: a lightweight probe estimates the compensation error for each block, and we compute exactly the blocks with the highest error-to-cost ratio while compensating for skipped blocks. We provide theoretical guarantees that relate attention reconstruction error to clustering quality, and empirically show that SVG-EAR improves the quality-efficiency trade-off and increases throughput at the same generation fidelity on video diffusion tasks. Overall, SVG-EAR establishes a clear Pareto frontier over prior approaches, achieving up to 1.77times and 1.93times speedups while maintaining PSNRs of up to 29.759 and 31.043 on Wan2.2 and HunyuanVideo, respectively.

Vision-language model (VLM) based GUI agents show promise for automating complex desktop and mobile tasks, but face significant challenges in applying reinforcement learning (RL): (1) slow multi-turn interactions with GUI environments for policy rollout, and (2) insufficient high-quality agent-environment interactions for policy learning. To address these challenges, we propose DART, a Decoupled Agentic RL Training framework for GUI agents, which coordinates heterogeneous modules in a highly decoupled manner. DART separates the training system into four asynchronous modules: environment cluster, rollout service, data manager, and trainer. This design enables non-blocking communication, asynchronous training, rollout-wise trajectory sampling, and per-worker model synchronization, significantly improving the system efficiency: 1.6*GPU utilization for rollout, 1.9* training throughput, and 5.5* environment utilization. To facilitate effective learning from abundant samples, we introduce an adaptive data curation scheme: (1) pre-collecting successful trajectories for challenging tasks to supplement sparse success in online sampling; (2) dynamically adjusting rollout numbers and trajectory lengths based on task difficulty; (3) training selectively on high-entropy steps to prioritize critical decisions; (4) stabilizing learning via truncated importance sampling for policy mismatch between policy rollout and updating. On the OSWorld benchmark, DART-GUI-7B achieves a 42.13% task success rate, a 14.61% absolute gain over the base model, and 7.34% higher than open-source SOTA. We will fully open-source our training framework, data, and model checkpoints via computer-use-agents.github.io/dart-gui, which we believe is a timely contribution to the open-source community of agentic RL training.

Transformer models trained on long sequences often achieve higher accuracy than short sequences. Unfortunately, conventional transformers struggle with long sequence training due to the overwhelming computation and memory requirements. Existing methods for long sequence training offer limited speedup and memory reduction, and may compromise accuracy. This paper presents a novel and efficient distributed training method, the Long Short-Sequence Transformer (LSS Transformer), for training transformer with long sequences. It distributes a long sequence into segments among GPUs, with each GPU computing a partial self-attention for its segment. Then, it uses a fused communication and a novel double gradient averaging technique to avoid the need to aggregate partial self-attention and minimize communication overhead. We evaluated the performance between LSS Transformer and the state-of-the-art Nvidia sequence parallelism on a Wikipedia enwik8 dataset. Results show that our proposed method lead to 5.6x faster and 10.2x more memory-efficient implementation compared to state-of-the-art sequence parallelism on 144 Nvidia V100 GPUs. Moreover, our algorithm scales to an extreme sequence length of 50,112 at 3,456 GPUs, achieving 161% super-linear parallel efficiency and a throughput of 32 petaflops.

We present MegaScale-MoE, a production system tailored for the efficient training of large-scale mixture-of-experts (MoE) models. MoE emerges as a promising architecture to scale large language models (LLMs) to unprecedented sizes, thereby enhancing model performance. However, existing MoE training systems experience a degradation in training efficiency, exacerbated by the escalating scale of MoE models and the continuous evolution of hardware. Recognizing the pivotal role of efficient communication in enhancing MoE training, MegaScale-MoE customizes communication-efficient parallelism strategies for attention and FFNs in each MoE layer and adopts a holistic approach to overlap communication with computation at both inter- and intra-operator levels. Additionally, MegaScale-MoE applies communication compression with adjusted communication patterns to lower precision, further improving training efficiency. When training a 352B MoE model on 1,440 NVIDIA Hopper GPUs, MegaScale-MoE achieves a training throughput of 1.41M tokens/s, improving the efficiency by 1.88times compared to Megatron-LM. We share our operational experience in accelerating MoE training and hope that by offering our insights in system design, this work will motivate future research in MoE systems.

In this work, we propose Causal Autoregressive Diffusion (CARD), a novel framework that unifies the training efficiency of ARMs with the high-throughput inference of diffusion models. CARD reformulates the diffusion process within a strictly causal attention mask, enabling dense, per-token supervision in a single forward pass. To address the optimization instability of causal diffusion, we introduce a soft-tailed masking schema to preserve local context and a context-aware reweighting mechanism derived from signal-to-noise principles. This design enables dynamic parallel decoding, where the model leverages KV-caching to adaptively generate variable-length token sequences based on confidence. Empirically, CARD outperforms existing discrete diffusion baselines while reducing training latency by 3 times compared to block diffusion methods. Our results demonstrate that CARD achieves ARM-level data efficiency while unlocking the latency benefits of parallel generation, establishing a robust paradigm for next-generation efficient LLMs.

Medical images are acquired at high resolutions with large fields of view in order to capture fine-grained features necessary for clinical decision-making. Consequently, training deep learning models on medical images can incur large computational costs. In this work, we address the challenge of downsizing medical images in order to improve downstream computational efficiency while preserving clinically-relevant features. We introduce MedVAE, a family of six large-scale 2D and 3D autoencoders capable of encoding medical images as downsized latent representations and decoding latent representations back to high-resolution images. We train MedVAE autoencoders using a novel two-stage training approach with 1,052,730 medical images. Across diverse tasks obtained from 20 medical image datasets, we demonstrate that (1) utilizing MedVAE latent representations in place of high-resolution images when training downstream models can lead to efficiency benefits (up to 70x improvement in throughput) while simultaneously preserving clinically-relevant features and (2) MedVAE can decode latent representations back to high-resolution images with high fidelity. Our work demonstrates that large-scale, generalizable autoencoders can help address critical efficiency challenges in the medical domain. Our code is available at https://github.com/StanfordMIMI/MedVAE.

Research in efficient vision backbones is evolving into models that are a mixture of convolutions and transformer blocks. A smart combination of both, architecture-wise and component-wise is mandatory to excel in the speedaccuracy trade-off. Most publications focus on maximizing accuracy and utilize MACs (multiply accumulate operations) as an efficiency metric. The latter however often do not measure accurately how fast a model actually is due to factors like memory access cost and degree of parallelism. We analyzed common modules and architectural design choices for backbones not in terms of MACs, but rather in actual throughput and latency, as the combination of the latter two is a better representation of the efficiency of models in real applications. We applied the conclusions taken from that analysis to create a recipe for increasing hardware-efficiency in macro design. Additionally we introduce a simple slimmed-down version of MultiHead Self-Attention, that aligns with our analysis. We combine both macro and micro design to create a new family of hardware-efficient backbone networks called LowFormer. LowFormer achieves a remarkable speedup in terms of throughput and latency, while achieving similar or better accuracy than current state-of-the-art efficient backbones. In order to prove the generalizability of our hardware-efficient design, we evaluate our method on GPU, mobile GPU and ARM CPU. We further show that the downstream tasks object detection and semantic segmentation profit from our hardware-efficient architecture. Code and models are available at https://github.com/ altair199797/LowFormer.

Increasing demand for Large Language Models (LLMs) services imposes substantial deployment and computation costs on providers. LLM routing offers a cost-efficient solution by directing queries to the optimal LLM based on model and query features. However, existing works primarily focus on offline scenarios and struggle to adapt to online settings with high query volume and constrained token budgets. In this work, we introduce the first training-free algorithm for online routing scenarios. Our algorithm leverages approximate nearest neighbor search to efficiently estimate query features and performs a one-time optimization over a small set of initial queries to learn a routing strategy that guides future routing. We provide theoretical guarantees demonstrating that our algorithm achieves a competitive ratio of 1 - o(1) under natural assumptions, which is further validated by extensive experiments across 3 benchmark datasets and 8 baselines, showing an average improvement of 3.55times in overall performance, 1.85times in cost efficiency, and nearly 4.25times in throughput. Our code is available at https://github.com/fzwark/PORT.

Existing visual token compression methods for Multimodal Large Language Models (MLLMs) predominantly operate as post-encoder modules, limiting their potential for efficiency gains. To address this limitation, we propose LaCo (Layer-wise Visual Token Compression), a novel framework that enables effective token compression within the intermediate layers of the vision encoder. LaCo introduces two core components: 1) a layer-wise pixel-shuffle mechanism that systematically merges adjacent tokens through space-to-channel transformations, and 2) a residual learning architecture with non-parametric shortcuts that preserves critical visual information during compression. Extensive experiments indicate that our LaCo outperforms all existing methods when compressing tokens in the intermediate layers of the vision encoder, demonstrating superior effectiveness. In addition, compared to external compression, our method improves training efficiency beyond 20% and inference throughput over 15% while maintaining strong performance.

Human-driven vehicles (HVs) exhibit complex and diverse behaviors. Accurately modeling such behavior is crucial for validating Robot Vehicles (RVs) in simulation and realizing the potential of mixed traffic control. However, existing approaches like parameterized models and data-driven techniques struggle to capture the full complexity and diversity. To address this, in this work, we introduce CARL, a hybrid technique combining imitation learning for close proximity car-following and probabilistic sampling for larger headways. We also propose two classes of RL-based RVs: a safety RV focused on maximizing safety and an efficiency RV focused on maximizing efficiency. Our experiments show that the safety RV increases Time-to-Collision above the critical 4 second threshold and reduces Deceleration Rate to Avoid a Crash by up to 80%, while the efficiency RV achieves improvements in throughput of up to 49%. These results demonstrate the effectiveness of CARL in enhancing both safety and efficiency in mixed traffic.

Deep learning datasets are expanding at an unprecedented pace, creating new challenges for data processing in model training pipelines. A crucial aspect of these pipelines is dataset shuffling, which significantly improves unbiased learning and convergence accuracy by adhering to the principles of random sampling. However, loading shuffled data for large datasets incurs significant overhead in the deep learning pipeline and severely impacts the end-to-end training throughput. To mitigate this, current deep learning systems often resort to partial dataset shuffling, sacrificing global randomness to maintain acceptable training throughput on large datasets, still leaving global shuffling efficiency issues not fully explored. In this work, we present RINAS, a data loading framework that systematically addresses the performance bottleneck of loading global shuffled datasets. Our key contribution is to offer an intra-batch unordered data fetching approach, which unleashes unexplored parallelism of data loading. We implement RINAS under the PyTorch framework for common dataset libraries HuggingFace and TorchVision. Our experimental results show that RINAS improves the throughput of general language model training and vision model training by up to 59% and 89%, respectively.

Large Language Models (LLMs) have demonstrated remarkable capabilities in leveraging extensive external knowledge to enhance responses in multi-turn and agentic applications, such as retrieval-augmented generation (RAG). However, processing long-context inputs introduces significant system latency and demands substantial memory for the key-value cache, resulting in reduced throughput and a fundamental trade-off between knowledge enrichment and system efficiency. While minimizing latency for long-context inputs is a primary objective for LLMs, we contend that RAG require specialized consideration. In RAG, much of the LLM context consists of concatenated passages from retrieval, with only a small subset directly relevant to the query. These passages often exhibit low semantic similarity due to diversity or deduplication during re-ranking, leading to block-diagonal attention patterns that differ from those in standard LLM generation tasks. Based on this observation, we argue that most computations over the RAG context during decoding are unnecessary and can be eliminated with minimal impact on performance. To this end, we propose REFRAG, an efficient decoding framework that compresses, senses, and expands to improve latency in RAG applications. By exploiting the sparsity structure, we demonstrate a 30.85 the time-to-first-token acceleration (3.75 improvement to previous work) without loss in perplexity. In addition, our optimization framework for large context enables REFRAG to extend the context size of LLMs by 16. We provide rigorous validation of REFRAG across diverse long-context tasks, including RAG, multi-turn conversations, and long document summarization, spanning a wide range of datasets. Experimental results confirm that REFRAG delivers substantial speedup with no loss in accuracy compared to LLaMA models and other state-of-the-art baselines across various context sizes.

This technical report introduces Uni-Parser, an industrial-grade document parsing engine tailored for scientific literature and patents, delivering high throughput, robust accuracy, and cost efficiency. Unlike pipeline-based document parsing methods, Uni-Parser employs a modular, loosely coupled multi-expert architecture that preserves fine-grained cross-modal alignments across text, equations, tables, figures, and chemical structures, while remaining easily extensible to emerging modalities. The system incorporates adaptive GPU load balancing, distributed inference, dynamic module orchestration, and configurable modes that support either holistic or modality-specific parsing. Optimized for large-scale cloud deployment, Uni-Parser achieves a processing rate of up to 20 PDF pages per second on 8 x NVIDIA RTX 4090D GPUs, enabling cost-efficient inference across billions of pages. This level of scalability facilitates a broad spectrum of downstream applications, ranging from literature retrieval and summarization to the extraction of chemical structures, reaction schemes, and bioactivity data, as well as the curation of large-scale corpora for training next-generation large language models and AI4Science models.

Parameter-efficient fine-tuning (PEFT) has become the standard approach for adapting large language models under limited compute and memory budgets. Although previous methods improve efficiency through low-rank updates, quantization, or heuristic budget reallocation, they often decouple the allocation of capacity from the way updates evolve during training. In this work, we introduce CTR-LoRA, a framework guided by curvature trust region that integrates rank scheduling with stability-aware optimization. CTR-LoRA allocates parameters based on marginal utility derived from lightweight second-order proxies and constrains updates using a Fisher/Hessian-metric trust region. Experiments on multiple open-source backbones (7B-13B), evaluated on both in-distribution and out-of-distribution benchmarks, show consistent improvements over strong PEFT baselines. In addition to increased accuracy, CTR-LoRA enhances training stability, reduces memory requirements, and achieves higher throughput, positioning it on the Pareto frontier of performance and efficiency. These results highlight a principled path toward more robust and deployable PEFT.

Sparse observations and coarse-resolution climate models limit effective regional decision-making, underscoring the need for robust downscaling. However, existing AI methods struggle with generalization across variables and geographies and are constrained by the quadratic complexity of Vision Transformer (ViT) self-attention. We introduce ORBIT-2, a scalable foundation model for global, hyper-resolution climate downscaling. ORBIT-2 incorporates two key innovations: (1) Residual Slim ViT (Reslim), a lightweight architecture with residual learning and Bayesian regularization for efficient, robust prediction; and (2) TILES, a tile-wise sequence scaling algorithm that reduces self-attention complexity from quadratic to linear, enabling long-sequence processing and massive parallelism. ORBIT-2 scales to 10 billion parameters across 65,536 GPUs, achieving up to 4.1 exaFLOPS sustained throughput and 74--98% strong scaling efficiency. It supports downscaling to 0.9 km global resolution and processes sequences up to 4.2 billion tokens. On 7 km resolution benchmarks, ORBIT-2 achieves high accuracy with R^2 scores in the range of 0.98--0.99 against observational data.

Increasing the scale of reinforcement learning experiments has allowed researchers to achieve unprecedented results in both training sophisticated agents for video games, and in sim-to-real transfer for robotics. Typically such experiments rely on large distributed systems and require expensive hardware setups, limiting wider access to this exciting area of research. In this work we aim to solve this problem by optimizing the efficiency and resource utilization of reinforcement learning algorithms instead of relying on distributed computation. We present the "Sample Factory", a high-throughput training system optimized for a single-machine setting. Our architecture combines a highly efficient, asynchronous, GPU-based sampler with off-policy correction techniques, allowing us to achieve throughput higher than 10^5 environment frames/second on non-trivial control problems in 3D without sacrificing sample efficiency. We extend Sample Factory to support self-play and population-based training and apply these techniques to train highly capable agents for a multiplayer first-person shooter game. The source code is available at https://github.com/alex-petrenko/sample-factory

Efficient deployment of small language models (SLMs) is essential for numerous real-world applications with stringent latency constraints. While previous work on SLM design has primarily focused on reducing the number of parameters to achieve parameter-optimal SLMs, parameter efficiency does not necessarily translate into proportional real-device speed-ups. This work aims to identify the key determinants of SLMs' real-device latency and offer generalizable principles and methodologies for SLM design and training when real-device latency is the primary consideration. Specifically, we identify two central architectural factors: depth-width ratios and operator choices. The former is crucial for small-batch-size latency, while the latter affects both latency and large-batch-size throughput. In light of this, we first study latency-optimal depth-width ratios, with the key finding that although deep-thin models generally achieve better accuracy under the same parameter budget, they may not lie on the accuracy-latency trade-off frontier. Next, we explore emerging efficient attention alternatives to evaluate their potential as candidate building operators. Using the identified promising operators, we construct an evolutionary search framework to automatically discover latency-optimal combinations of these operators within hybrid SLMs, thereby advancing the accuracy-latency frontier. In addition to architectural improvements, we further enhance SLM training using a weight normalization technique that enables more effective weight updates and improves final convergence. Combining these methods, we introduce a new family of hybrid SLMs, called Nemotron-Flash, which significantly advances the accuracy-efficiency frontier of state-of-the-art SLMs, e.g., achieving over +5.5% average accuracy, 1.3x/1.9x lower latency, and 18.7x/45.6x higher throughput compared to Qwen3-1.7B/0.6B, respectively.

The computational sparsity of Mixture-of-Experts (MoE) models enables sub-linear growth in compute cost as model size increases, thus offering a scalable path to training massive neural networks. However, existing implementations suffer from low GPU utilization, significant latency overhead, and a fundamental inability to leverage task locality, primarily due to CPU-managed scheduling, host-initiated communication, and frequent kernel launches. To overcome these limitations, we develop FlashMoE, a fully GPU-resident MoE operator that fuses expert computation and inter-GPU communication into a single persistent GPU kernel. FlashMoE enables fine-grained pipelining of dispatch, compute, and combine phases, eliminating launch overheads and reducing idle gaps. Unlike existing work, FlashMoE eliminates bulk-synchronous collectives for one-sided, device-initiated, inter-GPU (R)DMA transfers, thereby unlocking payload efficiency by eliminating bloated or redundant network payloads in sparsely activated layers. When evaluated on an 8-H100 GPU node with MoE models comprising up to 128 experts and 16K token sequences, FlashMoE achieves up to 9x higher GPU utilization, 6x lower latency, 5.7x higher throughput, and 4x better overlap efficiency compared to state-of-the-art baselines, despite using FP32, whereas the baselines use FP16. FlashMoE shows that principled GPU kernel-hardware co-design is key to unlocking the performance ceiling of large-scale distributed ML. We provide code at https://github.com/osayamenja/FlashMoE.

In this technical report, we present the Zamba2 series -- a suite of 1.2B, 2.7B, and 7.4B parameter hybrid Mamba2-transformer models that achieve state of the art performance against the leading open-weights models of their class, while achieving substantial gains in inference latency, throughput, and memory efficiency. The Zamba2 series builds upon our initial work with Zamba1-7B, optimizing its architecture, training and annealing datasets, and training for up to three trillion tokens. We provide open-source weights for all models of the Zamba2 series as well as instruction-tuned variants that are strongly competitive against comparable instruct-tuned models of their class. We additionally open-source the pretraining dataset, which we call Zyda-2, used to train the Zamba2 series of models. The models and datasets used in this work are openly available at https://huggingface.co/Zyphra

We present the design, implementation and engineering experience in building and deploying MegaScale, a production system for training large language models (LLMs) at the scale of more than 10,000 GPUs. Training LLMs at this scale brings unprecedented challenges to training efficiency and stability. We take a full-stack approach that co-designs the algorithmic and system components across model block and optimizer design, computation and communication overlapping, operator optimization, data pipeline, and network performance tuning. Maintaining high efficiency throughout the training process (i.e., stability) is an important consideration in production given the long extent of LLM training jobs. Many hard stability issues only emerge at large scale, and in-depth observability is the key to address them. We develop a set of diagnosis tools to monitor system components and events deep in the stack, identify root causes, and derive effective techniques to achieve fault tolerance and mitigate stragglers. MegaScale achieves 55.2% Model FLOPs Utilization (MFU) when training a 175B LLM model on 12,288 GPUs, improving the MFU by 1.34x compared to Megatron-LM. We share our operational experience in identifying and fixing failures and stragglers. We hope by articulating the problems and sharing our experience from a systems perspective, this work can inspire future LLM systems research.

A vital step towards the widespread adoption of neural retrieval models is their resource efficiency throughout the training, indexing and query workflows. The neural IR community made great advancements in training effective dual-encoder dense retrieval (DR) models recently. A dense text retrieval model uses a single vector representation per query and passage to score a match, which enables low-latency first stage retrieval with a nearest neighbor search. Increasingly common, training approaches require enormous compute power, as they either conduct negative passage sampling out of a continuously updating refreshing index or require very large batch sizes for in-batch negative sampling. Instead of relying on more compute capability, we introduce an efficient topic-aware query and balanced margin sampling technique, called TAS-Balanced. We cluster queries once before training and sample queries out of a cluster per batch. We train our lightweight 6-layer DR model with a novel dual-teacher supervision that combines pairwise and in-batch negative teachers. Our method is trainable on a single consumer-grade GPU in under 48 hours (as opposed to a common configuration of 8x V100s). We show that our TAS-Balanced training method achieves state-of-the-art low-latency (64ms per query) results on two TREC Deep Learning Track query sets. Evaluated on NDCG@10, we outperform BM25 by 44%, a plainly trained DR by 19%, docT5query by 11%, and the previous best DR model by 5%. Additionally, TAS-Balanced produces the first dense retriever that outperforms every other method on recall at any cutoff on TREC-DL and allows more resource intensive re-ranking models to operate on fewer passages to improve results further.

Long-context modeling is crucial for next-generation language models, yet the high computational cost of standard attention mechanisms poses significant computational challenges. Sparse attention offers a promising direction for improving efficiency while maintaining model capabilities. We present NSA, a Natively trainable Sparse Attention mechanism that integrates algorithmic innovations with hardware-aligned optimizations to achieve efficient long-context modeling. NSA employs a dynamic hierarchical sparse strategy, combining coarse-grained token compression with fine-grained token selection to preserve both global context awareness and local precision. Our approach advances sparse attention design with two key innovations: (1) We achieve substantial speedups through arithmetic intensity-balanced algorithm design, with implementation optimizations for modern hardware. (2) We enable end-to-end training, reducing pretraining computation without sacrificing model performance. As shown in Figure 1, experiments show the model pretrained with NSA maintains or exceeds Full Attention models across general benchmarks, long-context tasks, and instruction-based reasoning. Meanwhile, NSA achieves substantial speedups over Full Attention on 64k-length sequences across decoding, forward propagation, and backward propagation, validating its efficiency throughout the model lifecycle.

LLM-based coding agents have shown strong performance on automated issue resolution benchmarks, yet existing evaluations largely focus on final task success, providing limited insight into how agents retrieve and use code context during problem solving. We introduce ContextBench, a process-oriented evaluation of context retrieval in coding agents. ContextBench consists of 1,136 issue-resolution tasks from 66 repositories across eight programming languages, each augmented with human-annotated gold contexts. We further implement an automated evaluation framework that tracks agent trajectories and measures context recall, precision, and efficiency throughout issue resolution. Using ContextBench, we evaluate four frontier LLMs and five coding agents. Our results show that sophisticated agent scaffolding yields only marginal gains in context retrieval ("The Bitter Lesson" of coding agents), LLMs consistently favor recall over precision, and substantial gaps exist between explored and utilized context. ContextBench augments existing end-to-end benchmarks with intermediate gold-context metrics that unbox the issue-resolution process. These contexts offer valuable intermediate signals for guiding LLM reasoning in software tasks.

The rapid advancement of Large Language Models (LLMs) has driven the need for more efficient serving strategies. In this context, efficiency refers to the proportion of requests that meet their Service Level Objectives (SLOs), particularly for Time To First Token (TTFT) and Time Between Tokens (TBT). However, existing systems often prioritize one metric at the cost of the other. We present Ascendra, an LLM serving system designed to meet both TTFT and TBT SLOs simultaneously. The core insight behind Ascendra is that a request's urgency evolves as it approaches its deadline. To leverage this, Ascendra partitions GPU resources into two types of instances: low-priority and high-priority. Low-priority instances maximize throughput by processing requests out of arrival order, but at the risk of request starvation. To address this, Ascendra employs a performance model to predict requests at risk of missing their SLOs and proactively offloads them to high-priority instances. High-priority instances are optimized for low-latency execution and handle urgent requests nearing their deadlines. This partitioned architecture enables Ascendra to effectively balance high throughput and low latency. Extensive evaluation shows that Ascendra improves system throughput by up to 1.7x compared to vLLM and Sarathi-Serve while meeting both TTFT and TBT SLOs.

The increasing usage of Large Language Models (LLMs) has resulted in a surging demand for planet-scale serving systems, where tens of thousands of GPUs continuously serve hundreds of millions of users. Consequently, throughput (under reasonable latency constraints) has emerged as a key metric that determines serving systems' performance. To boost throughput, various methods of inter-device parallelism (e.g., data, tensor, pipeline) have been explored. However, existing methods do not consider overlapping the utilization of different resources within a single device, leading to underutilization and sub-optimal performance. We propose NanoFlow, a novel serving framework that exploits intra-device parallelism, which overlaps the usage of resources including compute, memory, and network within a single device through operation co-scheduling. To exploit intra-device parallelism, NanoFlow introduces two key innovations: First, NanoFlow splits requests into nano-batches at the granularity of operations, which breaks the dependency of sequential operations in LLM inference and enables overlapping; then, to get benefit from overlapping, NanoFlow uses an operation-level pipeline with execution unit scheduling, which partitions the device's functional units and simultaneously executes different operations in each unit. NanoFlow automates the pipeline setup using a parameter search algorithm, which enables easily porting NanoFlow to different models. We implement NanoFlow on NVIDIA GPUs and evaluate end-to-end serving throughput on several popular models such as LLaMA-2-70B, Mixtral 8x7B, LLaMA-3-8B, etc.. With practical workloads, NanoFlow provides 1.91x throughput boost compared to state-of-the-art serving systems achieving 59% to 72% of optimal throughput across ported models.

Recent years have witnessed increasing interest in extending large language models into agentic systems. While the effectiveness of agents has continued to improve, efficiency, which is crucial for real-world deployment, has often been overlooked. This paper therefore investigates efficiency from three core components of agents: memory, tool learning, and planning, considering costs such as latency, tokens, steps, etc. Aimed at conducting comprehensive research addressing the efficiency of the agentic system itself, we review a broad range of recent approaches that differ in implementation yet frequently converge on shared high-level principles including but not limited to bounding context via compression and management, designing reinforcement learning rewards to minimize tool invocation, and employing controlled search mechanisms to enhance efficiency, which we discuss in detail. Accordingly, we characterize efficiency in two complementary ways: comparing effectiveness under a fixed cost budget, and comparing cost at a comparable level of effectiveness. This trade-off can also be viewed through the Pareto frontier between effectiveness and cost. From this perspective, we also examine efficiency oriented benchmarks by summarizing evaluation protocols for these components and consolidating commonly reported efficiency metrics from both benchmark and methodological studies. Moreover, we discuss the key challenges and future directions, with the goal of providing promising insights.