Daily Papers

Singular Value Decomposition (SVD) has recently seen a surge of interest as a simple yet powerful tool for large language models (LLMs) compression, with a growing number of works demonstrating 20-80% parameter reductions at minimal accuracy loss. Previous SVD-based approaches have focused primarily on reducing the memory footprint of model weights, largely overlooking the additional activation memory overhead incurred during inference when applying truncated factors via standard dense CUDA kernels. Our experiments demonstrate that this activation overhead, scaling with sequence length and hidden dimension, prevents current SVD compression techniques from achieving any reduction in peak inference memory, thereby limiting their viability for real-world, on-device deployments. We introduce FlashSVD, a novel, end-to-end rank-aware streaming inference framework specifically designed for SVD-compressed large language models. FlashSVD can be seamlessly integrated with any model that employs SVD-based methods for parameter reduction. By fusing low-rank projection kernels directly into both the self-attention and feed-forward network (FFN) pipelines, FlashSVD avoid materializing full-size activation buffers. Instead, small tiles of the truncated factors are loaded into on-chip SRAM, multiplied and reduced on the fly, and immediately evicted, preserving high GPU occupancy and adding no extra latency. On standard encoder benchmarks (e.g., BERT-Base), FlashSVD cuts peak activation memory by up to 70.2% and intermediate transient memory by 75%, all while incur no accuracy loss with upstreaming compression methods, offering a practical path toward memory-constrained deployment of low-rank LLMs.

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

Machine learning models deployed on edge devices have enabled numerous exciting new applications, such as humanoid robots, AR glasses, and autonomous vehicles. However, the computing resources available on these edge devices are not catching up with the ever-growing number of parameters in these models. As the models become bigger and more complicated, the novel yet sophisticated structure challenges the inference runtime optimization. We present FluidML, a generic runtime memory management and optimization framework that can flexibly transform the model execution blueprint to achieve faster and more memory-efficient inference. Evaluations across different platforms show that FluidML can consistently reduce the end-to-end inference latency by up to 25.38% for popular language models and reduce peak memory usage by up to 41.47%, compared to state-of-the-art approaches. FluidML is of ~30K line of codes, built for general-purpose usage, and will be released as an open-source inference runtime optimization framework to the community.

In recent years, Mixture-of-Experts (MoE) has emerged as an effective approach for enhancing the capacity of deep neural network (DNN) with sub-linear computational costs. However, storing all experts on GPUs incurs significant memory overhead, increasing the monetary cost of MoE-based inference. To address this, we propose eMoE, a memory efficient inference system for MoE-based large language models (LLMs) by leveraging our observations from experiment measurements. eMoE reduces memory usage by predicting and loading only the required experts based on recurrent patterns in expert routing. To reduce loading latency while maintaining accuracy, as we found using the same experts for subsequent prompts has minimal impact on perplexity, eMoE invokes the expert predictor every few prompts rather than for each prompt. In addition, it skips predictions for tasks less sensitive to routing accuracy. Finally, it has task-aware scheduling to minimize inference latency by considering Service Level Objectives (SLOs), task-specific output lengths, and expert loading latencies. Experimental results show that compared to existing systems, eMoE reduces memory consumption by up to 80% while maintaining accuracy and reduces inference latency by up to 17%. It also enables processing prompts 40x longer, batches 4.5x larger, and achieves 1.5x higher throughput.

Transformer model with multi-head attention requires caching intermediate results for efficient inference in generation tasks. However, cache brings new memory-related costs and prevents leveraging larger batch size for faster speed. We propose memory-efficient lossless attention (called EL-attention) to address this issue. It avoids heavy operations for building multi-head keys and values, cache for them is not needed. EL-attention constructs an ensemble of attention results by expanding query while keeping key and value shared. It produces the same result as multi-head attention with less GPU memory and faster inference speed. We conduct extensive experiments on Transformer, BART, and GPT-2 for summarization and question generation tasks. The results show EL-attention speeds up existing models by 1.6x to 5.3x without accuracy loss.

We demonstrate that a deep model can be reparametrized as a linear combination of several randomly initialized and frozen deep models in the weight space. During training, we seek local minima that reside within the subspace spanned by these random models (i.e., `basis' networks). Our framework, PRANC, enables significant compaction of a deep model. The model can be reconstructed using a single scalar `seed,' employed to generate the pseudo-random `basis' networks, together with the learned linear mixture coefficients. In practical applications, PRANC addresses the challenge of efficiently storing and communicating deep models, a common bottleneck in several scenarios, including multi-agent learning, continual learners, federated systems, and edge devices, among others. In this study, we employ PRANC to condense image classification models and compress images by compacting their associated implicit neural networks. PRANC outperforms baselines with a large margin on image classification when compressing a deep model almost 100 times. Moreover, we show that PRANC enables memory-efficient inference by generating layer-wise weights on the fly. The source code of PRANC is here: https://github.com/UCDvision/PRANC

We present LFM2, a family of Liquid Foundation Models designed for efficient on-device deployment and strong task capabilities. Using hardware-in-the-loop architecture search under edge latency and memory constraints, we obtain a compact hybrid backbone that combines gated short convolutions with a small number of grouped query attention blocks, delivering up to 2x faster prefill and decode on CPUs compared to similarly sized models. The LFM2 family covers 350M-8.3B parameters, including dense models (350M, 700M, 1.2B, 2.6B) and a mixture-of-experts variant (8.3B total, 1.5B active), all with 32K context length. LFM2's training pipeline includes a tempered, decoupled Top-K knowledge distillation objective that avoids support mismatch; curriculum learning with difficulty-ordered data; and a three-stage post-training recipe of supervised fine-tuning, length-normalized preference optimization, and model merging. Pre-trained on 10-12T tokens, LFM2 models achieve strong results across diverse benchmarks; for example, LFM2-2.6B reaches 79.56% on IFEval and 82.41% on GSM8K. We further build multimodal and retrieval variants: LFM2-VL for vision-language tasks, LFM2-Audio for speech, and LFM2-ColBERT for retrieval. LFM2-VL supports tunable accuracy-latency tradeoffs via token-efficient visual processing, while LFM2-Audio separates audio input and output pathways to enable real-time speech-to-speech interaction competitive with models 3x larger. LFM2-ColBERT provides a low-latency encoder for queries and documents, enabling high-performance retrieval across multiple languages. All models are released with open weights and deployment packages for ExecuTorch, llama.cpp, and vLLM, making LFM2 a practical base for edge applications that need fast, memory-efficient inference and strong task capabilities.

Effective reasoning is crucial to solving complex mathematical problems. Recent large language models (LLMs) have boosted performance by scaling test-time computation through long chain-of-thought reasoning. However, transformer-based models are inherently limited in extending context length due to their quadratic computational complexity and linear memory requirements. In this paper, we introduce a novel hybrid linear RNN reasoning model, M1, built on the Mamba architecture, which allows memory-efficient inference. Our approach leverages a distillation process from existing reasoning models and is further enhanced through RL training. Experimental results on the AIME and MATH benchmarks show that M1 not only outperforms previous linear RNN models but also matches the performance of state-of-the-art Deepseek R1 distilled reasoning models at a similar scale. We also compare our generation speed with a highly performant general purpose inference engine, vLLM, and observe more than a 3x speedup compared to a same size transformer. With throughput speedup, we are able to achieve higher accuracy compared to DeepSeek R1 distilled transformer reasoning models under a fixed generation time budget using self-consistency voting. Overall, we introduce a hybrid Mamba reasoning model and provide a more effective approach to scaling test-time generation using self-consistency or long chain of thought reasoning.

Transformer-based large language models (LLMs) demonstrate impressive performance in long context generation. Extending the context length has disproportionately shifted the memory footprint of LLMs during inference to the key-value cache (KV cache). In this paper, we propose HEADINFER, which offloads the KV cache to CPU RAM while avoiding the need to fully store the KV cache for any transformer layer on the GPU. HEADINFER employs a fine-grained, head-wise offloading strategy, maintaining only selective attention heads KV cache on the GPU while computing attention output dynamically. Through roofline analysis, we demonstrate that HEADINFER maintains computational efficiency while significantly reducing memory footprint. We evaluate HEADINFER on the Llama-3-8B model with a 1-million-token sequence, reducing the GPU memory footprint of the KV cache from 128 GB to 1 GB and the total GPU memory usage from 207 GB to 17 GB, achieving a 92% reduction compared to BF16 baseline inference. Notably, HEADINFER enables 4-million-token inference with an 8B model on a single consumer GPU with 24GB memory (e.g., NVIDIA RTX 4090) without approximation methods.

The application of Transformer-based large models has achieved numerous success in recent years. However, the exponential growth in the parameters of large models introduces formidable memory challenge for edge deployment. Prior works to address this challenge mainly focus on optimizing the model structure and adopting memory swapping methods. However, the former reduces the inference accuracy, and the latter raises the inference latency. This paper introduces PIPELOAD, a novel memory-efficient pipeline execution mechanism. It reduces memory usage by incorporating dynamic memory management and minimizes inference latency by employing parallel model loading. Based on PIPELOAD mechanism, we present Hermes, a framework optimized for large model inference on edge devices. We evaluate Hermes on Transformer-based models of different sizes. Our experiments illustrate that Hermes achieves up to 4.24 X increase in inference speed and 86.7% lower memory consumption than the state-of-the-art pipeline mechanism for BERT and ViT models, 2.58 X increase in inference speed and 90.3% lower memory consumption for GPT-style models.

Language model training in distributed settings is limited by the communication cost of gradient exchanges. In this short note, we extend recent work from Malladi et al. (2023), using shared randomness to perform distributed fine-tuning with low bandwidth. The method is a natural decentralized extension of memory-efficient Simultaneous Perturbation Stochastic Approximation (SPSA). Each iteration, each machine seeds a Random Number Generator (RNG) to perform local reproducible perturbations on model weights and calculate and exchange scalar projected gradients, which are then used to update each model. By using a (machine, sample) identifier as the random seed, each model can regenerate one another's perturbations. As machines only exchange single-byte projected gradients, this is highly communication efficient. There are also potential privacy benefits, as projected gradients may be calculated on different training data, and models never access the other's data. Our approach not only drastically reduces communication bandwidth requirements but also accommodates dynamic addition or removal of machines during the training process and retains the memory-efficient and inference-only advantages of recent work. We perform proof-of-concept experiments to demonstrate the potential usefulness of this method, building off of rich literature on distributed optimization and memory-efficient training.

Tiny deep learning on microcontroller units (MCUs) is challenging due to the limited memory size. We find that the memory bottleneck is due to the imbalanced memory distribution in convolutional neural network (CNN) designs: the first several blocks have an order of magnitude larger memory usage than the rest of the network. To alleviate this issue, we propose a generic patch-by-patch inference scheduling, which operates only on a small spatial region of the feature map and significantly cuts down the peak memory. However, naive implementation brings overlapping patches and computation overhead. We further propose network redistribution to shift the receptive field and FLOPs to the later stage and reduce the computation overhead. Manually redistributing the receptive field is difficult. We automate the process with neural architecture search to jointly optimize the neural architecture and inference scheduling, leading to MCUNetV2. Patch-based inference effectively reduces the peak memory usage of existing networks by 4-8x. Co-designed with neural networks, MCUNetV2 sets a record ImageNet accuracy on MCU (71.8%), and achieves >90% accuracy on the visual wake words dataset under only 32kB SRAM. MCUNetV2 also unblocks object detection on tiny devices, achieving 16.9% higher mAP on Pascal VOC compared to the state-of-the-art result. Our study largely addressed the memory bottleneck in tinyML and paved the way for various vision applications beyond image classification.

The performance of neural networks improves when more parameters are used. However, the model sizes are constrained by the available on-device memory during training and inference. Although applying techniques like quantization can alleviate the constraint, they suffer from performance degradation. In this work, we introduce NeuZip, a new weight compression scheme based on the entropy of floating-point numbers in neural networks. With NeuZip, we are able to achieve memory-efficient training and inference without sacrificing performance. Notably, we significantly reduce the memory footprint of training a Llama-3 8B model from 31GB to less than 16GB, while keeping the training dynamics fully unchanged. In inference, our method can reduce memory usage by more than half while maintaining near-lossless performance. Our code is publicly available.

Large deep learning models have achieved impressive performance across a range of applications. However, their large memory requirements, including parameter memory and activation memory, have become a significant challenge for their practical serving. While existing methods mainly address parameter memory, the importance of activation memory has been overlooked. Especially for long input sequences, activation memory is expected to experience a significant exponential growth as the length of sequences increases. In this approach, we propose AutoChunk, an automatic and adaptive compiler system that efficiently reduces activation memory for long sequence inference by chunk strategies. The proposed system generates chunk plans by optimizing through multiple stages. In each stage, the chunk search pass explores all possible chunk candidates and the chunk selection pass identifies the optimal one. At runtime, AutoChunk employs code generation to automatically apply chunk strategies. The experiments demonstrate that AutoChunk can reduce over 80\% of activation memory while maintaining speed loss within 10%, extend max sequence length by 3.2x to 11.7x, and outperform state-of-the-art methods by a large margin.

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

We present a novel and practical deep fully convolutional neural network architecture for semantic pixel-wise segmentation termed SegNet. This core trainable segmentation engine consists of an encoder network, a corresponding decoder network followed by a pixel-wise classification layer. The architecture of the encoder network is topologically identical to the 13 convolutional layers in the VGG16 network. The role of the decoder network is to map the low resolution encoder feature maps to full input resolution feature maps for pixel-wise classification. The novelty of SegNet lies is in the manner in which the decoder upsamples its lower resolution input feature map(s). Specifically, the decoder uses pooling indices computed in the max-pooling step of the corresponding encoder to perform non-linear upsampling. This eliminates the need for learning to upsample. The upsampled maps are sparse and are then convolved with trainable filters to produce dense feature maps. We compare our proposed architecture with the widely adopted FCN and also with the well known DeepLab-LargeFOV, DeconvNet architectures. This comparison reveals the memory versus accuracy trade-off involved in achieving good segmentation performance. SegNet was primarily motivated by scene understanding applications. Hence, it is designed to be efficient both in terms of memory and computational time during inference. It is also significantly smaller in the number of trainable parameters than other competing architectures. We also performed a controlled benchmark of SegNet and other architectures on both road scenes and SUN RGB-D indoor scene segmentation tasks. We show that SegNet provides good performance with competitive inference time and more efficient inference memory-wise as compared to other architectures. We also provide a Caffe implementation of SegNet and a web demo at http://mi.eng.cam.ac.uk/projects/segnet/.

Large model inference is shifting from cloud to edge due to concerns about the privacy of user interaction data. However, edge devices often struggle with limited computing power, memory, and bandwidth, requiring collaboration across multiple devices to run and speed up LLM inference. Pipeline parallelism, the mainstream solution, is inefficient for single-user scenarios, while tensor parallelism struggles with frequent communications. In this paper, we argue that tensor parallelism can be more effective than pipeline on low-resource devices, and present a compute- and memory-efficient tensor parallel inference system, named TPI-LLM, to serve 70B-scale models. TPI-LLM keeps sensitive raw data local in the users' devices and introduces a sliding window memory scheduler to dynamically manage layer weights during inference, with disk I/O latency overlapped with the computation and communication. This allows larger models to run smoothly on memory-limited devices. We analyze the communication bottleneck and find that link latency, not bandwidth, emerges as the main issue, so a star-based allreduce algorithm is implemented. Through extensive experiments on both emulated and real testbeds, TPI-LLM demonstrated over 80% less time-to-first-token and token latency compared to Accelerate, and over 90% compared to Transformers and Galaxy, while cutting the peak memory footprint of Llama 2-70B by 90%, requiring only 3.1 GB of memory for 70B-scale models.

Large language models (LLMs) are central to modern natural language processing, delivering exceptional performance in various tasks. However, their intensive computational and memory requirements present challenges, especially for devices with limited DRAM capacity. This paper tackles the challenge of efficiently running LLMs that exceed the available DRAM capacity by storing the model parameters on flash memory but bringing them on demand to DRAM. Our method involves constructing an inference cost model that harmonizes with the flash memory behavior, guiding us to optimize in two critical areas: reducing the volume of data transferred from flash and reading data in larger, more contiguous chunks. Within this flash memory-informed framework, we introduce two principal techniques. First, "windowing'" strategically reduces data transfer by reusing previously activated neurons, and second, "row-column bundling", tailored to the sequential data access strengths of flash memory, increases the size of data chunks read from flash memory. These methods collectively enable running models up to twice the size of the available DRAM, with a 4-5x and 20-25x increase in inference speed compared to naive loading approaches in CPU and GPU, respectively. Our integration of sparsity awareness, context-adaptive loading, and a hardware-oriented design paves the way for effective inference of LLMs on devices with limited memory.

The rapid progress in artificial intelligence-generated content (AIGC), especially with diffusion models, has significantly advanced development of high-quality video generation. However, current video diffusion models exhibit demanding computational requirements and high peak memory usage, especially for generating longer and higher-resolution videos. These limitations greatly hinder the practical application of video diffusion models on standard hardware platforms. To tackle this issue, we present a novel, training-free framework named Streamlined Inference, which leverages the temporal and spatial properties of video diffusion models. Our approach integrates three core components: Feature Slicer, Operator Grouping, and Step Rehash. Specifically, Feature Slicer effectively partitions input features into sub-features and Operator Grouping processes each sub-feature with a group of consecutive operators, resulting in significant memory reduction without sacrificing the quality or speed. Step Rehash further exploits the similarity between adjacent steps in diffusion, and accelerates inference through skipping unnecessary steps. Extensive experiments demonstrate that our approach significantly reduces peak memory and computational overhead, making it feasible to generate high-quality videos on a single consumer GPU (e.g., reducing peak memory of AnimateDiff from 42GB to 11GB, featuring faster inference on 2080Ti).

The auto-regressive decoding of Large Language Models (LLMs) results in significant overheads in their hardware performance. While recent research has investigated various speculative decoding techniques for multi-token generation, these efforts have primarily focused on improving processing speed such as throughput. Crucially, they often neglect other metrics essential for real-life deployments, such as memory consumption and training cost. To overcome these limitations, we propose a novel parallel prompt decoding that requires only 0.0002% trainable parameters, enabling efficient training on a single A100-40GB GPU in just 16 hours. Inspired by the human natural language generation process, PPD approximates outputs generated at future timesteps in parallel by using multiple prompt tokens. This approach partially recovers the missing conditional dependency information necessary for multi-token generation, resulting in up to a 28% higher acceptance rate for long-range predictions. Furthermore, we present a hardware-aware dynamic sparse tree technique that adaptively optimizes this decoding scheme to fully leverage the computational capacities on different GPUs. Through extensive experiments across LLMs ranging from MobileLlama to Vicuna-13B on a wide range of benchmarks, our approach demonstrates up to 2.49times speedup and maintains a minimal runtime memory overhead of just 0.0004%. More importantly, our parallel prompt decoding can serve as an orthogonal optimization for synergistic integration with existing speculative decoding, showing up to 1.22times further speed improvement. Our code is available at https://github.com/hmarkc/parallel-prompt-decoding.

Encoder-only transformer models such as BERT offer a great performance-size tradeoff for retrieval and classification tasks with respect to larger decoder-only models. Despite being the workhorse of numerous production pipelines, there have been limited Pareto improvements to BERT since its release. In this paper, we introduce ModernBERT, bringing modern model optimizations to encoder-only models and representing a major Pareto improvement over older encoders. Trained on 2 trillion tokens with a native 8192 sequence length, ModernBERT models exhibit state-of-the-art results on a large pool of evaluations encompassing diverse classification tasks and both single and multi-vector retrieval on different domains (including code). In addition to strong downstream performance, ModernBERT is also the most speed and memory efficient encoder and is designed for inference on common GPUs.

Visual Autoregressive (VAR) modeling has garnered significant attention for its innovative next-scale prediction approach, which yields substantial improvements in efficiency, scalability, and zero-shot generalization. Nevertheless, the coarse-to-fine methodology inherent in VAR results in exponential growth of the KV cache during inference, causing considerable memory consumption and computational redundancy. To address these bottlenecks, we introduce ScaleKV, a novel KV cache compression framework tailored for VAR architectures. ScaleKV leverages two critical observations: varying cache demands across transformer layers and distinct attention patterns at different scales. Based on these insights, ScaleKV categorizes transformer layers into two functional groups: drafters and refiners. Drafters exhibit dispersed attention across multiple scales, thereby requiring greater cache capacity. Conversely, refiners focus attention on the current token map to process local details, consequently necessitating substantially reduced cache capacity. ScaleKV optimizes the multi-scale inference pipeline by identifying scale-specific drafters and refiners, facilitating differentiated cache management tailored to each scale. Evaluation on the state-of-the-art text-to-image VAR model family, Infinity, demonstrates that our approach effectively reduces the required KV cache memory to 10% while preserving pixel-level fidelity.

The development of large language models (LLMs) has significantly expanded model sizes, resulting in substantial GPU memory requirements during inference. The key and value storage of the attention map in the KV (key-value) cache accounts for more than 80\% of this memory consumption. Nowadays, most existing KV cache compression methods focus on intra-layer compression within a single Transformer layer but few works consider layer-wise compression. In this paper, we propose a plug-and-play method called KVSharer, which shares the KV cache between layers to achieve layer-wise compression. Rather than intuitively sharing based on higher similarity, we discover a counterintuitive phenomenon: sharing dissimilar KV caches better preserves the model performance. Experiments show that KVSharer can reduce KV cache computation by 30\%, thereby lowering memory consumption without significantly impacting model performance and it can also achieve at least 1.3 times generation acceleration. Additionally, we verify that KVSharer is compatible with existing intra-layer KV cache compression methods, and combining both can further save memory.

Training-free acceleration has emerged as an advanced research area in video generation based on diffusion models. The redundancy of latents in diffusion model inference provides a natural entry point for acceleration. In this paper, we decompose the inference process into the encoding, denoising, and decoding stages, and observe that cache-based acceleration methods often lead to substantial memory surges in the latter two stages. To address this problem, we analyze the characteristics of inference across different stages and propose stage-specific strategies for reducing memory consumption: 1) Asynchronous Cache Swapping. 2) Feature chunk. 3) Slicing latents to decode. At the same time, we ensure that the time overhead introduced by these three strategies remains lower than the acceleration gains themselves. Compared with the baseline, our approach achieves faster inference speed and lower memory usage, while maintaining quality degradation within an acceptable range. The Code is available at https://github.com/NKUShaw/LightCache .

The recently released NLLB-200 is a set of multilingual Neural Machine Translation models that cover 202 languages. The largest model is based on a Mixture of Experts architecture and achieves SoTA results across many language pairs. It contains 54.5B parameters and requires at least four 32GB GPUs just for inference. In this work, we propose a pruning method that enables the removal of up to 80% of experts without further finetuning and with a negligible loss in translation quality, which makes it feasible to run the model on a single 32GB GPU. Further analysis suggests that our pruning metrics can identify language-specific experts.

Huge memory consumption has been a major bottleneck for deploying high-throughput large language models in real-world applications. In addition to the large number of parameters, the key-value (KV) cache for the attention mechanism in the transformer architecture consumes a significant amount of memory, especially when the number of layers is large for deep language models. In this paper, we propose a novel method that only computes and caches the KVs of a small number of layers, thus significantly saving memory consumption and improving inference throughput. Our experiments on large language models show that our method achieves up to 26times higher throughput than standard transformers and competitive performance in language modeling and downstream tasks. In addition, our method is orthogonal to existing transformer memory-saving techniques, so it is straightforward to integrate them with our model, achieving further improvement in inference efficiency. Our code is available at https://github.com/whyNLP/LCKV.

Diffusion models have achieved significant visual generation quality. However, their significant computational and memory costs pose challenge for their application on resource-constrained mobile devices or even desktop GPUs. Recent few-step diffusion models reduces the inference time by reducing the denoising steps. However, their memory consumptions are still excessive. The Post Training Quantization (PTQ) replaces high bit-width FP representation with low-bit integer values (INT4/8) , which is an effective and efficient technique to reduce the memory cost. However, when applying to few-step diffusion models, existing quantization methods face challenges in preserving both the image quality and text alignment. To address this issue, we propose an mixed-precision quantization framework - MixDQ. Firstly, We design specialized BOS-aware quantization method for highly sensitive text embedding quantization. Then, we conduct metric-decoupled sensitivity analysis to measure the sensitivity of each layer. Finally, we develop an integer-programming-based method to conduct bit-width allocation. While existing quantization methods fall short at W8A8, MixDQ could achieve W8A8 without performance loss, and W4A8 with negligible visual degradation. Compared with FP16, we achieve 3-4x reduction in model size and memory cost, and 1.45x latency speedup.

Large language models (LLMs) face the challenges in fine-tuning and deployment due to their high memory demands and computational costs. While parameter-efficient fine-tuning (PEFT) methods aim to reduce the memory usage of the optimizer state during fine-tuning, the inherent size of pre-trained LLM weights continues to be a pressing concern. Even though quantization techniques are widely proposed to ease memory demands and accelerate LLM inference, most of these techniques are geared towards the deployment phase. To bridge this gap, this paper presents Parameter-Efficient and Quantization-aware Adaptation (PEQA) - a simple yet effective method that combines the advantages of PEFT with quantized LLMs. By updating solely the quantization scales, PEQA can be directly applied to quantized LLMs, ensuring seamless task transitions. Parallel to existing PEFT methods, PEQA significantly reduces the memory overhead associated with the optimizer state. Furthermore, it leverages the advantages of quantization to substantially reduce model sizes. Even after fine-tuning, the quantization structure of a PEQA-tuned LLM remains intact, allowing for accelerated inference on the deployment stage. We employ PEQA-tuning for task-specific adaptation on LLMs with up to 65 billion parameters. To assess the logical reasoning and language comprehension of PEQA-tuned LLMs, we fine-tune low-bit quantized LLMs using a instruction dataset. Our results show that even when LLMs are quantized to below 4-bit precision, their capabilities in language modeling, few-shot in-context learning, and comprehension can be resiliently restored to (or even improved over) their full-precision original performances with PEQA.

Auto-regressive inference of transformers benefit greatly from Key-Value (KV) caching, but can lead to major memory bottlenecks as model size, batch size, and sequence length grow at scale. We introduce Multi-Layer Key-Value (MLKV) sharing, a novel approach extending KV sharing across transformer layers to reduce memory usage beyond what was possible with Multi-Query Attention (MQA) and Grouped-Query Attention (GQA). Evaluations on various NLP benchmarks and inference metrics using uptrained Pythia-160M variants demonstrate that MLKV significantly reduces memory usage with minimal performance loss, reducing KV cache size down to a factor of 6x compared to MQA. These results highlight MLKV's potential for efficient deployment of transformer models at scale. We provide code at https://github.com/zaydzuhri/pythia-mlkv

Model merging aggregates Large Language Models (LLMs) finetuned on different tasks into a stronger one. However, parameter conflicts between models leads to performance degradation in averaging. While model routing addresses this issue by selecting individual models during inference, it imposes excessive storage and compute costs, and fails to leverage the common knowledge from different models. In this work, we observe that different layers exhibit varying levels of parameter conflicts. Building on this insight, we average layers with minimal parameter conflicts and use a novel task-level expert routing for layers with significant conflicts. To further reduce storage costs, inspired by task arithmetic sparsity, we decouple multiple fine-tuned experts into a dense expert and several sparse experts. Considering the out-of-distribution samples, we select and merge appropriate experts based on the task uncertainty of the input data. We conduct extensive experiments on both LLaMA and Qwen with varying parameter scales, and evaluate on real-world reasoning tasks. Results demonstrate that our method consistently achieves significant performance improvements while requiring less system cost compared to existing methods.

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

The rise of billion-parameter diffusion models like Stable Diffusion XL, Imagen, and Dall-E3 markedly advances the field of generative AI. However, their large-scale nature poses challenges in fine-tuning and deployment due to high resource demands and slow inference speed. This paper ventures into the relatively unexplored yet promising realm of fine-tuning quantized diffusion models. We establish a strong baseline by customizing three models: PEQA for fine-tuning quantization parameters, Q-Diffusion for post-training quantization, and DreamBooth for personalization. Our analysis reveals a notable trade-off between subject and prompt fidelity within the baseline model. To address these issues, we introduce two strategies, inspired by the distinct roles of different timesteps in diffusion models: S1 optimizing a single set of fine-tuning parameters exclusively at selected intervals, and S2 creating multiple fine-tuning parameter sets, each specialized for different timestep intervals. Our approach not only enhances personalization but also upholds prompt fidelity and image quality, significantly outperforming the baseline qualitatively and quantitatively. The code will be made publicly available.

Large Reasoning Models (LRMs) significantly improve the reasoning ability of Large Language Models (LLMs) by learning to reason, exhibiting promising performance in complex task-solving. However, their deliberative reasoning process leads to inefficiencies in token usage, memory consumption, and inference time. Thus, this survey provides a review of efficient inference methods designed specifically for LRMs, focusing on mitigating token inefficiency while preserving the reasoning quality. First, we introduce a taxonomy to group the recent methods into two main categories: (a) explicit compact Chain-of-Thought (CoT), which reduces tokens while keeping the explicit reasoning structure, and (b) implicit latent CoT, which encodes reasoning steps within hidden representations instead of explicit tokens. Meanwhile, we discuss their strengths and weaknesses. Then, we conduct empirical analyses on existing methods from performance and efficiency aspects. Besides, we present open challenges in this field, including human-centric controllable reasoning, trade-off between interpretability and efficiency of reasoning, ensuring safety of efficient reasoning, and broader applications of efficient reasoning. In addition, we highlight key insights for enhancing LRMs' inference efficiency via techniques such as model merging, new architectures, and agent routers. We hope this work serves as a valuable guide, helping researchers overcome challenges in this vibrant fieldhttps://github.com/yueliu1999/Awesome-Efficient-Inference-for-LRMs.

Efficient inference in Large Language Models (LLMs) is impeded by the growing memory demands of key-value (KV) caching, especially for longer sequences. Traditional KV cache eviction strategies, which prioritize less critical KV-pairs based on attention scores, often degrade generation quality, leading to issues such as context loss or hallucinations. To address this, we introduce Dynamic Discriminative Operations (D2O), a novel method that utilizes two-level discriminative strategies to optimize KV cache size without fine-tuning, while preserving essential context. Initially, by observing varying densities of attention weights between shallow and deep layers, we use this insight to determine which layers should avoid excessive eviction to minimize information loss. Subsequently, for the eviction strategy in each layer, D2O innovatively incorporates a compensation mechanism that maintains a similarity threshold to re-discriminate the importance of previously discarded tokens, determining whether they should be recalled and merged with similar tokens. Our approach not only achieves significant memory savings and enhances inference throughput by more than 3 times but also maintains high-quality long-text generation. Extensive experiments across various benchmarks and LLM architectures have demonstrated that D2O significantly enhances performance with a constrained KV cache budget.

We introduce ProVideLLM, an end-to-end framework for real-time procedural video understanding. ProVideLLM integrates a multimodal cache configured to store two types of tokens - verbalized text tokens, which provide compressed textual summaries of long-term observations, and visual tokens, encoded with DETR-QFormer to capture fine-grained details from short-term observations. This design reduces token count by 22x over existing methods in representing one hour of long-term observations while effectively encoding fine-granularity of the present. By interleaving these tokens in our multimodal cache, ProVideLLM ensures sub-linear scaling of memory and compute with video length, enabling per-frame streaming inference at 10 FPS and streaming dialogue at 25 FPS, with a minimal 2GB GPU memory footprint. ProVideLLM also sets new state-of-the-art results on six procedural tasks across four datasets.

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

With the increasing diversity of ML infrastructures nowadays, distributed training over heterogeneous computing systems is desired to facilitate the production of big models. Mixture-of-Experts (MoE) models have been proposed to lower the cost of training subject to the overall size of models/data through gating and parallelism in a divide-and-conquer fashion. While DeepSpeed has made efforts in carrying out large-scale MoE training over heterogeneous infrastructures, the efficiency of training and inference could be further improved from several system aspects, including load balancing, communication/computation efficiency, and memory footprint limits. In this work, we present SE-MoE that proposes Elastic MoE training with 2D prefetch and Fusion communication over Hierarchical storage, so as to enjoy efficient parallelisms in various types. For scalable inference in a single node, especially when the model size is larger than GPU memory, SE-MoE forms the CPU-GPU memory jointly into a ring of sections to load the model, and executes the computation tasks across the memory sections in a round-robin manner for efficient inference. We carried out extensive experiments to evaluate SE-MoE, where SE-MoE successfully trains a Unified Feature Optimization (UFO) model with a Sparsely-Gated Mixture-of-Experts model of 12B parameters in 8 days on 48 A100 GPU cards. The comparison against the state-of-the-art shows that SE-MoE outperformed DeepSpeed with 33% higher throughput (tokens per second) in training and 13% higher throughput in inference in general. Particularly, under unbalanced MoE Tasks, e.g., UFO, SE-MoE achieved 64% higher throughput with 18% lower memory footprints. The code of the framework will be released on: https://github.com/PaddlePaddle/Paddle.

Semantic top-K selection with cross-encoder rerankers underpins of on-device AI services, such as retrieval-augmented generation, agent memory, and personalized recommendation. However, its latency and memory demands dominate end-to-end budgets on edge hardware. Revisiting the objective of top-K selection, we reveal that only relative rankings matter, not exact per-candidate scores. We further observe sequence-level sparsity: relative rankings stabilize early in intermediate layers, allowing pruning opportunities prior to completing full inference. Building on this insight, we propose monolithic forwarding and develop a training-free inference system, GRATING. By maintaining a global view of all candidates, it reduces latency through progressive cluster pruning. It also bounds peak memory usage by strategically overlapping I/O with computation via dual-layer sliding window and chunked execution. We evaluate GRATING against state-of-the-art baselines on rerankers from 0.6B to 8B parameters across Apple M2 and RTX 5070. GRATING consistently reduces latency by up to 89.0% and peak memory by up to 94.9% in microbenchmarks, without any loss in precision. Across three real-world on-device AI applications, GRATING lowers latency by 11.6%-51.0% and peak memory by 18.6%-77.8%, demonstrating substantial improvements in efficiency and deployability.

In the field of instruction-following large vision-language models (LVLMs), the efficient deployment of these models faces challenges, notably due to the high memory demands of their key-value (KV) caches. Conventional cache management strategies for LLMs focus on cache eviction, which often fails to address the specific needs of multimodal instruction-following models. Recognizing this gap, in this paper, we introduce Elastic Cache, a novel approach that benefits from applying distinct acceleration methods for instruction encoding and output generation stages. We investigate the metrics of importance in different stages and propose an importance-driven cache merging strategy to prune redundancy caches. Instead of discarding less important caches, our strategy identifies important key/value vectors as anchor points. Surrounding less important caches are then merged with these anchors, enhancing the preservation of contextual information in the KV caches while yielding an arbitrary acceleration ratio. For instruction encoding, we utilize the frequency to evaluate the importance of caches. Regarding output generation, we prioritize tokens based on their distance with an offset, by which both the initial and most recent tokens are retained. Results on a range of LVLMs demonstrate that Elastic Cache not only boosts efficiency but also notably outperforms existing pruning methods in language generation across various tasks. Code is available at https://github.com/liuzuyan/ElasticCache

Retrieval-augmented generation (RAG) extends large language models (LLMs) with external data sources to enhance factual correctness and domain coverage. Modern RAG pipelines rely on large datastores, leading to system challenges in latency-sensitive deployments, especially when limited GPU memory is available. To address these challenges, we propose TeleRAG, an efficient inference system that reduces RAG latency with minimal GPU memory requirements. The core innovation of TeleRAG is lookahead retrieval, a prefetching mechanism that anticipates required data and transfers it from CPU to GPU in parallel with LLM generation. By leveraging the modularity of RAG pipelines, the inverted file index (IVF) search algorithm and similarities between queries, TeleRAG optimally overlaps data movement and computation. Experimental results show that TeleRAG reduces end-to-end RAG inference latency by up to 1.72x on average compared to state-of-the-art systems, enabling faster, more memory-efficient deployments of advanced RAG applications.

Large Language Models (LLMs) have attracted extensive attention due to their remarkable performance across various tasks. However, the substantial computational and memory requirements of LLM inference pose challenges for deployment in resource-constrained scenarios. Efforts within the field have been directed towards developing techniques aimed at enhancing the efficiency of LLM inference. This paper presents a comprehensive survey of the existing literature on efficient LLM inference. We start by analyzing the primary causes of the inefficient LLM inference, i.e., the large model size, the quadratic-complexity attention operation, and the auto-regressive decoding approach. Then, we introduce a comprehensive taxonomy that organizes the current literature into data-level, model-level, and system-level optimization. Moreover, the paper includes comparative experiments on representative methods within critical sub-fields to provide quantitative insights. Last but not least, we provide some knowledge summary and discuss future research directions.

Convolutional neural networks (CNNs) are the current state-of-the-art meta-algorithm for volumetric segmentation of medical data, for example, to localize COVID-19 infected tissue on computer tomography scans or the detection of tumour volumes in magnetic resonance imaging. A key limitation of 3D CNNs on voxelised data is that the memory consumption grows cubically with the training data resolution. Occupancy networks (O-Nets) are an alternative for which the data is represented continuously in a function space and 3D shapes are learned as a continuous decision boundary. While O-Nets are significantly more memory efficient than 3D CNNs, they are limited to simple shapes, are relatively slow at inference, and have not yet been adapted for 3D semantic segmentation of medical data. Here, we propose Occupancy Networks for Semantic Segmentation (OSS-Nets) to accurately and memory-efficiently segment 3D medical data. We build upon the original O-Net with modifications for increased expressiveness leading to improved segmentation performance comparable to 3D CNNs, as well as modifications for faster inference. We leverage local observations to represent complex shapes and prior encoder predictions to expedite inference. We showcase OSS-Net's performance on 3D brain tumour and liver segmentation against a function space baseline (O-Net), a performance baseline (3D residual U-Net), and an efficiency baseline (2D residual U-Net). OSS-Net yields segmentation results similar to the performance baseline and superior to the function space and efficiency baselines. In terms of memory efficiency, OSS-Net consumes comparable amounts of memory as the function space baseline, somewhat more memory than the efficiency baseline and significantly less than the performance baseline. As such, OSS-Net enables memory-efficient and accurate 3D semantic segmentation that can scale to high resolutions.

Large Language Models (LLMs), while demonstrating remarkable capabilities across various applications, present significant challenges during inference due to their substantial model size, especially when deployed on edge devices. Activation sparsity offers a promising solution to reduce computation and memory movement, enabling more efficient inference, particularly for small-batch on-device applications. However, current approaches face limitations with non-ReLU activation function, which are foundational to most advanced LLMs, or require heavy continual training. Additionally, the difficulty in predicting active channels and limited achievable sparsity ratios constrain the effectiveness of activation sparsity-based methods. In this paper, we introduce R-Sparse, a training-free activation sparsity approach capable of achieving high sparsity levels in advanced LLMs. We conducted two preliminary investigations into how different components contribute to the output within a single linear layer and found two key observations: (i) the non-sparse components of the input function can be regarded as a few bias terms, and (ii) The full computation can be effectively approximated by an appropriate combination of input channels and weight singular values. Building on this, we replace the linear layers in LLMs with a rank-aware sparse inference method that leverages the sparsity of input channels and singular value components, eliminating the need for active channel prediction like the output sparsity based approaches. Experiments on Llama-2/3 and Mistral models across ten diverse tasks demonstrate that R-Sparse achieves comparable performance at 50% model-level sparsity, resulting in a significant 43% end-to-end efficient improvements with customized kernels.

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.

The past several years have witnessed the success of transformer-based models, and their scale and application scenarios continue to grow aggressively. The current landscape of transformer models is increasingly diverse: the model size varies drastically with the largest being of hundred-billion parameters; the model characteristics differ due to the sparsity introduced by the Mixture-of-Experts; the target application scenarios can be latency-critical or throughput-oriented; the deployment hardware could be single- or multi-GPU systems with different types of memory and storage, etc. With such increasing diversity and the fast-evolving pace of transformer models, designing a highly performant and efficient inference system is extremely challenging. In this paper, we present DeepSpeed Inference, a comprehensive system solution for transformer model inference to address the above-mentioned challenges. DeepSpeed Inference consists of (1) a multi-GPU inference solution to minimize latency while maximizing the throughput of both dense and sparse transformer models when they fit in aggregate GPU memory, and (2) a heterogeneous inference solution that leverages CPU and NVMe memory in addition to the GPU memory and compute to enable high inference throughput with large models which do not fit in aggregate GPU memory. DeepSpeed Inference reduces latency by up to 7.3X over the state-of-the-art for latency-oriented scenarios and increases throughput by over 1.5x for throughput-oriented scenarios. Moreover, it enables trillion parameter scale inference under real-time latency constraints by leveraging hundreds of GPUs, an unprecedented scale for inference. It can inference 25x larger models than with GPU-only solutions, while delivering a high throughput of 84 TFLOPS (over 50% of A6000 peak).

This paper presents SANA-1.5, a linear Diffusion Transformer for efficient scaling in text-to-image generation. Building upon SANA-1.0, we introduce three key innovations: (1) Efficient Training Scaling: A depth-growth paradigm that enables scaling from 1.6B to 4.8B parameters with significantly reduced computational resources, combined with a memory-efficient 8-bit optimizer. (2) Model Depth Pruning: A block importance analysis technique for efficient model compression to arbitrary sizes with minimal quality loss. (3) Inference-time Scaling: A repeated sampling strategy that trades computation for model capacity, enabling smaller models to match larger model quality at inference time. Through these strategies, SANA-1.5 achieves a text-image alignment score of 0.72 on GenEval, which can be further improved to 0.80 through inference scaling, establishing a new SoTA on GenEval benchmark. These innovations enable efficient model scaling across different compute budgets while maintaining high quality, making high-quality image generation more accessible.

Open-source foundation models have seen rapid adoption and development, enabling powerful general-purpose capabilities across diverse domains. However, fine-tuning large foundation models for domain-specific or personalized tasks remains prohibitively expensive for most users due to the significant memory overhead beyond that of inference. We introduce EMLoC, an Emulator-based Memory-efficient fine-tuning framework with LoRA Correction, which enables model fine-tuning within the same memory budget required for inference. EMLoC constructs a task-specific light-weight emulator using activation-aware singular value decomposition (SVD) on a small downstream calibration set. Fine-tuning then is performed on this lightweight emulator via LoRA. To tackle the misalignment between the original model and the compressed emulator, we propose a novel compensation algorithm to correct the fine-tuned LoRA module, which thus can be merged into the original model for inference. EMLoC supports flexible compression ratios and standard training pipelines, making it adaptable to a wide range of applications. Extensive experiments demonstrate that EMLoC outperforms other baselines across multiple datasets and modalities. Moreover, without quantization, EMLoC enables fine-tuning of a 38B model on a single 24GB consumer GPU-bringing efficient and practical model adaptation to individual users.

Industrial anomaly detection is an important task within computer vision with a wide range of practical use cases. The small size of anomalous regions in many real-world datasets necessitates processing the images at a high resolution. This frequently poses significant challenges concerning memory consumption during the model training and inference stages, leaving some existing methods impractical for widespread adoption. To overcome this challenge, we present the tiled ensemble approach, which reduces memory consumption by dividing the input images into a grid of tiles and training a dedicated model for each tile location. The tiled ensemble is compatible with any existing anomaly detection model without the need for any modification of the underlying architecture. By introducing overlapping tiles, we utilize the benefits of traditional stacking ensembles, leading to further improvements in anomaly detection capabilities beyond high resolution alone. We perform a comprehensive analysis using diverse underlying architectures, including Padim, PatchCore, FastFlow, and Reverse Distillation, on two standard anomaly detection datasets: MVTec and VisA. Our method demonstrates a notable improvement across setups while remaining within GPU memory constraints, consuming only as much GPU memory as a single model needs to process a single tile.

Diffusion models have shown remarkable performance in image generation in recent years. However, due to a quadratic increase in memory during generating ultra-high-resolution images (e.g. 4096*4096), the resolution of generated images is often limited to 1024*1024. In this work. we propose a unidirectional block attention mechanism that can adaptively adjust the memory overhead during the inference process and handle global dependencies. Building on this module, we adopt the DiT structure for upsampling and develop an infinite super-resolution model capable of upsampling images of various shapes and resolutions. Comprehensive experiments show that our model achieves SOTA performance in generating ultra-high-resolution images in both machine and human evaluation. Compared to commonly used UNet structures, our model can save more than 5x memory when generating 4096*4096 images. The project URL is https://github.com/THUDM/Inf-DiT.

Generative transformer models have become increasingly complex, with large numbers of parameters and the ability to process multiple input modalities. Current methods for explaining their predictions are resource-intensive. Most crucially, they require prohibitively large amounts of extra memory, since they rely on backpropagation which allocates almost twice as much GPU memory as the forward pass. This makes it difficult, if not impossible, to use them in production. We present AtMan that provides explanations of generative transformer models at almost no extra cost. Specifically, AtMan is a modality-agnostic perturbation method that manipulates the attention mechanisms of transformers to produce relevance maps for the input with respect to the output prediction. Instead of using backpropagation, AtMan applies a parallelizable token-based search method based on cosine similarity neighborhood in the embedding space. Our exhaustive experiments on text and image-text benchmarks demonstrate that AtMan outperforms current state-of-the-art gradient-based methods on several metrics while being computationally efficient. As such, AtMan is suitable for use in large model inference deployments.

The linear growth of key-value (KV) cache memory and quadratic computational complexity pose significant bottlenecks for large language models (LLMs) in long-context processing. While existing KV cache optimization methods address these challenges through token pruning or feature merging, they often suffer from irreversible information loss or require costly parameter retraining. We propose ZeroMerge, a dynamic zero-shot compression framework that achieves efficient cache management through three key innovations: (1) Fine-grained memory allocation guided by multi-dimensional token importance metrics at head-level granularity, (2) A residual merging mechanism that preserves critical context through compensated attention scoring, and (3) Parameter-free adaptation compatible with diverse LLM architectures without retraining. Comprehensive evaluations across LLaMA-2 model demonstrate that ZeroMerge maintains full-cache performance at 5\% compression ratios while doubling inference throughput at 40K token lengths. The method effectively balances memory efficiency, generation quality, and deployment flexibility, advancing practical long-context LLM applications. The code is available at https://github.com/SusCom-Lab/ZeroMerge.

Recent advances in self-supervised learning and the Transformer architecture have significantly improved natural language processing (NLP), achieving remarkably low perplexity. However, the growing size of NLP models introduces a memory wall problem during the generation phase. To mitigate this issue, recent efforts have focused on quantizing model weights to sub-4-bit precision while preserving full precision for activations, resulting in practical speed-ups during inference on a single GPU. However, these improvements primarily stem from reduced memory movement, which necessitates a resource-intensive dequantization process rather than actual computational reduction. In this paper, we introduce LUT-GEMM, an efficient kernel for quantized matrix multiplication, which not only eliminates the resource-intensive dequantization process but also reduces computational costs compared to previous kernels for weight-only quantization. Furthermore, we proposed group-wise quantization to offer a flexible trade-off between compression ratio and accuracy. The impact of LUT-GEMM is facilitated by implementing high compression ratios through low-bit quantization and efficient LUT-based operations. We show experimentally that when applied to the OPT-175B model with 3-bit quantization, LUT-GEMM substantially accelerates token generation latency, achieving a remarkable 2.1times improvement on a single GPU when compared to OPTQ, which relies on the costly dequantization process.

Large-scale text encoders in text-to-image (T2I) diffusion models have demonstrated exceptional performance in generating high-quality images from textual prompts. Unlike denoising modules that rely on multiple iterative steps, text encoders require only a single forward pass to produce text embeddings. However, despite their minimal contribution to total inference time and floating-point operations (FLOPs), text encoders demand significantly higher memory usage, up to eight times more than denoising modules. To address this inefficiency, we propose Skip and Re-use layers (Skrr), a simple yet effective pruning strategy specifically designed for text encoders in T2I diffusion models. Skrr exploits the inherent redundancy in transformer blocks by selectively skipping or reusing certain layers in a manner tailored for T2I tasks, thereby reducing memory consumption without compromising performance. Extensive experiments demonstrate that Skrr maintains image quality comparable to the original model even under high sparsity levels, outperforming existing blockwise pruning methods. Furthermore, Skrr achieves state-of-the-art memory efficiency while preserving performance across multiple evaluation metrics, including the FID, CLIP, DreamSim, and GenEval scores.

Recent breakthroughs in solving reasoning, math and coding problems with Large Language Models (LLMs) have been enabled by investing substantial computation budgets at inference time. Therefore, inference speed is one of the most critical properties of LLM architectures, and there is a growing need for LLMs that are efficient and fast at inference. Recently, LLMs built on the xLSTM architecture have emerged as a powerful alternative to Transformers, offering linear compute scaling with sequence length and constant memory usage, both highly desirable properties for efficient inference. However, such xLSTM-based LLMs have yet to be scaled to larger models and assessed and compared with respect to inference speed and efficiency. In this work, we introduce xLSTM 7B, a 7-billion-parameter LLM that combines xLSTM's architectural benefits with targeted optimizations for fast and efficient inference. Our experiments demonstrate that xLSTM 7B achieves performance on downstream tasks comparable to other similar-sized LLMs, while providing significantly faster inference speeds and greater efficiency compared to Llama- and Mamba-based LLMs. These results establish xLSTM 7B as the fastest and most efficient 7B LLM, offering a solution for tasks that require large amounts of test-time computation. Our work highlights xLSTM's potential as a foundational architecture for methods building on heavy use of LLM inference. Our model weights, model code and training code are open-source.

Modern deep learning-based recommendation systems exploit hundreds to thousands of different categorical features, each with millions of different categories ranging from clicks to posts. To respect the natural diversity within the categorical data, embeddings map each category to a unique dense representation within an embedded space. Since each categorical feature could take on as many as tens of millions of different possible categories, the embedding tables form the primary memory bottleneck during both training and inference. We propose a novel approach for reducing the embedding size in an end-to-end fashion by exploiting complementary partitions of the category set to produce a unique embedding vector for each category without explicit definition. By storing multiple smaller embedding tables based on each complementary partition and combining embeddings from each table, we define a unique embedding for each category at smaller memory cost. This approach may be interpreted as using a specific fixed codebook to ensure uniqueness of each category's representation. Our experimental results demonstrate the effectiveness of our approach over the hashing trick for reducing the size of the embedding tables in terms of model loss and accuracy, while retaining a similar reduction in the number of parameters.

High-resolution Vision-Language Models (VLMs) have been widely used in multimodal tasks to enhance accuracy by preserving detailed image information. However, these models often generate excessive visual tokens due to encoding multiple partitions of the input image. Processing these excessive visual tokens is computationally challenging, especially in resource-constrained environments with commodity GPUs. To support high-resolution images while meeting resource constraints, we propose High-Resolution Early Dropping (HiRED), a token-dropping scheme that operates within a fixed token budget before the Large Language Model (LLM) stage. HiRED can be integrated with existing high-resolution VLMs in a plug-and-play manner, as it requires no additional training while still maintaining superior accuracy. We strategically use the vision encoder's attention in the initial layers to assess the visual content of each image partition and allocate the token budget accordingly. Then, using the attention in the final layer, we select the most important visual tokens from each partition within the allocated budget, dropping the rest. Empirically, when applied to LLaVA-Next-7B on NVIDIA TESLA P40 GPU, HiRED with a 20% token budget increases token generation throughput by 4.7, reduces first-token generation latency by 15 seconds, and saves 2.3 GB of GPU memory for a single inference.

Large Language Models (LLMs) have significantly advanced natural language processing with exceptional task generalization capabilities. Low-Rank Adaption (LoRA) offers a cost-effective fine-tuning solution, freezing the original model parameters and training only lightweight, low-rank adapter matrices. However, the memory footprint of LoRA is largely dominated by the original model parameters. To mitigate this, we propose LoRAM, a memory-efficient LoRA training scheme founded on the intuition that many neurons in over-parameterized LLMs have low training utility but are essential for inference. LoRAM presents a unique twist: it trains on a pruned (small) model to obtain pruned low-rank matrices, which are then recovered and utilized with the original (large) model for inference. Additionally, minimal-cost continual pre-training, performed by the model publishers in advance, aligns the knowledge discrepancy between pruned and original models. Our extensive experiments demonstrate the efficacy of LoRAM across various pruning strategies and downstream tasks. For a model with 70 billion parameters, LoRAM enables training on a GPU with only 20G HBM, replacing an A100-80G GPU for LoRA training and 15 GPUs for full fine-tuning. Specifically, QLoRAM implemented by structured pruning combined with 4-bit quantization, for LLaMA-3.1-70B (LLaMA-2-70B), reduces the parameter storage cost that dominates the memory usage in low-rank matrix training by 15.81times (16.95times), while achieving dominant performance gains over both the original LLaMA-3.1-70B (LLaMA-2-70B) and LoRA-trained LLaMA-3.1-8B (LLaMA-2-13B).

The efficiency of large vision-language models (LVLMs) is constrained by the computational bottleneck of the attention mechanism during the prefill phase and the memory bottleneck of fetching the key-value (KV) cache in the decoding phase, particularly in scenarios involving high-resolution images or videos. Visual content often exhibits substantial redundancy, resulting in highly sparse attention maps within LVLMs. This sparsity can be leveraged to accelerate attention computation or compress the KV cache through various approaches. However, most studies focus on addressing only one of these bottlenecks and do not adequately support dynamic adjustment of sparsity concerning distinct layers or tasks. In this paper, we present ZipVL, an efficient inference framework designed for LVLMs that resolves both computation and memory bottlenecks through a dynamic ratio allocation strategy of important tokens. This ratio is adaptively determined based on the layer-specific distribution of attention scores, rather than fixed hyper-parameters, thereby improving efficiency for less complex tasks while maintaining high performance for more challenging ones. Then we select important tokens based on their normalized attention scores and perform attention mechanism solely on those important tokens to accelerate the prefill phase. To mitigate the memory bottleneck in the decoding phase, we employ mixed-precision quantization to the KV cache, where high-bit quantization is used for caches of important tokens, while low-bit quantization is applied to those of less importance. Our experiments demonstrate that ZipVL can accelerate the prefill phase by 2.6times and reduce GPU memory usage by 50.0%, with a minimal accuracy reduction of only 0.2% on Video-MME benchmark over LongVA-7B model, effectively enhancing the generation efficiency of LVLMs.

We present a very simple algorithm for attention that requires O(1) memory with respect to sequence length and an extension to self-attention that requires O(log n) memory. This is in contrast with the frequently stated belief that self-attention requires O(n^2) memory. While the time complexity is still O(n^2), device memory rather than compute capability is often the limiting factor on modern accelerators. Thus, reducing the memory requirements of attention allows processing of longer sequences than might otherwise be feasible. We provide a practical implementation for accelerators that requires O(n) memory, is numerically stable, and is within a few percent of the runtime of the standard implementation of attention. We also demonstrate how to differentiate the function while remaining memory-efficient. For sequence length 16384, the memory overhead of self-attention is reduced by 59X for inference and by 32X for differentiation.

Large pre-trained models (LPMs), such as LLaMA and GLM, have shown exceptional performance across various tasks through fine-tuning. Although low-rank adaption (LoRA) has emerged to cheaply fine-tune these LPMs on downstream tasks, their deployment is still hindered by the vast model scale and computational costs. Neural network pruning offers a way to compress LPMs. However, the current pruning methods designed for LPMs are not compatible with LoRA. This is due to their utilization of unstructured pruning on LPMs, impeding the merging of LoRA weights, or their dependence on the gradients of pre-trained weights to guide pruning, which can impose significant memory overhead. To this end, we propose LoRAPrune, a new framework that delivers an accurate, compact model for efficient inference in a highly memory-effective manner. Specifically, we first design a LoRA-guided pruning criterion, which uses the weights and gradients of LoRA, rather than the gradients of pre-trained weights for importance estimation. We then propose a structured iterative pruning procedure, to remove redundant channels and heads. Extensive experimental results demonstrate the superior performance of our LoRAPrune over existing approaches on the LLaMA series models. For instance, at a 50\% compression rate, LoRAPrune outperforms LLM-Pruner by a perplexity reduction of 8.0 on WikiText2 and 16.05 on PTB datasets, while concurrently reducing memory usage by 52.6\%. The code will be released after review

Mixture-of-experts (MoE) architectures enable scaling large language models (LLMs) to vast parameter counts without a proportional rise in computational costs. However, the significant memory demands of large MoE models hinder their deployment across various computational environments, from cloud servers to consumer devices. This study first demonstrates pronounced task-specific specialization in expert activation patterns within MoE layers. Building on this, we introduce PreMoe, a novel framework that enables efficient deployment of massive MoE models in memory-constrained environments. PreMoe features two main components: probabilistic expert pruning (PEP) and task-adaptive expert retrieval (TAER). PEP employs a new metric, the task-conditioned expected selection score (TCESS), derived from router logits to quantify expert importance for specific tasks, thereby identifying a minimal set of critical experts. TAER leverages these task-specific expert importance profiles for efficient inference. It pre-computes and stores compact expert patterns for diverse tasks. When a user query is received, TAER rapidly identifies the most relevant stored task pattern and reconstructs the model by loading only the small subset of experts crucial for that task. This approach dramatically reduces the memory footprint across all deployment scenarios. DeepSeek-R1 671B maintains 97.2\% accuracy on MATH500 when pruned to 8/128 configuration (50\% expert reduction), and still achieves 72.0\% with aggressive 8/32 pruning (87.5\% expert reduction). Pangu-Ultra-MoE 718B achieves 97.15\% on MATH500 and 81.3\% on AIME24 with 8/128 pruning, while even more aggressive pruning to 4/64 (390GB memory) preserves 96.95\% accuracy on MATH500. We make our code publicly available at https://github.com/JarvisPei/PreMoe.

We introduce RecurrentGemma, an open language model which uses Google's novel Griffin architecture. Griffin combines linear recurrences with local attention to achieve excellent performance on language. It has a fixed-sized state, which reduces memory use and enables efficient inference on long sequences. We provide a pre-trained model with 2B non-embedding parameters, and an instruction tuned variant. Both models achieve comparable performance to Gemma-2B despite being trained on fewer tokens.

We present Mobile Video Networks (MoViNets), a family of computation and memory efficient video networks that can operate on streaming video for online inference. 3D convolutional neural networks (CNNs) are accurate at video recognition but require large computation and memory budgets and do not support online inference, making them difficult to work on mobile devices. We propose a three-step approach to improve computational efficiency while substantially reducing the peak memory usage of 3D CNNs. First, we design a video network search space and employ neural architecture search to generate efficient and diverse 3D CNN architectures. Second, we introduce the Stream Buffer technique that decouples memory from video clip duration, allowing 3D CNNs to embed arbitrary-length streaming video sequences for both training and inference with a small constant memory footprint. Third, we propose a simple ensembling technique to improve accuracy further without sacrificing efficiency. These three progressive techniques allow MoViNets to achieve state-of-the-art accuracy and efficiency on the Kinetics, Moments in Time, and Charades video action recognition datasets. For instance, MoViNet-A5-Stream achieves the same accuracy as X3D-XL on Kinetics 600 while requiring 80% fewer FLOPs and 65% less memory. Code will be made available at https://github.com/tensorflow/models/tree/master/official/vision.

The growing context lengths of large language models (LLMs) pose significant challenges for efficient inference, primarily due to GPU memory and bandwidth constraints. We present RetroInfer, a novel system that reconceptualizes the key-value (KV) cache as a vector storage system which exploits the inherent attention sparsity to accelerate long-context LLM inference. At its core is the wave index, an Attention-aWare VEctor index that enables efficient and accurate retrieval of critical tokens through techniques such as tripartite attention approximation, accuracy-bounded attention estimation, and segmented clustering. Complementing this is the wave buffer, which coordinates KV cache placement and overlaps computation and data transfer across GPU and CPU to sustain high throughput. Unlike prior sparsity-based methods that struggle with token selection and hardware coordination, RetroInfer delivers robust performance without compromising model accuracy. Experiments on long-context benchmarks show up to 4.5X speedup over full attention within GPU memory limits and up to 10.5X over sparse attention baselines when KV cache is extended to CPU memory, all while preserving full-attention-level accuracy.

As Large Language Models (LLMs) can now process extremely long contexts, efficient inference over these extended inputs has become increasingly important, especially for emerging applications like LLM agents that highly depend on this capability. Speculative decoding (SD) offers a promising lossless acceleration technique compared to lossy alternatives such as quantization and model cascades. However, most state-of-the-art SD methods are trained on short texts (typically fewer than 4k tokens), making them unsuitable for long-context scenarios. Specifically, adapting these methods to long contexts presents three key challenges: (1) the excessive memory demands posed by draft models due to large Key-Value (KV) cache; (2) performance degradation resulting from the mismatch between short-context training and long-context inference; and (3) inefficiencies in tree attention mechanisms when managing long token sequences. This work introduces LongSpec, a framework that addresses these challenges through three core innovations: a memory-efficient draft model with a constant-sized KV cache; novel position indices that mitigate the training-inference mismatch; and an attention aggregation strategy that combines fast prefix computation with standard tree attention to enable efficient decoding. Experimental results confirm the effectiveness of LongSpec, achieving up to a 3.26x speedup over strong Flash Attention baselines across five long-context understanding datasets, as well as a 2.25x reduction in wall-clock time on the AIME24 long reasoning task with the QwQ model, demonstrating significant latency improvements for long-context applications. The code is available at https://github.com/sail-sg/LongSpec.

Powerful foundation models, including large language models (LLMs), with Transformer architectures have ushered in a new era of Generative AI across various industries. Industry and research community have witnessed a large number of new applications, based on those foundation models. Such applications include question and answer, customer services, image and video generation, and code completions, among others. However, as the number of model parameters reaches to hundreds of billions, their deployment incurs prohibitive inference costs and high latency in real-world scenarios. As a result, the demand for cost-effective and fast inference using AI accelerators is ever more higher. To this end, our tutorial offers a comprehensive discussion on complementary inference optimization techniques using AI accelerators. Beginning with an overview of basic Transformer architectures and deep learning system frameworks, we deep dive into system optimization techniques for fast and memory-efficient attention computations and discuss how they can be implemented efficiently on AI accelerators. Next, we describe architectural elements that are key for fast transformer inference. Finally, we examine various model compression and fast decoding strategies in the same context.

Autoregressive decoding of large language models (LLMs) is memory bandwidth bounded, resulting in high latency and significant wastes of the parallel processing power of modern accelerators. Existing methods for accelerating LLM decoding often require a draft model (e.g., speculative decoding), which is nontrivial to obtain and unable to generalize. In this paper, we introduce Lookahead decoding, an exact, parallel decoding algorithm that accelerates LLM decoding without needing auxiliary models or data stores. It allows trading per-step log(FLOPs) to reduce the number of total decoding steps, is more parallelizable on single or multiple modern accelerators, and is compatible with concurrent memory-efficient attention (e.g., FlashAttention). Our implementation of Lookahead decoding can speed up autoregressive decoding by up to 1.8x on MT-bench and 4x with strong scaling on multiple GPUs in code completion tasks. Our code is avialable at https://github.com/hao-ai-lab/LookaheadDecoding

Contextual memory integration remains a high challenge in the development of language models, particularly in tasks that require maintaining coherence over extended sequences. Traditional approaches, such as self-attention mechanisms and memory-augmented architectures, often prioritize short-term dependencies, leading to fragmentation and inconsistency in long-range contextual understanding. Inspired by principles of synaptic plasticity observed in biological neural systems, a novel mechanism, Synaptic Resonance, is introduced to dynamically reinforce relevant memory pathways during training and inference. Unlike static memory representations, this mechanism continuously adjusts synaptic weight matrices based on contextual relevance, allowing for improved information retention without excessive computational overhead. Evaluations conducted on an open-source language model demonstrate reductions in perplexity, enhancements in contextual coherence, and increased robustness against input noise, highlighting the effectiveness of reinforcement-driven memory modulation. Comparative analysis against baseline models further reveals that the proposed approach achieves higher memory retention efficiency while maintaining computational feasibility. The architectural modifications integrate seamlessly into existing transformer-based frameworks, ensuring stable convergence and efficient inference without sacrificing scalability. Applications benefiting from improved long-term contextual consistency, such as dialogue systems and document summarization, stand to gain from this approach. Empirical findings suggest that dynamically reinforced memory pathways offer a promising alternative to conventional memory mechanisms, addressing longstanding limitations in extended sequence modeling.

Federated learning (FL) is a general principle for decentralized clients to train a server model collectively without sharing local data. FL is a promising framework with practical applications, but its standard training paradigm requires the clients to backpropagate through the model to compute gradients. Since these clients are typically edge devices and not fully trusted, executing backpropagation on them incurs computational and storage overhead as well as white-box vulnerability. In light of this, we develop backpropagation-free federated learning, dubbed BAFFLE, in which backpropagation is replaced by multiple forward processes to estimate gradients. BAFFLE is 1) memory-efficient and easily fits uploading bandwidth; 2) compatible with inference-only hardware optimization and model quantization or pruning; and 3) well-suited to trusted execution environments, because the clients in BAFFLE only execute forward propagation and return a set of scalars to the server. Empirically we use BAFFLE to train deep models from scratch or to finetune pretrained models, achieving acceptable results. Code is available in https://github.com/FengHZ/BAFFLE.

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

We introduce pixelSplat, a feed-forward model that learns to reconstruct 3D radiance fields parameterized by 3D Gaussian primitives from pairs of images. Our model features real-time and memory-efficient rendering for scalable training as well as fast 3D reconstruction at inference time. To overcome local minima inherent to sparse and locally supported representations, we predict a dense probability distribution over 3D and sample Gaussian means from that probability distribution. We make this sampling operation differentiable via a reparameterization trick, allowing us to back-propagate gradients through the Gaussian splatting representation. We benchmark our method on wide-baseline novel view synthesis on the real-world RealEstate10k and ACID datasets, where we outperform state-of-the-art light field transformers and accelerate rendering by 2.5 orders of magnitude while reconstructing an interpretable and editable 3D radiance field.

We propose the In-context Autoencoder (ICAE) for context compression in a large language model (LLM). The ICAE has two modules: a learnable encoder adapted with LoRA from an LLM for compressing a long context into a limited number of memory slots, and a fixed decoder which is the target LLM that can condition on the memory slots for various purposes. We first pretrain the ICAE using both autoencoding and language modeling objectives on massive text data, enabling it to generate memory slots that accurately and comprehensively represent the original context. Then, we fine-tune the pretrained ICAE on a small amount of instruct data to enhance its interaction with various prompts for producing desirable responses. Our experimental results demonstrate that the ICAE learned with our proposed pretraining and fine-tuning paradigm can effectively produce memory slots with 4times context compression, which can be well conditioned on by the target LLM to respond to various prompts. The promising results demonstrate significant implications of the ICAE for its novel approach to the long context problem and its potential to reduce computation and memory overheads for LLM inference in practice, suggesting further research effort in context management for an LLM. Our code and data will be released shortly.

Transformer based models, like BERT and RoBERTa, have achieved state-of-the-art results in many Natural Language Processing tasks. However, their memory footprint, inference latency, and power consumption are prohibitive efficient inference at the edge, and even at the data center. While quantization can be a viable solution for this, previous work on quantizing Transformer based models use floating-point arithmetic during inference, which cannot efficiently utilize integer-only logical units such as the recent Turing Tensor Cores, or traditional integer-only ARM processors. In this work, we propose I-BERT, a novel quantization scheme for Transformer based models that quantizes the entire inference with integer-only arithmetic. Based on lightweight integer-only approximation methods for nonlinear operations, e.g., GELU, Softmax, and Layer Normalization, I-BERT performs an end-to-end integer-only BERT inference without any floating point calculation. We evaluate our approach on GLUE downstream tasks using RoBERTa-Base/Large. We show that for both cases, I-BERT achieves similar (and slightly higher) accuracy as compared to the full-precision baseline. Furthermore, our preliminary implementation of I-BERT shows a speedup of 2.4-4.0x for INT8 inference on a T4 GPU system as compared to FP32 inference. The framework has been developed in PyTorch and has been open-sourced.

Quantization is an effective method for reducing memory footprint and inference time of Neural Networks, e.g., for efficient inference in the cloud, especially at the edge. However, ultra low precision quantization could lead to significant degradation in model generalization. A promising method to address this is to perform mixed-precision quantization, where more sensitive layers are kept at higher precision. However, the search space for a mixed-precision quantization is exponential in the number of layers. Recent work has proposed HAWQ, a novel Hessian based framework, with the aim of reducing this exponential search space by using second-order information. While promising, this prior work has three major limitations: (i) HAWQV1 only uses the top Hessian eigenvalue as a measure of sensitivity and do not consider the rest of the Hessian spectrum; (ii) HAWQV1 approach only provides relative sensitivity of different layers and therefore requires a manual selection of the mixed-precision setting; and (iii) HAWQV1 does not consider mixed-precision activation quantization. Here, we present HAWQV2 which addresses these shortcomings. For (i), we perform a theoretical analysis showing that a better sensitivity metric is to compute the average of all of the Hessian eigenvalues. For (ii), we develop a Pareto frontier based method for selecting the exact bit precision of different layers without any manual selection. For (iii), we extend the Hessian analysis to mixed-precision activation quantization. We have found this to be very beneficial for object detection. We show that HAWQV2 achieves new state-of-the-art results for a wide range of tasks.

In modern large language models (LLMs), handling very long context lengths presents significant challenges as it causes slower inference speeds and increased memory costs. Additionally, most existing pre-trained LLMs fail to generalize beyond their original training sequence lengths. To enable efficient and practical long-context utilization, we introduce InfiniteHiP, a novel, and practical LLM inference framework that accelerates processing by dynamically eliminating irrelevant context tokens through a modular hierarchical token pruning algorithm. Our method also allows generalization to longer sequences by selectively applying various RoPE adjustment methods according to the internal attention patterns within LLMs. Furthermore, we offload the key-value cache to host memory during inference, significantly reducing GPU memory pressure. As a result, InfiniteHiP enables the processing of up to 3 million tokens on a single L40s 48GB GPU -- 3x larger -- without any permanent loss of context information. Our framework achieves an 18.95x speedup in attention decoding for a 1 million token context without requiring additional training. We implement our method in the SGLang framework and demonstrate its effectiveness and practicality through extensive evaluations.

Transformers have emerged as the architecture of choice for many state-of-the-art AI models, showcasing exceptional performance across a wide range of AI applications. However, the memory demands imposed by Transformers limit their ability to handle long sequences, thereby creating challenges for tasks involving extended sequences or long-term dependencies. We present a distinct approach, Ring Attention, which leverages blockwise computation of self-attention to distribute long sequences across multiple devices while concurrently overlapping the communication of key-value blocks with the computation of blockwise attention. By processing longer input sequences while maintaining memory efficiency, Ring Attention enables training and inference of sequences that are device count times longer than those of prior memory-efficient Transformers, effectively eliminating the memory constraints imposed by individual devices. Extensive experiments on language modeling tasks demonstrate the effectiveness of Ring Attention in allowing large sequence input size and improving performance.

We present MAGI-1, a world model that generates videos by autoregressively predicting a sequence of video chunks, defined as fixed-length segments of consecutive frames. Trained to denoise per-chunk noise that increases monotonically over time, MAGI-1 enables causal temporal modeling and naturally supports streaming generation. It achieves strong performance on image-to-video (I2V) tasks conditioned on text instructions, providing high temporal consistency and scalability, which are made possible by several algorithmic innovations and a dedicated infrastructure stack. MAGI-1 facilitates controllable generation via chunk-wise prompting and supports real-time, memory-efficient deployment by maintaining constant peak inference cost, regardless of video length. The largest variant of MAGI-1 comprises 24 billion parameters and supports context lengths of up to 4 million tokens, demonstrating the scalability and robustness of our approach. The code and models are available at https://github.com/SandAI-org/MAGI-1 and https://github.com/SandAI-org/MagiAttention. The product can be accessed at https://sand.ai.

Large Language Models (LLMs) demonstrate remarkable emergent abilities across various tasks, yet fall short of complex reasoning and planning tasks. The tree-search-based reasoning methods address this by surpassing the capabilities of chain-of-thought prompting, encouraging exploration of intermediate steps. However, such methods introduce significant inference latency due to the systematic exploration and evaluation of multiple thought paths. This paper introduces SeeD, a novel and efficient inference framework to optimize runtime speed and GPU memory management concurrently. By employing a scheduled speculative execution, SeeD efficiently handles multiple iterations for the thought generation and the state evaluation, leveraging a rounds-scheduled strategy to manage draft model dispatching. Extensive experimental evaluations on three reasoning datasets demonstrate superior speedup performance of SeeD, providing a viable path for batched inference in training-free speculative decoding.

We present Megrez2, a novel lightweight and high-performance language model architecture optimized for device native deployment. Megrez2 introduces a novel cross-layer expert sharing mechanism, which significantly reduces total parameter count by reusing expert modules across adjacent transformer layers while maintaining most of the model's capacity. It also incorporates pre-gated routing, enabling memory-efficient expert loading and faster inference. As the first instantiation of the Megrez2 architecture, we introduce the Megrez2-Preview model, which is pre-trained on a 5-trillion-token corpus and further enhanced through supervised fine-tuning and reinforcement learning with verifiable rewards. With only 3B activated and 7.5B stored parameters, Megrez2-Preview demonstrates competitive or superior performance compared to larger models on a wide range of tasks, including language understanding, instruction following, mathematical reasoning, and code generation. These results highlight the effectiveness of the Megrez2 architecture to achieve a balance between accuracy, efficiency, and deployability, making it a strong candidate for real-world, resource-constrained applications.

Transformers have revolutionized deep learning across various tasks, including audio representation learning, due to their powerful modeling capabilities. However, they often suffer from quadratic complexity in both GPU memory usage and computational inference time, affecting their efficiency. Recently, state space models (SSMs) like Mamba have emerged as a promising alternative, offering a more efficient approach by avoiding these complexities. Given these advantages, we explore the potential of SSM-based models in audio tasks. In this paper, we introduce Self-Supervised Audio Mamba (SSAMBA), the first self-supervised, attention-free, and SSM-based model for audio representation learning. SSAMBA leverages the bidirectional Mamba to capture complex audio patterns effectively. We incorporate a self-supervised pretraining framework that optimizes both discriminative and generative objectives, enabling the model to learn robust audio representations from large-scale, unlabeled datasets. We evaluated SSAMBA on various tasks such as audio classification, keyword spotting, and speaker identification. Our results demonstrate that SSAMBA outperforms the Self-Supervised Audio Spectrogram Transformer (SSAST) in most tasks. Notably, SSAMBA is approximately 92.7% faster in batch inference speed and 95.4% more memory-efficient than SSAST for the tiny model size with an input token size of 22k. These efficiency gains, combined with superior performance, underscore the effectiveness of SSAMBA's architectural innovation, making it a compelling choice for a wide range of audio processing applications.

Fine-tuning language models (LMs) has yielded success on diverse downstream tasks, but as LMs grow in size, backpropagation requires a prohibitively large amount of memory. Zeroth-order (ZO) methods can in principle estimate gradients using only two forward passes but are theorized to be catastrophically slow for optimizing large models. In this work, we propose a memory-efficient zerothorder optimizer (MeZO), adapting the classical ZO-SGD method to operate in-place, thereby fine-tuning LMs with the same memory footprint as inference. For example, with a single A100 80GB GPU, MeZO can train a 30-billion parameter model, whereas fine-tuning with backpropagation can train only a 2.7B LM with the same budget. We conduct comprehensive experiments across model types (masked and autoregressive LMs), model scales (up to 66B), and downstream tasks (classification, multiple-choice, and generation). Our results demonstrate that (1) MeZO significantly outperforms in-context learning and linear probing; (2) MeZO achieves comparable performance to fine-tuning with backpropagation across multiple tasks, with up to 12x memory reduction; (3) MeZO is compatible with both full-parameter and parameter-efficient tuning techniques such as LoRA and prefix tuning; (4) MeZO can effectively optimize non-differentiable objectives (e.g., maximizing accuracy or F1). We support our empirical findings with theoretical insights, highlighting how adequate pre-training and task prompts enable MeZO to fine-tune huge models, despite classical ZO analyses suggesting otherwise.

Attention is a key part of the transformer architecture. It is a sequence-to-sequence mapping that transforms each sequence element into a weighted sum of values. The weights are typically obtained as the softmax of dot products between keys and queries. Recent work has explored alternatives to softmax attention in transformers, such as ReLU and sigmoid activations. In this work, we revisit sigmoid attention and conduct an in-depth theoretical and empirical analysis. Theoretically, we prove that transformers with sigmoid attention are universal function approximators and benefit from improved regularity compared to softmax attention. Through detailed empirical analysis, we identify stabilization of large initial attention norms during the early stages of training as a crucial factor for the successful training of models with sigmoid attention, outperforming prior attempts. We also introduce FLASHSIGMOID, a hardware-aware and memory-efficient implementation of sigmoid attention yielding a 17% inference kernel speed-up over FLASHATTENTION2 on H100 GPUs. Experiments across language, vision, and speech show that properly normalized sigmoid attention matches the strong performance of softmax attention on a wide range of domains and scales, which previous attempts at sigmoid attention were unable to fully achieve. Our work unifies prior art and establishes best practices for sigmoid attention as a drop-in softmax replacement in transformers.

Rectified Flow (RF) models have advanced high-quality image and video synthesis via optimal transport theory. However, when applied to image-to-image translation, they still depend on costly multi-step denoising, hindering real-time applications. Although the recent adversarial training paradigm, CycleGAN-Turbo, works in pretrained diffusion models for one-step image translation, we find that directly applying it to RF models leads to severe convergence issues. In this paper, we analyze these challenges and propose TReFT, a novel method to Tame Rectified Flow models for one-step image Translation. Unlike previous works, TReFT directly uses the velocity predicted by pretrained DiT or UNet as output-a simple yet effective design that tackles the convergence issues under adversarial training with one-step inference. This design is mainly motivated by a novel observation that, near the end of the denoising process, the velocity predicted by pretrained RF models converges to the vector from origin to the final clean image, a property we further justify through theoretical analysis. When applying TReFT to large pretrained RF models such as SD3.5 and FLUX, we introduce memory-efficient latent cycle-consistency and identity losses during training, as well as lightweight architectural simplifications for faster inference. Pretrained RF models finetuned with TReFT achieve performance comparable to sota methods across multiple image translation datasets while enabling real-time inference.

The "one-shot" technique represents a distinct and sophisticated aesthetic in filmmaking. However, its practical realization is often hindered by prohibitive costs and complex real-world constraints. Although emerging video generation models offer a virtual alternative, existing approaches typically rely on naive clip concatenation, which frequently fails to maintain visual smoothness and temporal coherence. In this paper, we introduce DreaMontage, a comprehensive framework designed for arbitrary frame-guided generation, capable of synthesizing seamless, expressive, and long-duration one-shot videos from diverse user-provided inputs. To achieve this, we address the challenge through three primary dimensions. (i) We integrate a lightweight intermediate-conditioning mechanism into the DiT architecture. By employing an Adaptive Tuning strategy that effectively leverages base training data, we unlock robust arbitrary-frame control capabilities. (ii) To enhance visual fidelity and cinematic expressiveness, we curate a high-quality dataset and implement a Visual Expression SFT stage. In addressing critical issues such as subject motion rationality and transition smoothness, we apply a Tailored DPO scheme, which significantly improves the success rate and usability of the generated content. (iii) To facilitate the production of extended sequences, we design a Segment-wise Auto-Regressive (SAR) inference strategy that operates in a memory-efficient manner. Extensive experiments demonstrate that our approach achieves visually striking and seamlessly coherent one-shot effects while maintaining computational efficiency, empowering users to transform fragmented visual materials into vivid, cohesive one-shot cinematic experiences.

Transformers have demonstrated strong performance across a wide range of sequence modeling tasks, but their quadratic attention complexity limits scalability to long sequences. Linear models such as Mamba and sliding-window attention (SWA) address this by mixing tokens through recurrent or localized operations with fixed-size memory, achieving efficient inference. However, these methods risk degrading performance on long sequences due to their inability to retain detailed information from distant tokens. We propose SCOUT (Segment Compression for Optimized Utility in Transformers), a hybrid architecture that compresses tokens locally within fixed-size segments and applies attention only over these compressed representations. Each token embedding is first enriched via a linear local mixer, Mamba or SWA, that integrates recent context. Then, instead of attending to all previous tokens, each token sparsely attends to a small number of compressed checkpoint tokens that summarize the input history. This design retains much of the expressivity of full attention while substantially reducing the computational and memory cost. By attending to compressed history rather than all previous tokens, SCOUT incurs slightly higher memory than purely linear models, but its growth rate remains sub-quadratic and far more scalable than that of full Transformers. We analyze SCOUT's computational and memory efficiency and evaluate it empirically on long-context language modeling and reasoning tasks. SCOUT with both Mamba and SWA mixers outperforms strong long-sequence baselines under the same computational budget, matches full-attention Transformers on language modeling and common-sense reasoning tasks at 400M and 1.3B scales. Moreover, our SCOUT achieves higher end-to-end throughput than SOTA models, while delivering comparable results on long sequence benchmarks.

Processing long-context inputs with large language models presents a significant challenge due to the enormous memory requirements of the Key-Value (KV) cache during inference. Existing KV cache compression methods exhibit noticeable performance degradation when memory is reduced by more than 85%. Additionally, strategies that leverage GPU-CPU collaboration for approximate attention remain underexplored in this setting. We propose HCAttention, a heterogeneous attention computation framework that integrates key quantization, value offloading, and dynamic KV eviction to enable efficient inference under extreme memory constraints. The method is compatible with existing transformer architectures and does not require model fine-tuning. Experimental results on the LongBench benchmark demonstrate that our approach preserves the accuracy of full-attention model while shrinking the KV cache memory footprint to 25% of its original size. Remarkably, it stays competitive with only 12.5% of the cache, setting a new state-of-the-art in LLM KV cache compression. To the best of our knowledge, HCAttention is the first to extend the Llama-3-8B model to process 4 million tokens on a single A100 GPU with 80GB memory.

The time complexity of the standard attention mechanism in a transformer scales quadratically with the length of the sequence. We introduce a method to reduce this to linear scaling with time, based on defining attention via latent vectors. The method is readily usable as a drop-in replacement for the standard attention mechanism. Our "Latte Transformer" model can be implemented for both bidirectional and unidirectional tasks, with the causal version allowing a recurrent implementation which is memory and time-efficient during inference of language generation tasks. Whilst next token prediction scales linearly with the sequence length for a standard transformer, a Latte Transformer requires constant time to compute the next token. The empirical performance of our method is comparable to standard attention, yet allows scaling to context windows much larger than practical in standard attention.

The emergence of long-context text applications utilizing large language models (LLMs) has presented significant scalability challenges, particularly in memory footprint. The linear growth of the Key-Value (KV) cache responsible for storing attention keys and values to minimize redundant computations can lead to substantial increases in memory consumption, potentially causing models to fail to serve with limited memory resources. To address this issue, we propose a novel approach called Cache Sparse Representation (CSR), which converts the KV cache by transforming the dense Key-Value cache tensor into sparse indexes and weights, offering a more memory-efficient representation during LLM inference. Furthermore, we introduce NeuralDict, a novel neural network-based method for automatically generating the dictionary used in our sparse representation. Our extensive experiments demonstrate that CSR achieves performance comparable to state-of-the-art KV cache quantization algorithms while maintaining robust functionality in memory-constrained environments.

Long-video understanding has emerged as a crucial capability in real-world applications such as video surveillance, meeting summarization, educational lecture analysis, and sports broadcasting. However, it remains computationally prohibitive for VideoLLMs, primarily due to two bottlenecks: 1) sequential video decoding, the process of converting the raw bit stream to RGB frames can take up to a minute for hour-long video inputs, and 2) costly prefilling of up to several million tokens for LLM inference, resulting in high latency and memory use. To address these challenges, we propose QuickVideo, a system-algorithm co-design that substantially accelerates long-video understanding to support real-time downstream applications. It comprises three key innovations: QuickDecoder, a parallelized CPU-based video decoder that achieves 2-3 times speedup by splitting videos into keyframe-aligned intervals processed concurrently; QuickPrefill, a memory-efficient prefilling method using KV-cache pruning to support more frames with less GPU memory; and an overlapping scheme that overlaps CPU video decoding with GPU inference. Together, these components infernece time reduce by a minute on long video inputs, enabling scalable, high-quality video understanding even on limited hardware. Experiments show that QuickVideo generalizes across durations and sampling rates, making long video processing feasible in practice.

Gated Linear Units (GLUs) have become essential components in the feed-forward networks of state-of-the-art Large Language Models (LLMs). However, they require twice as many memory reads compared to feed-forward layers without gating, due to the use of separate weight matrices for the gate and value streams. To address this bottleneck, we introduce Masked Gated Linear Units (MGLUs), a novel family of GLUs with an efficient kernel implementation. The core contribution of MGLUs include: (1) the Mixture of Element-wise Gating (MoEG) architecture that learns multiple binary masks, each determining gate or value assignments at the element level on a single shared weight matrix resulting in reduced memory transfer, and (2) FlashMGLU, a hardware-friendly kernel that yields up to a 19.7 times inference-time speed-up over a naive PyTorch MGLU and is 47% more memory-efficient and 34% faster than standard GLUs despite added architectural complexity on an RTX5090 GPU. In LLM experiments, the Swish-activated variant SwiMGLU preserves its memory advantages while matching - or even surpassing - the downstream accuracy of the SwiGLU baseline.

The growing popularity of neuro symbolic reasoning has led to the adoption of various forms of differentiable (i.e., fuzzy) first order logic. We introduce PyReason, a software framework based on generalized annotated logic that both captures the current cohort of differentiable logics and temporal extensions to support inference over finite periods of time with capabilities for open world reasoning. Further, PyReason is implemented to directly support reasoning over graphical structures (e.g., knowledge graphs, social networks, biological networks, etc.), produces fully explainable traces of inference, and includes various practical features such as type checking and a memory-efficient implementation. This paper reviews various extensions of generalized annotated logic integrated into our implementation, our modern, efficient Python-based implementation that conducts exact yet scalable deductive inference, and a suite of experiments. PyReason is available at: github.com/lab-v2/pyreason.

While deep neural net inference was considered a task for servers only, latest advances in technology allow the task of inference to be moved to mobile and embedded devices, desired for various reasons ranging from latency to privacy. These devices are not only limited by their compute power and battery, but also by their inferior physical memory and cache, and thus, an efficient memory manager becomes a crucial component for deep neural net inference at the edge. We explore various strategies to smartly share memory buffers among intermediate tensors in deep neural nets. Employing these can result in up to 11% smaller memory footprint than the state of the art.

Large language models face significant computational and memory challenges when processing long contexts. During inference, efficient management of the key-value (KV) cache, which stores intermediate activations for autoregressive generation, is critical to reducing memory overhead and improving computational efficiency. Traditional token-level efficient KV caching methods overlook semantic information, treating tokens independently without considering their semantic relationships. Meanwhile, existing semantic-preserving KV cache management approaches often suffer from substantial memory usage and high time-to-first-token. To address these limitations, we propose SentenceKV, a novel sentence-level semantic KV caching approach designed to enhance inference efficiency while preserving semantic coherence. During prefilling, SentenceKV groups tokens based on sentence-level semantic similarity, compressing sentence representations into concise semantic vectors stored directly on the GPU, while individual KV pairs are offloaded to CPU. During decoding, SentenceKV generates tokens by selectively retrieving semantically relevant sentence-level KV entries, leveraging the semantic similarity between the prefilling-stage semantic vectors and decoding-stage queries. This ensures efficient and contextually accurate predictions, minimizing the loading of redundant or irrelevant data into GPU memory and significantly reducing memory overhead while maintaining stable inference latency, even for extremely long contexts. Extensive evaluations on benchmarks including PG-19, LongBench, and Needle-In-A-Haystack demonstrate that SentenceKV significantly outperforms state-of-the-art methods in both efficiency and memory usage, without compromising model accuracy.

Long-form video question answering (VQA) overwhelms current vision-language models (VLMs) because attention and key-value (KV) caches grow with runtime, forcing either expensive inference or near-sighted sliding windows. We introduce CacheFlow, a training-free pipeline that pairs Dynamic Token Dropping (DTD) with a compressive long-term memory. DTD prunes per-patch tokens online via cosine similarity to the previous frame, and surviving tokens are packed into fixed-size blocks. This online, per-frame processing makes our approach fundamentally suited for live streaming VQA. As blocks are processed, each one's keys are summarized by a tiny recurrent encoder to form a retrieval index, while the block's full KV pairs are offloaded and later rehydrated for generation, preserving answer fidelity. At inference, a consensus-based retrieval mechanism retrieves only the Top-K most relevant blocks and attends over both the retrieved and local context for precise, long-range reasoning. CacheFlow is drop-in, architecture-agnostic, and requires no fine-tuning. Experiments on both offline and streaming VQA benchmarks demonstrate that CacheFlow outperforms current strong baselines, while processing up to 87% less tokens. Our dual approach enables VLMs to be both efficient and context-aware, paving the way for practical long-form video understanding.

Generative large language models (LLMs) have opened up numerous novel possibilities, but due to their significant computational requirements their ubiquitous use remains challenging. Some of the most useful applications require processing large numbers of samples at a time and using long contexts, both significantly increasing the memory communication load of the models. We introduce SparQ Attention, a technique for increasing the inference throughput of LLMs by reducing the memory bandwidth requirements within the attention blocks through selective fetching of the cached history. Our proposed technique can be applied directly to off-the-shelf LLMs during inference, without requiring any modification to the pre-training setup or additional fine-tuning. We show how SparQ Attention can decrease the attention memory bandwidth requirements up to eight times without any loss in accuracy by evaluating Llama 2 and Pythia models on a wide range of downstream tasks.

Large Language Models(LLMs) have had a profound impact on AI applications, particularly in the domains of long-text comprehension and generation. KV Cache technology is one of the most widely used techniques in the industry. It ensures efficient sequence generation by caching previously computed KV states. However, it also introduces significant memory overhead. We discovered that KV Cache is not necessary and proposed a novel KCache technique to alleviate the memory bottleneck issue during the LLMs inference process. KCache can be used directly for inference without any training process, Our evaluations show that KCache improves the throughput of popular LLMs by 40% with the baseline, while keeping accuracy.

Large Language Models (LLMs) with hundreds of billions of parameters have transformed the field of machine learning. However, serving these models at inference time is both compute and memory intensive, where a single request can require multiple GPUs and tens of Gigabytes of memory. Multi-Head Attention is one of the key components of LLMs, which can account for over 50% of LLMs memory and compute requirement. We observe that there is a high amount of redundancy across heads on which tokens they pay attention to. Based on this insight, we propose Clustered Head Attention (CHAI). CHAI combines heads with a high amount of correlation for self-attention at runtime, thus reducing both memory and compute. In our experiments, we show that CHAI is able to reduce the memory requirements for storing K,V cache by up to 21.4% and inference time latency by up to 1.73x without any fine-tuning required. CHAI achieves this with a maximum 3.2% deviation in accuracy across 3 different models (i.e. OPT-66B, LLAMA-7B, LLAMA-33B) and 5 different evaluation datasets.

Transformer based Large Language Models (LLMs) have been widely used in many fields, and the efficiency of LLM inference becomes hot topic in real applications. However, LLMs are usually complicatedly designed in model structure with massive operations and perform inference in the auto-regressive mode, making it a challenging task to design a system with high efficiency. In this paper, we propose an efficient LLM inference solution with low latency and high throughput. Firstly, we simplify the LLM decoder layer by fusing data movement and element-wise operations to reduce the memory access frequency and lower system latency. We also propose a segment KV cache policy to keep key/value of the request and response tokens in separate physical memory for effective device memory management, helping enlarge the runtime batch size and improve system throughput. A customized Scaled-Dot-Product-Attention kernel is designed to match our fusion policy based on the segment KV cache solution. We implement our LLM inference solution on Intel GPU and publish it publicly. Compared with the standard HuggingFace implementation, the proposed solution achieves up to 7x lower token latency and 27x higher throughput for some popular LLMs on Intel GPU.

Transformer-based large language models (LLMs) demonstrate impressive performance across various natural language processing tasks. Serving LLM inference for generating long contents, however, poses a challenge due to the enormous memory footprint of the transient state, known as the key-value (KV) cache, which scales with the sequence length and batch size. In this paper, we present InfiniGen, a novel KV cache management framework tailored for long-text generation, which synergistically works with modern offloading-based inference systems. InfiniGen leverages the key insight that a few important tokens that are essential for computing the subsequent attention layer in the Transformer can be speculated by performing a minimal rehearsal with the inputs of the current layer and part of the query weight and key cache of the subsequent layer. This allows us to prefetch only the essential KV cache entries (without fetching them all), thereby mitigating the fetch overhead from the host memory in offloading-based LLM serving systems. Our evaluation on several representative LLMs shows that InfiniGen improves the overall performance of a modern offloading-based system by up to 3.00x compared to prior KV cache management methods while offering substantially better model accuracy.

Inference with Transformer-based Large Language Models (LLMs) on long sequences is both costly and slow due to the quadratic complexity of the self-attention mechanism. We introduce Star Attention, a two-phase block-sparse approximation that improves computational efficiency by sharding attention across multiple hosts while minimizing communication overhead. In the first phase, the context is processed using blockwise-local attention across hosts, in parallel. In the second phase, query and response tokens attend to all prior cached tokens through sequence-global attention. Star Attention integrates seamlessly with most Transformer-based LLMs trained with global attention, reducing memory requirements and inference time by up to 11x while preserving 95-100% of accuracy.

Large language models (LLMs) have demonstrated remarkable performance and tremendous potential across a wide range of tasks. However, deploying these models has been challenging due to the astronomical amount of model parameters, which requires a demand for large memory capacity and high memory bandwidth. In this paper, we propose an effective approach that can make the deployment of LLMs more efficiently. We support an automatic INT4 weight-only quantization flow and design a special LLM runtime with highly-optimized kernels to accelerate the LLM inference on CPUs. We demonstrate the general applicability of our approach on popular LLMs including Llama2, Llama, GPT-NeoX, and showcase the extreme inference efficiency on CPUs. The code is publicly available at: https://github.com/intel/intel-extension-for-transformers.

Deploying a large language model (LLM) inference service remains costly because centralized serving depends on specialized GPU clusters and high-bandwidth interconnects in datacenters. An appealing alternative is to leverage collaborative decentralized GPU pools. However, heterogeneity in GPU and limited interconnected network bandwidth, along with potentially dynamic availability, make efficient scheduling the central challenge in this scenario. In this paper, we present Parallax, a decentralized LLM serving system that turns a pool of heterogeneous GPUs into an efficient inference platform via a two-phase scheduler. Parallax decomposes planning into (i) model allocation, which places layers of each replica across diverse GPUs to jointly optimize latency and throughput under memory and link-bandwidth constraints, and (ii) request-time GPU pipeline selection, which stitches layers from different replicas into end-to-end execution chains that balance load and adapt to current conditions. We implement Parallax and evaluate it on open-source LLMs deployed over real volunteer nodes. Parallax consistently reduces latency and increases throughput relative to decentralized baselines, demonstrating that principled scheduling can make volunteer compute a practical, affordable substrate for LLM inference. Github Repo at: https://github.com/GradientHQ/parallax.

Transformers have emerged as the underpinning architecture for Large Language Models (LLMs). In generative language models, the inference process involves two primary phases: prompt processing and token generation. Token generation, which constitutes the majority of the computational workload, primarily entails vector-matrix multiplications and interactions with the Key-Value (KV) Cache. This phase is constrained by memory bandwidth due to the overhead of transferring weights and KV cache values from the memory system to the computing units. This memory bottleneck becomes particularly pronounced in applications that require long-context and extensive text generation, both of which are increasingly crucial for LLMs. This paper introduces "Keyformer", an innovative inference-time approach, to mitigate the challenges associated with KV cache size and memory bandwidth utilization. Keyformer leverages the observation that approximately 90% of the attention weight in generative inference focuses on a specific subset of tokens, referred to as "key" tokens. Keyformer retains only the key tokens in the KV cache by identifying these crucial tokens using a novel score function. This approach effectively reduces both the KV cache size and memory bandwidth usage without compromising model accuracy. We evaluate Keyformer's performance across three foundational models: GPT-J, Cerebras-GPT, and MPT, which employ various positional embedding algorithms. Our assessment encompasses a variety of tasks, with a particular emphasis on summarization and conversation tasks involving extended contexts. Keyformer's reduction of KV cache reduces inference latency by 2.1x and improves token generation throughput by 2.4x, while preserving the model's accuracy.

Large Language Models (LLMs) for Generative AI have achieved remarkable progress, evolving into sophisticated and versatile tools widely adopted across various domains and applications. However, the substantial memory overhead caused by their vast number of parameters, combined with the high computational demands of the attention mechanism, poses significant challenges in achieving low latency and high throughput for LLM inference services. Recent advancements, driven by groundbreaking research, have significantly accelerated progress in this field. This paper provides a comprehensive survey of these methods, covering fundamental instance-level approaches, in-depth cluster-level strategies, emerging scenario directions, and other miscellaneous but important areas. At the instance level, we review model placement, request scheduling, decoding length prediction, storage management, and the disaggregation paradigm. At the cluster level, we explore GPU cluster deployment, multi-instance load balancing, and cloud service solutions. For emerging scenarios, we organize the discussion around specific tasks, modules, and auxiliary methods. To ensure a holistic overview, we also highlight several niche yet critical areas. Finally, we outline potential research directions to further advance the field of LLM inference serving.

We introduce QUICK, a group of novel optimized CUDA kernels for the efficient inference of quantized Large Language Models (LLMs). QUICK addresses the shared memory bank-conflict problem of state-of-the-art mixed precision matrix multiplication kernels. Our method interleaves the quantized weight matrices of LLMs offline to skip the shared memory write-back after the dequantization. We demonstrate up to 1.91x speedup over existing kernels of AutoAWQ on larger batches and up to 1.94x throughput gain on representative LLM models on various NVIDIA GPU devices.

MoE (Mixture of Experts) prevails as a neural architecture that can scale modern transformer-based LLMs (Large Language Models) to unprecedented scales. Nevertheless, large MoEs' great demands of computing power, memory capacity and memory bandwidth make scalable serving a fundamental challenge and efficient parallel inference has become a requisite to attain adequate throughput under latency constraints. DeepSpeed-MoE, one state-of-the-art MoE inference framework, adopts a 3D-parallel paradigm including EP (Expert Parallelism), TP (Tensor Parallel) and DP (Data Parallelism). However, our analysis shows DeepSpeed-MoE's inference efficiency is largely bottlenecked by EP, which is implemented with costly all-to-all collectives to route token activation. Our work aims to boost DeepSpeed-MoE by strategically reducing EP's communication overhead with a technique named Speculative MoE. Speculative MoE has two speculative parallelization schemes, speculative token shuffling and speculative expert grouping, which predict outstanding tokens' expert routing paths and pre-schedule tokens and experts across devices to losslessly trim EP's communication volume. Besides DeepSpeed-MoE, we also build Speculative MoE into a prevailing MoE inference engine SGLang. Experiments show Speculative MoE can significantly boost state-of-the-art MoE inference frameworks on fast homogeneous and slow heterogeneous interconnects.

Despite their remarkable capabilities, Large Language Models (LLMs) struggle to effectively leverage historical interaction information in dynamic and complex environments. Memory systems enable LLMs to move beyond stateless interactions by introducing persistent information storage, retrieval, and utilization mechanisms. However, existing memory systems often introduce substantial time and computational overhead. To this end, we introduce a new memory system called LightMem, which strikes a balance between the performance and efficiency of memory systems. Inspired by the Atkinson-Shiffrin model of human memory, LightMem organizes memory into three complementary stages. First, cognition-inspired sensory memory rapidly filters irrelevant information through lightweight compression and groups information according to their topics. Next, topic-aware short-term memory consolidates these topic-based groups, organizing and summarizing content for more structured access. Finally, long-term memory with sleep-time update employs an offline procedure that decouples consolidation from online inference. Experiments on LongMemEval with GPT and Qwen backbones show that LightMem outperforms strong baselines in accuracy (up to 10.9% gains) while reducing token usage by up to 117x, API calls by up to 159x, and runtime by over 12x. The code is available at https://github.com/zjunlp/LightMem.

While Chain-of-Thought (CoT) prompting improves reasoning in large language models (LLMs), the excessive length of reasoning tokens increases latency and KV cache memory usage, and may even truncate final answers under context limits. We propose ThinkLess, an inference-efficient framework that terminates reasoning generation early and maintains output quality without modifying the model. Atttention analysis reveals that answer tokens focus minimally on earlier reasoning steps and primarily attend to the reasoning terminator token, due to information migration under causal masking. Building on this insight, ThinkLess inserts the terminator token at earlier positions to skip redundant reasoning while preserving the underlying knowledge transfer. To prevent format discruption casued by early termination, ThinkLess employs a lightweight post-regulation mechanism, relying on the model's natural instruction-following ability to produce well-structured answers. Without fine-tuning or auxiliary data, ThinkLess achieves comparable accuracy to full-length CoT decoding while greatly reducing decoding time and memory consumption.

With the advancements in long-context inference capabilities of large language models (LLMs), the KV cache has become one of the foundational components. However, its substantial GPU memory consumption makes KV cache compression a key technique for enabling efficient LLM inference in industrial scenarios. While recent studies have focused on optimizing the memory occupied by the KV cache, they overlook two critical factors: preserving semantic coherence and considering task-specific characteristic during compression. To address these limitations, we propose a novel task-adaptive KV cache window selection method, WindowKV. WindowKV dynamically selects local semantic windows consisting of consecutive tokens, according to task-specific characteristics, ensuring the retained KV cache captures continuous, essential context. Additionally, we introduce an intra-group layer KV cache indices sharing strategy to reduce computational overhead, achieving a balance between performance and efficiency. We rigorously evaluate WindowKV on the LongBench benchmark, and the results demonstrate that it maintains a performance comparable to full KV cache retention while using only 12% of the original KV cache, significantly reducing memory requirements. Furthermore, our method also achieves state-of-the-art results in the Needle-in-a-Haystack evaluation, highlighting its effectiveness and robustness.

With the development of large language models (LLMs), efficient inference through Key-Value (KV) cache compression has attracted considerable attention, especially for long-context generation. To compress the KV cache, recent methods identify critical KV tokens through heuristic ranking with attention scores. However, these methods often struggle to accurately determine critical tokens as they neglect the temporal patterns in attention scores, resulting in a noticeable degradation in LLM performance. To address this challenge, we propose AttentionPredictor, which is the first learning-based critical token identification approach. Specifically, AttentionPredictor learns a lightweight convolution model to capture spatiotemporal patterns and predict the next-token attention score. An appealing feature of AttentionPredictor is that it accurately predicts the attention score while consuming negligible memory. Moreover, we propose a cross-token critical cache prefetching framework that hides the token estimation time overhead to accelerate the decoding stage. By retaining most of the attention information, AttentionPredictor achieves 16times KV cache compression with comparable LLM performance, significantly outperforming the state-of-the-art.

Large Language Models (LLMs) have grown rapidly in size, creating significant challenges for efficient deployment on resource-constrained hardware. In this paper, we introduce Dynamic-Length Float (DFloat11), a lossless compression framework that reduces LLM size by 30% while preserving outputs that are bit-for-bit identical to the original model. DFloat11 is motivated by the low entropy in the BFloat16 weight representation of LLMs, which reveals significant inefficiency in existing storage format. By applying entropy coding, DFloat11 assigns dynamic-length encodings to weights based on frequency, achieving near information-optimal compression without any loss of precision. To facilitate efficient inference with dynamic-length encodings, we develop a custom GPU kernel for fast online decompression. Our design incorporates the following: (i) decomposition of memory-intensive lookup tables (LUTs) into compact LUTs that fit in GPU SRAM, (ii) a two-phase kernel for coordinating thread read/write positions using lightweight auxiliary variables, and (iii) transformer-block-level decompression to minimize latency. Experiments on recent models, including Llama-3.1, Qwen-2.5, and Gemma-3, validates our hypothesis that DFloat11 achieves around 30% model size reduction while preserving bit-for-bit exact outputs. Compared to a potential alternative of offloading parts of an uncompressed model to the CPU to meet memory constraints, DFloat11 achieves 1.9-38.8x higher throughput in token generation. With a fixed GPU memory budget, DFloat11 enables 5.3-13.17x longer context lengths than uncompressed models. Notably, our method enables lossless inference of Llama-3.1-405B, an 810GB model, on a single node equipped with 8x80GB GPUs. Our code and models are available at https://github.com/LeanModels/DFloat11.

Large Language Models (LLMs), despite their recent impressive accomplishments, are notably cost-prohibitive to deploy, particularly for applications involving long-content generation, such as dialogue systems and story writing. Often, a large amount of transient state information, referred to as the KV cache, is stored in GPU memory in addition to model parameters, scaling linearly with the sequence length and batch size. In this paper, we introduce a novel approach for implementing the KV cache which significantly reduces its memory footprint. Our approach is based on the noteworthy observation that a small portion of tokens contributes most of the value when computing attention scores. We call these tokens Heavy Hitters (H_2). Through a comprehensive investigation, we find that (i) the emergence of H_2 is natural and strongly correlates with the frequent co-occurrence of tokens in the text, and (ii) removing them results in significant performance degradation. Based on these insights, we propose Heavy Hitter Oracle (H_2O), a KV cache eviction policy that dynamically retains a balance of recent and H_2 tokens. We formulate the KV cache eviction as a dynamic submodular problem and prove (under mild assumptions) a theoretical guarantee for our novel eviction algorithm which could help guide future work. We validate the accuracy of our algorithm with OPT, LLaMA, and GPT-NeoX across a wide range of tasks. Our implementation of H_2O with 20% heavy hitters improves the throughput over three leading inference systems DeepSpeed Zero-Inference, Hugging Face Accelerate, and FlexGen by up to 29times, 29times, and 3times on OPT-6.7B and OPT-30B. With the same batch size, H2O can reduce the latency by up to 1.9times. The code is available at https://github.com/FMInference/H2O.

In-context learning (ICL) approaches typically leverage prompting to condition decoder-only language model generation on reference information. Just-in-time processing of a context is inefficient due to the quadratic cost of self-attention operations, and caching is desirable. However, caching transformer states can easily require almost as much space as the model parameters. When the right context isn't known in advance, caching ICL can be challenging. This work addresses these limitations by introducing models that, inspired by the encoder-decoder architecture, use cross-attention to condition generation on reference text without the prompt. More precisely, we leverage pre-trained decoder-only models and only train a small number of added layers. We use Question-Answering (QA) as a testbed to evaluate the ability of our models to perform conditional generation and observe that they outperform ICL, are comparable to fine-tuned prompted LLMs, and drastically reduce the space footprint relative to standard KV caching by two orders of magnitude.

Residual transformations enhance the representational depth and expressive power of large language models (LLMs). However, applying static residual transformations across all tokens in auto-regressive generation leads to a suboptimal trade-off between inference efficiency and generation fidelity. Existing methods, including Early Exiting, Skip Decoding, and Mixture-of-Depth address this by modulating the residual transformation based on token-level complexity. Nevertheless, these approaches predominantly consider the distance traversed by tokens through the model layers, neglecting the underlying velocity of residual evolution. We introduce Mixture of Multi-rate Residuals (M2R2), a framework that dynamically modulates residual velocity to improve early alignment, enhancing inference efficiency. Evaluations on reasoning oriented tasks such as Koala, Self-Instruct, WizardLM, and MT-Bench show M2R2 surpasses state-of-the-art distance-based strategies, balancing generation quality and speedup. In self-speculative decoding setup, M2R2 achieves up to 2.8x speedups on MT-Bench, outperforming methods like 2-model speculative decoding, Medusa, LookAhead Decoding, and DEED. In Mixture-of-Experts (MoE) architectures, integrating early residual alignment with ahead-of-time expert loading into high-bandwidth memory (HBM) accelerates decoding, reduces expert-switching bottlenecks, and achieves a 2.9x speedup, making it highly effective in resource-constrained environments.

Transformer-based large language models (LLMs) have achieved remarkable success as model sizes continue to grow, yet their deployment remains challenging due to significant computational and memory demands. Quantization has emerged as a promising solution, and state-of-the-art quantization algorithms for LLMs introduce the need for mixed-precision matrix multiplication (mpGEMM), where lower-precision weights are multiplied with higher-precision activations. Despite its benefits, current hardware accelerators such as GPUs and TPUs lack native support for efficient mpGEMM, leading to inefficient dequantization operations in the main sequential loop. To address this limitation, we introduce MixPE, a specialized mixed-precision processing element designed for efficient low-bit quantization in LLM inference. MixPE leverages two key innovations to minimize dequantization overhead and unlock the full potential of low-bit quantization. First, recognizing that scale and zero point are shared within each quantization group, we propose performing dequantization after per-group mpGEMM, significantly reducing dequantization overhead. Second, instead of relying on conventional multipliers, MixPE utilizes efficient shift\&add operations for multiplication, optimizing both computation and energy efficiency. Our experimental results demonstrate that MixPE surpasses the state-of-the-art quantization accelerators by 2.6times speedup and 1.4times energy reduction.

Many computational factors limit broader deployment of large language models. In this paper, we focus on a memory bottleneck imposed by the key-value (KV) cache, a computational shortcut that requires storing previous KV pairs during decoding. While existing KV cache methods approach this problem by pruning or evicting large swaths of relatively less important KV pairs to dramatically reduce the memory footprint of the cache, they can have limited success in tasks that require recollecting a majority of previous tokens. To alleviate this issue, we propose LESS, a simple integration of a (nearly free) constant sized cache with eviction-based cache methods, such that all tokens can be queried at later decoding steps. Its ability to retain information throughout time shows merit on a variety of tasks where we demonstrate LESS can help reduce the performance gap from caching everything, sometimes even matching it, all while being efficient.

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

The recent surge of open-source large language models (LLMs) enables developers to create AI-based solutions while maintaining control over aspects such as privacy and compliance, thereby providing governance and ownership of the model deployment process. To utilize these LLMs, inference engines are needed. These engines load the model's weights onto available resources, such as GPUs, and process queries to generate responses. The speed of inference, or performance, of the LLM, is critical for real-time applications, as it computes millions or billions of floating point operations per inference. Recently, advanced inference engines such as vLLM have emerged, incorporating novel mechanisms such as efficient memory management to achieve state-of-the-art performance. In this paper, we analyze the performance, particularly the throughput (tokens generated per unit of time), of 20 LLMs using two inference libraries: vLLM and HuggingFace's pipelines. We investigate how various hyperparameters, which developers must configure, influence inference performance. Our results reveal that throughput landscapes are irregular, with distinct peaks, highlighting the importance of hyperparameter optimization to achieve maximum performance. We also show that applying hyperparameter optimization when upgrading or downgrading the GPU model used for inference can improve throughput from HuggingFace pipelines by an average of 9.16% and 13.7%, respectively.

Large Language Model (LLM) inference requires substantial computational resources, yet CPU-based inference remains essential for democratizing AI due to the widespread availability of CPUs compared to specialized accelerators. However, efficient LLM inference on CPUs faces two fundamental challenges: (1) existing CPU architectures struggle with low-precision arithmetic required by quantized models, where optimal bit precision varies across models and layers; and (2) the memory-bound nature of the token generation phase creates severe performance bottlenecks. To address these challenges, we propose SAIL (SRAM-Accelerated Inference of LLMs), a CPU-based inference solution that efficiently supports arbitrary bit precisions with minimal overhead. SAIL integrates three key innovations: First, we introduce Batched LUT-based General Matrix-Vector Multiplication (LUT-GEMV) with SRAM-based processing-in-memory, enabling high data reuse through lookup tables and reducing memory movement. Second, our Pattern-Aware LUT optimization identifies and exploits redundancy in input activation patterns, reducing computation cycles by 13.8\%. Third, we develop an in-memory type conversion algorithm that leverages PIM's parallelism for efficient de-/quantization operations, alleviating pressure on CPU's vector units. Our architecture requires only 2\% hardware overhead and a single new instruction, while maintaining dual functionality as both compute and storage units. Experimental evaluations using a modified gem5 simulator demonstrate that SAIL achieves up to 10.7x speedup and 19.9x higher tokens per dollar compared to ARM Neoverse-N1 CPU baselines, and up to 7.04x better cost efficiency than NVIDIA V100 GPUs, establishing a practical path for efficient CPU-based LLM inference.

Transformer models struggle with long-context inference due to their quadratic time and linear memory complexity. Recurrent Memory Transformers (RMTs) offer a solution by reducing the asymptotic cost to linear time and constant memory usage. However, their memory update mechanism leads to sequential execution, causing a performance bottleneck. We introduce Diagonal Batching, a scheduling scheme that unlocks parallelism across segments in RMTs while preserving exact recurrence. This approach eliminates the sequential constraint, enabling efficient GPU inference even for single long-context inputs without complex batching and pipelining techniques. Because the technique is purely a run-time computation reordering, existing RMT models adopt it with no retraining. Applied to a LLaMA-1B ARMT model, Diagonal Batching yields a 3.3x speedup over standard full-attention LLaMA-1B and a 1.8x speedup over the sequential RMT implementation on 131,072-token sequences. By removing sequential bottleneck, Diagonal Batching reduces inference cost and latency, thereby strengthening RMTs as a practical solution for real-world, long-context applications.