Daily Papers

Matrix multiplication optimization remains a fundamental challenge in computational mathematics. This work introduces a novel approach that discovers matrix multiplication schemes in the ternary field (Z_T), where coefficients are restricted to {-1, 0, 1} to minimize naive additive complexity. The core of the method is a GPU-accelerated meta flip graph algorithm that maintains ternary safety through specialized arithmetic operations and sign symmetry breaking. Key results include new best ranks for the formats 4 times 5 times 12, 5 times 6 times 10, and 6 times 7 times 9, the independent discovery of 32 schemes in Z_T that match known optimal ranks (including 8 previously known only with rational coefficients), and 30 rank improvements in the binary field. The analysis of 164 known schemes shows that 92 can be implemented in Z_T, while 72 could not be found in the ternary field with current methods, defining the current boundaries of this approach. All software, results, and discovered schemes are provided as open-source.

The deployment of large language models (LLMs) is often constrained by memory bandwidth, where the primary bottleneck is the cost of transferring model parameters from the GPU's global memory to its registers. When coupled with custom kernels that fuse the dequantization and matmul operations, weight-only quantization can thus enable faster inference by reducing the amount of memory movement. However, developing high-performance kernels for weight-quantized LLMs presents substantial challenges, especially when the weights are compressed to non-evenly-divisible bit widths (e.g., 3 bits) with non-uniform, lookup table (LUT) quantization. This paper describes FLUTE, a flexible lookup table engine for LUT-quantized LLMs, which uses offline restructuring of the quantized weight matrix to minimize bit manipulations associated with unpacking, and vectorization and duplication of the lookup table to mitigate shared memory bandwidth constraints. At batch sizes < 32 and quantization group size of 128 (typical in LLM inference), the FLUTE kernel can be 2-4x faster than existing GEMM kernels. As an application of FLUTE, we explore a simple extension to lookup table-based NormalFloat quantization and apply it to quantize LLaMA3 to various configurations, obtaining competitive quantization performance against strong baselines while obtaining an end-to-end throughput increase of 1.5 to 2 times.

Large language models have been widely adopted but require significant GPU memory for inference. We develop a procedure for Int8 matrix multiplication for feed-forward and attention projection layers in transformers, which cut the memory needed for inference by half while retaining full precision performance. With our method, a 175B parameter 16/32-bit checkpoint can be loaded, converted to Int8, and used immediately without performance degradation. This is made possible by understanding and working around properties of highly systematic emergent features in transformer language models that dominate attention and transformer predictive performance. To cope with these features, we develop a two-part quantization procedure, LLM.int8(). We first use vector-wise quantization with separate normalization constants for each inner product in the matrix multiplication, to quantize most of the features. However, for the emergent outliers, we also include a new mixed-precision decomposition scheme, which isolates the outlier feature dimensions into a 16-bit matrix multiplication while still more than 99.9% of values are multiplied in 8-bit. Using LLM.int8(), we show empirically it is possible to perform inference in LLMs with up to 175B parameters without any performance degradation. This result makes such models much more accessible, for example making it possible to use OPT-175B/BLOOM on a single server with consumer GPUs. We open-source our software.

We give explicit low-rank bilinear non-commutative schemes for multiplying structured n times n matrices with 2 leq n leq 5, which serve as building blocks for recursive algorithms with improved multiplicative factors in asymptotic complexity. Our schemes are discovered over F_2 or F_3 and lifted to Z or Q. Using a flip graph search over tensor decompositions, we derive schemes for general, upper-triangular, lower-triangular, symmetric, and skew-symmetric inputs, as well as products of a structured matrix with its transpose. In particular, we obtain 4 times 4 rank-34 schemes: (i) multiplying a general matrix by its transpose using 10 recursive calls, improving the factor from 26/41 (0.634) to 8/13 (0.615); and (ii) multiplying an upper-triangular matrix by a general matrix using 12 recursive calls, improving the factor from 8/13 (0.615) to 22/37 (0.595). Additionally, using F_3 flip graphs, we discover schemes over Q that fundamentally require the inverse of 2, including a 2 times 2 symmetric-symmetric multiplication of rank 5 and a 3 times 3 skew-symmetric-general multiplication of rank 14 (improving upon AlphaTensor's 15).

In this study, we propose a simple method for fault-tolerant Strassen-like matrix multiplications. The proposed method is based on using two distinct Strassen-like algorithms instead of replicating a given one. We have realized that using two different algorithms, new check relations arise resulting in more local computations. These local computations are found using computer aided search. To improve performance, special parity (extra) sub-matrix multiplications (PSMMs) are generated (two of them) at the expense of increasing communication/computation cost of the system. Our preliminary results demonstrate that the proposed method outperforms a Strassen-like algorithm with two copies and secures a very close performance to three copy version using only 2 PSMMs, reducing the total number of compute nodes by around 24\% i.e., from 21 to 16.

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.

The new barrier mode in Apache Spark allows embedding distributed deep learning training as a Spark stage to simplify the distributed training workflow. In Spark, a task in a stage does not depend on any other tasks in the same stage, and hence it can be scheduled independently. However, several algorithms require more sophisticated inter-task communications, similar to the MPI paradigm. By combining distributed message passing (using asynchronous network IO), OpenJDK's new auto-vectorization and Spark's barrier execution mode, we can add non-map/reduce based algorithms, such as Cannon's distributed matrix multiplication to Spark. We document an efficient distributed matrix multiplication using Cannon's algorithm, which improves significantly on the performance of the existing MLlib implementation. Used within a barrier task, the algorithm described herein results in an up to 24 percent performance increase on a 10,000x10,000 square matrix with a significantly lower memory footprint. Applications of efficient matrix multiplication include, among others, accelerating the training and implementation of deep convolutional neural network based workloads, and thus such efficient algorithms can play a ground-breaking role in faster, more efficient execution of even the most complicated machine learning tasks.

In this paper, we propose CUDA-L2, a system that combines large language models (LLMs) and reinforcement learning (RL) to automatically optimize Half-precision General Matrix Multiply (HGEMM) CUDA kernels. Using CUDA execution speed as the RL reward, CUDA-L2 automatically optimizes HGEMM kernels across 1,000 configurations. CUDA-L2 systematically outperforms major matmul baselines to date, from the widely-used {\it torch.matmul} to state-of-the-art Nvidia's closed-source libraries, i.e., {\it cuBLAS}, {\it cuBLASLt}. In offline mode, where kernels are executed consecutively without time intervals, CUDA-L2 yields +22.0\% over {\it torch.matmul} on average; +19.2\% over {\it cuBLAS} using the optimal layout configuration (normal-normal NN and transposed-normal TN); +16.8\% over {\it cuBLASLt-heuristic}, which queries {\it cuBLASLt} library and selects the algorithm based on the heuristic's suggestion; and +11.4\% over the most competitive {\it cuBLASLt-AutoTuning} model, which selects the fastest algorithm from up to 100 candidates from {\it cuBLASLt}'s suggestions. In server mode, where kernels are executed at random intervals simulating real-time inference, the speedups further increase to +28.7\%, +26.0\%, +22.4\%, and +15.9\% for {\it torch.matmul}, {\it cuBLAS}, {\it cuBLASLt-heuristic}, and {\it cuBLASLt-AutoTuning} respectively. CUDA-L2 shows that even the most performance-critical, heavily-optimized kernels like HGEMM can be improved through LLM-guided RL automation by systematically exploring configuration spaces at scales impractical for humans. Project and code can be found at github.com/deepreinforce-ai/CUDA-L2

Inspired by the latest developments in multilevel Monte Carlo (MLMC) methods and randomised sketching for linear algebra problems we propose a MLMC estimator for real-time processing of matrix structured random data. Our algorithm is particularly effective in handling high-dimensional inner products and matrix multiplication, in applications of image analysis and large-scale supervised learning.

Knowledge graph (KG) learning offers a powerful framework for generating new knowledge and making inferences. Training KG embedding can take a significantly long time, especially for larger datasets. Our analysis shows that the gradient computation of embedding is one of the dominant functions in the translation-based KG embedding training loop. We address this issue by replacing the core embedding computation with SpMM (Sparse-Dense Matrix Multiplication) kernels. This allows us to unify multiple scatter (and gather) operations as a single operation, reducing training time and memory usage. We create a general framework for training KG models using sparse kernels and implement four models, namely TransE, TransR, TransH, and TorusE. Our sparse implementations exhibit up to 5.3x speedup on the CPU and up to 4.2x speedup on the GPU with a significantly low GPU memory footprint. The speedups are consistent across large and small datasets for a given model. Our proposed sparse approach can be extended to accelerate other translation-based (such as TransC, TransM, etc.) and non-translational (such as DistMult, ComplEx, RotatE, etc.) models as well. An implementation of the SpTransX framework is publicly available as a Python package in https://github.com/HipGraph/SpTransX.

Multiplications are responsible for most of the computational cost involved in neural network training and inference. Recent research has thus looked for ways to reduce the cost associated with them. Inspired by Mogami (2020), we replace multiplication with a cheap piecewise affine approximation that is achieved by adding the bit representation of the floating point numbers together as integers. We show that transformers can be trained with the resulting modified matrix multiplications on both vision and language tasks with little to no performance impact, and without changes to the training hyperparameters. We further replace all non-linearities in the networks making them fully and jointly piecewise affine in both inputs and weights. Finally, we show that we can eliminate all multiplications in the entire training process, including operations in the forward pass, backward pass and optimizer update, demonstrating the first successful training of modern neural network architectures in a fully multiplication-free fashion.

Recent advancements in fine-tuning Vision-Language Foundation Models (VLMs) have garnered significant attention for their effectiveness in downstream few-shot learning tasks.While these recent approaches exhibits some performance improvements, they often suffer from excessive training parameters and high computational costs. To address these challenges, we propose a novel Block matrix-based low-rank adaptation framework, called Block-LoRA, for fine-tuning VLMs on downstream few-shot tasks. Inspired by recent work on Low-Rank Adaptation (LoRA), Block-LoRA partitions the original low-rank decomposition matrix of LoRA into a series of sub-matrices while sharing all down-projection sub-matrices. This structure not only reduces the number of training parameters, but also transforms certain complex matrix multiplication operations into simpler matrix addition, significantly lowering the computational cost of fine-tuning. Notably, Block-LoRA enables fine-tuning CLIP on the ImageNet few-shot benchmark using a single 24GB GPU. We also show that Block-LoRA has the more tighter bound of generalization error than vanilla LoRA. Without bells and whistles, extensive experiments demonstrate that Block-LoRA achieves competitive performance compared to state-of-the-art CLIP-based few-shot methods, while maintaining a low training parameters count and reduced computational overhead.

Recent works have demonstrated reasonable success of representation learning in hypercomplex space. Specifically, "fully-connected layers with Quaternions" (4D hypercomplex numbers), which replace real-valued matrix multiplications in fully-connected layers with Hamilton products of Quaternions, both enjoy parameter savings with only 1/4 learnable parameters and achieve comparable performance in various applications. However, one key caveat is that hypercomplex space only exists at very few predefined dimensions (4D, 8D, and 16D). This restricts the flexibility of models that leverage hypercomplex multiplications. To this end, we propose parameterizing hypercomplex multiplications, allowing models to learn multiplication rules from data regardless of whether such rules are predefined. As a result, our method not only subsumes the Hamilton product, but also learns to operate on any arbitrary nD hypercomplex space, providing more architectural flexibility using arbitrarily 1/n learnable parameters compared with the fully-connected layer counterpart. Experiments of applications to the LSTM and Transformer models on natural language inference, machine translation, text style transfer, and subject verb agreement demonstrate architectural flexibility and effectiveness of the proposed approach.

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

General matrix-matrix multiplication (GEMM) is a cornerstone of AI computations, making tensor processing engines (TPEs) increasingly critical in GPUs and domain-specific architectures. Existing architectures primarily optimize dataflow or operand reuse strategies. However, considering the interaction between matrix multiplication and multiply-accumulators (MACs) offers greater optimization potential. This work introduces a novel hardware perspective on matrix multiplication, focusing on the bit-weight dimension of MACs. We propose a finer-grained TPE notation using matrix triple loops as an example, introducing new methods for designing and optimizing PE microarchitectures. Based on this notation and its transformations, we propose four optimization techniques that improve timing, area, and power consumption. Implementing our design in RTL using the SMIC-28nm process, we evaluate its effectiveness across four classic TPE architectures: systolic array, 3D-Cube, multiplier-adder tree, and 2D-Matrix. Our techniques achieve area efficiency improvements of 1.27x, 1.28x, 1.56x, and 1.44x, and energy efficiency gains of 1.04x, 1.56x, 1.49x, and 1.20x, respectively. Applied to a bit-slice architecture, our approach achieves a 12.10x improvement in energy efficiency and 2.85x in area efficiency compared to Laconic. Our Verilog HDL code, along with timing, area, and power reports, is available at https://github.com/wqzustc/High-Performance-Tensor-Processing-Engines

Although quantization for linear layers has been widely used, its application to accelerate the attention process remains limited. SageAttention utilizes 8-bit matrix multiplication, 16-bit matrix multiplication with 16-bit accumulator, and precision-enhancing methods, implementing an accurate and 2x speedup kernel compared to FlashAttention2. To further enhance the efficiency of attention computation while maintaining precision, we propose SageAttention2, which utilizes significantly faster 4-bit matrix multiplication (Matmul) alongside additional precision-enhancing techniques. First, we propose to quantize matrixes (Q, K) to INT4 in a warp-level granularity and quantize matrixes (widetilde P, V) to FP8. Second, we propose a method to smooth Q and V, enhancing the accuracy of attention with INT4 QK and FP8 PV. Third, we analyze the quantization accuracy across timesteps and layers, then propose an adaptive quantization method to ensure the end-to-end metrics over various models. The operations per second (OPS) of SageAttention2 surpass FlashAttention2 and xformers by about 3x and 5x on RTX4090, respectively. Comprehensive experiments confirm that our approach incurs negligible end-to-end metrics loss across diverse models, including those for large language processing, image generation, and video generation. The codes are available at https://github.com/thu-ml/SageAttention.

The advent of 1-bit large language models (LLMs), led by BitNet b1.58, has spurred interest in ternary LLMs. Despite this, research and practical applications focusing on efficient edge inference for ternary LLMs remain scarce. To bridge this gap, we introduce Bitnet.cpp, an inference system optimized for BitNet b1.58 and ternary LLMs. Given that mixed-precision matrix multiplication (mpGEMM) constitutes the bulk of inference time in ternary LLMs, Bitnet.cpp incorporates a novel mpGEMM library to facilitate sub-2-bits-per-weight, efficient and lossless inference. The library features two core solutions: Ternary Lookup Table (TL), which addresses spatial inefficiencies of previous bit-wise methods, and Int2 with a Scale (I2_S), which ensures lossless edge inference, both enabling high-speed inference. Our experiments show that Bitnet.cpp achieves up to a 6.25x increase in speed over full-precision baselines and up to 2.32x over low-bit baselines, setting new benchmarks in the field. Additionally, we expand TL to element-wise lookup table (ELUT) for low-bit LLMs in the appendix, presenting both theoretical and empirical evidence of its considerable potential. Bitnet.cpp is publicly available at https://github.com/microsoft/BitNet/tree/paper , offering a sophisticated solution for the efficient and practical deployment of edge LLMs.

In order to reduce the computational complexity of large language models, great efforts have been made to to improve the efficiency of transformer models such as linear attention and flash-attention. However, the model size and corresponding computational complexity are constantly scaled up in pursuit of higher performance. In this work, we present MemoryFormer, a novel transformer architecture which significantly reduces the computational complexity (FLOPs) from a new perspective. We eliminate nearly all the computations of the transformer model except for the necessary computation required by the multi-head attention operation. This is made possible by utilizing an alternative method for feature transformation to replace the linear projection of fully-connected layers. Specifically, we first construct a group of in-memory lookup tables that store a large amount of discrete vectors to replace the weight matrix used in linear projection. We then use a hash algorithm to retrieve a correlated subset of vectors dynamically based on the input embedding. The retrieved vectors combined together will form the output embedding, which provides an estimation of the result of matrix multiplication operation in a fully-connected layer. Compared to conducting matrix multiplication, retrieving data blocks from memory is a much cheaper operation which requires little computations. We train MemoryFormer from scratch and conduct extensive experiments on various benchmarks to demonstrate the effectiveness of the proposed model.

Quantization is a widely-used compression technology to reduce the overhead of serving large language models (LLMs) on terminal devices and in cloud data centers. However, prevalent quantization methods, such as 8-bit weight-activation or 4-bit weight-only quantization, achieve limited performance improvements due to poor support for low-precision (e.g., 4-bit) activation. This work, for the first time, realizes practical W4A4KV4 serving for LLMs, fully utilizing the INT4 tensor cores on modern GPUs and reducing the memory bottleneck caused by the KV cache. Specifically, we propose a novel fine-grained mixed-precision quantization algorithm (FMPQ) that compresses most activations into 4-bit with negligible accuracy loss. To support mixed-precision matrix multiplication for W4A4 and W4A8, we develop a highly optimized W4Ax kernel. Our approach introduces a novel mixed-precision data layout to facilitate access and fast dequantization for activation and weight tensors, utilizing the GPU's software pipeline to hide the overhead of data loading and conversion. Additionally, we propose fine-grained streaming multiprocessor (SM) scheduling to achieve load balance across different SMs. We integrate the optimized W4Ax kernel into our inference framework, COMET, and provide efficient management to support popular LLMs such as LLaMA-3-70B. Extensive evaluations demonstrate that, when running LLaMA family models on a single A100-80G-SMX4, COMET achieves a kernel-level speedup of 2.88times over cuBLAS and a 2.02 times throughput improvement compared to TensorRT-LLM from an end-to-end framework perspective.

Training large transformers is slow, but recent innovations on GPU architecture give us an advantage. NVIDIA Ampere GPUs can execute a fine-grained 2:4 sparse matrix multiplication twice as fast as its dense equivalent. In the light of this property, we comprehensively investigate the feasibility of accelerating feed-forward networks (FFNs) of transformers in pre-training. First, we define a ``flip rate'' to monitor the stability of a 2:4 training process. Utilizing this metric, we propose three techniques to preserve accuracy: to modify the sparse-refined straight-through estimator by applying the masked decay term on gradients, to determine a feasible decay factor in warm-up stage, and to enhance the model's quality by a dense fine-tuning procedure near the end of pre-training. Besides, we devise two techniques to practically accelerate training: to calculate transposable 2:4 masks by convolution, and to accelerate gated activation functions by reducing GPU L2 cache miss. Experiments show that our 2:4 sparse training algorithm achieves similar convergence to dense training algorithms on several transformer pre-training tasks, while actual acceleration can be observed on different shapes of transformer block apparently. Our toolkit is available at https://github.com/huyz2023/2by4-pretrain.

Large language models (LLMs) have demonstrated state-of-the-art performance across various tasks. However, the latency of inference and the large GPU memory consumption of LLMs restrict their deployment performance. Recently, there have been some efficient attempts to quantize LLMs, yet inference with large batch size or long sequence still has the issue of being compute-bound. Fine-grained quantization methods have showcased their proficiency in achieving low-bit quantization for LLMs, while requiring FP16 data type for linear layer computations, which is time-consuming when dealing with large batch size or long sequence. In this paper, we introduce a method called FlattenQuant, which significantly reduces the maximum value of the tensor by flattening the large channels in the tensor, to achieve low bit per-tensor quantization with minimal accuracy loss. Our experiments show that FlattenQuant can directly use 4 bits to achieve 48.29% of the linear layer calculation in LLMs, with the remaining layers using 8 bits. The 4-bit matrix multiplication introduced in the FlattenQuant method can effectively address the compute-bound caused by large matrix calculation. Our work achieves up to 2times speedup and 2.3times memory reduction for LLMs with negligible loss in accuracy.

Fast approximations to matrix multiplication have the potential to dramatically reduce the cost of neural network inference. Recent work on approximate matrix multiplication proposed to replace costly multiplications with table-lookups by fitting a fast hash function from training data. In this work, we propose improvements to this previous work, targeted to the deep learning inference setting, where one has access to both training data and fixed (already learned) model weight matrices. We further propose a fine-tuning procedure for accelerating entire neural networks while minimizing loss in accuracy. Finally, we analyze the proposed method on a simple image classification task. While we show improvements to prior work, overall classification accuracy remains substantially diminished compared to exact matrix multiplication. Our work, despite this negative result, points the way towards future efforts to accelerate inner products with fast nonlinear hashing methods.

Deep learning on point clouds has received increased attention thanks to its wide applications in AR/VR and autonomous driving. These applications require low latency and high accuracy to provide real-time user experience and ensure user safety. Unlike conventional dense workloads, the sparse and irregular nature of point clouds poses severe challenges to running sparse CNNs efficiently on the general-purpose hardware. Furthermore, existing sparse acceleration techniques for 2D images do not translate to 3D point clouds. In this paper, we introduce TorchSparse, a high-performance point cloud inference engine that accelerates the sparse convolution computation on GPUs. TorchSparse directly optimizes the two bottlenecks of sparse convolution: irregular computation and data movement. It applies adaptive matrix multiplication grouping to trade computation for better regularity, achieving 1.4-1.5x speedup for matrix multiplication. It also optimizes the data movement by adopting vectorized, quantized and fused locality-aware memory access, reducing the memory movement cost by 2.7x. Evaluated on seven representative models across three benchmark datasets, TorchSparse achieves 1.6x and 1.5x measured end-to-end speedup over the state-of-the-art MinkowskiEngine and SpConv, respectively.

Low-rank optimization has emerged as a promising direction in training large language models (LLMs) to reduce the memory usage of adaptive optimizers by constraining learning to a lower-dimensional space. Prior work typically projects gradients of linear layers using approaches based on Singular Value Decomposition (SVD). However, applying SVD-based procedures individually to each layer in large models is computationally expensive and incurs additional memory costs due to storing the projection matrices. In this work, we propose a computationally efficient and conceptually simple two-step procedure to approximate SVD-based gradient projections into lower-dimensional spaces. First, we construct a complete orthogonal basis using predefined orthogonal matrices of the Discrete Cosine Transform (DCT). Second, we adaptively select basis columns based on their alignment with the gradient of each layer. Each projection matrix in our method is obtained via a single matrix multiplication followed by a lightweight sorting step to identify the most relevant basis vectors. Due to the predefined nature of the orthogonal bases, they are computed once at the start of training. During training, we store only the indices of the selected columns, avoiding the need to store full projection matrices for each layer. Our numerical experiments on both pre-training and fine-tuning tasks demonstrate the effectiveness of our dual strategy in approximating optimal low-rank projections, matching the performance of costly SVD-based methods while achieving faster runtime and reduced memory usage.

The deployment of Large Language Models (LLMs) on edge devices is increasingly important to enhance on-device intelligence. Weight quantization is crucial for reducing the memory footprint of LLMs on devices. However, low-bit LLMs necessitate mixed precision matrix multiplication (mpGEMM) of low precision weights and high precision activations during inference. Existing systems, lacking native support for mpGEMM, resort to dequantize weights for high precision computation. Such an indirect way can lead to a significant inference overhead. In this paper, we introduce T-MAC, an innovative lookup table(LUT)-based method designed for efficient low-bit LLM (i.e., weight-quantized LLM) inference on CPUs. T-MAC directly supports mpGEMM without dequantization, while simultaneously eliminating multiplications and reducing additions required. Specifically, T-MAC transforms the traditional data-type-centric multiplication to bit-wise table lookup, and enables a unified and scalable mpGEMM solution. Our LUT-based kernels scale linearly to the weight bit-width. Evaluated on low-bit Llama and BitNet models, T-MAC demonstrates up to 4x increase in throughput and 70% reduction in energy consumption compared to llama.cpp. For BitNet-b1.58-3B, T-MAC delivers a token generation throughput of 30 tokens/s with a single core and 71 tokens/s with eight cores on M2-Ultra, and 11 tokens/s on lower-end devices like Raspberry Pi 5, which significantly exceeds the adult average reading speed. T-MAC with LUT-based computing paradigm, paves the way for the practical deployment of low-bit LLMs on resource-constrained edge devices without compromising computational efficiency. The system is open-sourced at https://github.com/microsoft/T-MAC.

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

Analytical framework for predicting General Matrix Multiplication (GEMM) performance on modern GPUs, focusing on runtime, power consumption, and energy efficiency. Our study employs two approaches: a custom-implemented tiled matrix multiplication kernel for fundamental analysis, and NVIDIA's CUTLASS library for comprehensive performance data collection across advanced configurations. Using the NVIDIA RTX 4070 as our experimental platform, we developed a Random Forest-based prediction model with multi-output regression capability. Through analysis of both naive tiled matrix multiplication with varying tile sizes (1 to 32) and 16,128 CUTLASS GEMM operations across diverse configurations, we identified critical performance patterns related to matrix dimensions, thread block configurations, and memory access patterns. Our framework achieved exceptional accuracy with an R^2 score of 0.98 for runtime prediction (mean error 15.57%) and 0.78 for power prediction (median error 5.42%). The system successfully predicts performance across matrix sizes, demonstrating robust scaling behavior. Our results show that optimal tile size selection can improve performance by up to 3.2x while reducing power consumption by 22% compared to baseline configurations. Analysis of shared memory utilization and SM occupancy reveals that tile sizes of 16x16 achieve the best balance between parallelism and resource usage. The implementation of our framework, including prediction models and analysis tools, is available as an open-source project at GPPerf [https://github.com/pavlyhalim/GPPerf].

Large language models (LLMs) have been widely applied but face challenges in efficient inference. While quantization methods reduce computational demands, ultra-low bit quantization with arbitrary precision is hindered by limited GPU Tensor Core support and inefficient memory management, leading to suboptimal acceleration. To address these challenges, we propose a comprehensive acceleration scheme for arbitrary precision LLMs. At its core, we introduce a novel bipolar-INT data format that facilitates parallel computing and supports symmetric quantization, effectively reducing data redundancy. Building on this, we implement an arbitrary precision matrix multiplication scheme that decomposes and recovers matrices at the bit level, enabling flexible precision while maximizing GPU Tensor Core utilization. Furthermore, we develop an efficient matrix preprocessing method that optimizes data layout for subsequent computations. Finally, we design a data recovery-oriented memory management system that strategically utilizes fast shared memory, significantly enhancing kernel execution speed and minimizing memory access latency. Experimental results demonstrate our approach's effectiveness, with up to 2.4\times speedup in matrix multiplication compared to NVIDIA's CUTLASS. When integrated into LLMs, we achieve up to 6.7\times inference acceleration. These improvements significantly enhance LLM inference efficiency, enabling broader and more responsive applications of LLMs.

Self-attention is at the heart of the popular Transformer architecture, yet suffers from quadratic time and memory complexity. The breakthrough FlashAttention algorithm revealed I/O complexity as the true bottleneck in scaling Transformers. Given two levels of memory hierarchy, a fast cache (e.g. GPU on-chip SRAM) and a slow memory (e.g. GPU high-bandwidth memory), the I/O complexity measures the number of accesses to memory. FlashAttention computes attention using N^2d^2{M} I/O operations where N is the dimension of the attention matrix, d the head-dimension and M the cache size. However, is this I/O complexity optimal? The known lower bound only rules out an I/O complexity of o(Nd) when M=Theta(Nd), since the output that needs to be written to slow memory is Omega(Nd). This leads to the main question of our work: Is FlashAttention I/O optimal for all values of M? We resolve the above question in its full generality by showing an I/O complexity lower bound that matches the upper bound provided by FlashAttention for any values of M geq d^2 within any constant factors. Further, we give a better algorithm with lower I/O complexity for M < d^2, and show that it is optimal as well. Moreover, our lower bounds do not rely on using combinatorial matrix multiplication for computing the attention matrix. We show even if one uses fast matrix multiplication, the above I/O complexity bounds cannot be improved. We do so by introducing a new communication complexity protocol for matrix compression, and connecting communication complexity to I/O complexity. To the best of our knowledge, this is the first work to establish a connection between communication complexity and I/O complexity, and we believe this connection could be of independent interest and will find many more applications in proving I/O complexity lower bounds in the future.

Quantized training of Large Language Models (LLMs) remains an open challenge, as maintaining accuracy while performing all matrix multiplications in low precision has proven difficult. This is particularly the case when fine-tuning pre-trained models, which often already have large weight and activation outlier values that render quantized optimization difficult. We present HALO, a novel quantization-aware training approach for Transformers that enables accurate and efficient low-precision training by combining 1) strategic placement of Hadamard rotations in both forward and backward passes, to mitigate outliers during the low-precision computation, 2) FSDP integration for low-precision communication, and 3) high-performance kernel support. Our approach ensures that all large matrix multiplications during the forward and backward passes are executed in lower precision. Applied to LLAMA-family models, HALO achieves near-full-precision-equivalent results during fine-tuning on various tasks, while delivering up to 1.31x end-to-end speedup for full fine-tuning on RTX 4090 GPUs. Our method supports both standard and parameter-efficient fine-tuning (PEFT) methods, both backed by efficient kernel implementations. Our results demonstrate the first practical approach to fully quantized LLM fine-tuning that maintains accuracy in FP8 precision, while delivering performance benefits.

Quantization is a crucial technique for deploying deep learning models on resource-constrained devices, such as embedded FPGAs. Prior efforts mostly focus on quantizing matrix multiplications, leaving other layers like BatchNorm or shortcuts in floating-point form, even though fixed-point arithmetic is more efficient on FPGAs. A common practice is to fine-tune a pre-trained model to fixed-point for FPGA deployment, but potentially degrading accuracy. This work presents QFX, a novel trainable fixed-point quantization approach that automatically learns the binary-point position during model training. Additionally, we introduce a multiplier-free quantization strategy within QFX to minimize DSP usage. QFX is implemented as a PyTorch-based library that efficiently emulates fixed-point arithmetic, supported by FPGA HLS, in a differentiable manner during backpropagation. With minimal effort, models trained with QFX can readily be deployed through HLS, producing the same numerical results as their software counterparts. Our evaluation shows that compared to post-training quantization, QFX can quantize models trained with element-wise layers quantized to fewer bits and achieve higher accuracy on both CIFAR-10 and ImageNet datasets. We further demonstrate the efficacy of multiplier-free quantization using a state-of-the-art binarized neural network accelerator designed for an embedded FPGA (AMD Xilinx Ultra96 v2). We plan to release QFX in open-source format.

The increasing scale of vision transformers (ViT) has made the efficient fine-tuning of these large models for specific needs a significant challenge in various applications. This issue originates from the computationally demanding matrix multiplications required during the backpropagation process through linear layers in ViT. In this paper, we tackle this problem by proposing a new Low-rank BackPropagation via Walsh-Hadamard Transformation (LBP-WHT) method. Intuitively, LBP-WHT projects the gradient into a low-rank space and carries out backpropagation. This approach substantially reduces the computation needed for adapting ViT, as matrix multiplication in the low-rank space is far less resource-intensive. We conduct extensive experiments with different models (ViT, hybrid convolution-ViT model) on multiple datasets to demonstrate the effectiveness of our method. For instance, when adapting an EfficientFormer-L1 model on CIFAR100, our LBP-WHT achieves 10.4% higher accuracy than the state-of-the-art baseline, while requiring 9 MFLOPs less computation. As the first work to accelerate ViT adaptation with low-rank backpropagation, our LBP-WHT method is complementary to many prior efforts and can be combined with them for better performance.

Low-precision formats such as float8 have been introduced in machine learning accelerated hardware to improve computational efficiency for large language models training and inference. Nevertheless, adoption by the ML community has been slowed down by the complex, and sometimes brittle, techniques required to match higher precision training accuracy. In this work, we present Scalify, a end-to-end scale propagation paradigm for computational graphs, generalizing and formalizing existing tensor scaling methods. Experiment results show that Scalify supports out-of-the-box float8 matrix multiplication and gradients representation, as well as float16 optimizer state storage. Our JAX implementation of Scalify is open-sourced at https://github.com/graphcore-research/jax-scalify

Large Language Models have shown remarkable capabilities in the NLP domain. Their effectiveness can mainly be attributed to their ability to adapt to an array of downstream tasks. However, generally, full fine-tuning is a computationally expensive job. To mitigate this, many techniques have been developed that prime efficiency, a prominent one being Low-Rank Adaptation (LoRA). However, LoRA and its variants employ re-parametrized additive updates. In this paper, we propose Low-Rank Multiplicative Adaptation (LoRMA), which shifts the paradigm of additive updates to a richer space of matrix multiplicative transformations. We tackle challenges such as computational complexity and rank bottleneck of matrix multiplication by effectively re-ordering operations and introducing rank inflation strategies. We conduct extensive experiments to demonstrate the effectiveness of our approach in terms of various evaluation metrics.

Attention mechanism has emerged as a foundation module of modern deep learning models and has also empowered many milestones in various domains. Moreover, FlashAttention with IO-aware speedup resolves the efficiency issue of standard attention, further promoting its practicality. Beyond canonical attention, attention with bias also widely exists, such as relative position bias in vision and language models and pair representation bias in AlphaFold. In these works, prior knowledge is introduced as an additive bias term of attention weights to guide the learning process, which has been proven essential for model performance. Surprisingly, despite the common usage of attention with bias, its targeted efficiency optimization is still absent, which seriously hinders its wide applications in complex tasks. Diving into the computation of FlashAttention, we prove that its optimal efficiency is determined by the rank of the attention weight matrix. Inspired by this theoretical result, this paper presents FlashBias based on the low-rank compressed sensing theory, which can provide fast-exact computation for many widely used attention biases and a fast-accurate approximation for biases in general formalization. FlashBias can fully take advantage of the extremely optimized matrix multiplication operation in modern GPUs, achieving 1.5times speedup for AlphaFold, and over 2times speedup for attention with bias in vision and language models without loss of accuracy.

Large Language Models (LLMs) have revolutionized natural language processing tasks. However, their practical application is constrained by substantial memory and computational demands. Post-training quantization (PTQ) is considered an effective method to accelerate LLM inference. Despite its growing popularity in LLM model compression, PTQ deployment faces two major challenges. First, low-bit quantization leads to performance degradation. Second, restricted by the limited integer computing unit type on GPUs, quantized matrix operations with different precisions cannot be effectively accelerated. To address these issues, we introduce a novel arbitrary-bit quantization algorithm and inference framework, ABQ-LLM. It achieves superior performance across various quantization settings and enables efficient arbitrary-precision quantized inference on the GPU. ABQ-LLM introduces several key innovations: (1) a distribution correction method for transformer blocks to mitigate distribution differences caused by full quantization of weights and activations, improving performance at low bit-widths. (2) the bit balance strategy to counteract performance degradation from asymmetric distribution issues at very low bit-widths (e.g., 2-bit). (3) an innovative quantization acceleration framework that reconstructs the quantization matrix multiplication of arbitrary precision combinations based on BTC (Binary TensorCore) equivalents, gets rid of the limitations of INT4/INT8 computing units. ABQ-LLM can convert each component bit width gain into actual acceleration gain, maximizing performance under mixed precision(e.g., W6A6, W2A8). Based on W2*A8 quantization configuration on LLaMA-7B model, it achieved a WikiText2 perplexity of 7.59 (2.17downarrow vs 9.76 in AffineQuant). Compared to SmoothQuant, we realized 1.6times acceleration improvement and 2.7times memory compression gain.

Tensor computations, with matrix multiplication being the primary operation, serve as the fundamental basis for data analysis, physics, machine learning, and deep learning. As the scale and complexity of data continue to grow rapidly, the demand for tensor computations has also increased significantly. To meet this demand, several research institutions have started developing dedicated hardware for tensor computations. To further improve the computational performance of tensor process units, we have reexamined the issue of computation reuse that was previously overlooked in existing architectures. As a result, we propose a novel EN-T architecture that can reduce chip area and power consumption. Furthermore, our method is compatible with existing tensor processing units. We evaluated our method on prevalent microarchitectures, the results demonstrate an average improvement in area efficiency of 8.7\%, 12.2\%, and 11.0\% for tensor computing units at computational scales of 256 GOPS, 1 TOPS, and 4 TOPS, respectively. Similarly, there were energy efficiency enhancements of 13.0\%, 17.5\%, and 15.5\%.

Temporal Action Detection (TAD) is challenging but fundamental for real-world video applications. Recently, DETR-based models have been devised for TAD but have not performed well yet. In this paper, we point out the problem in the self-attention of DETR for TAD; the attention modules focus on a few key elements, called temporal collapse problem. It degrades the capability of the encoder and decoder since their self-attention modules play no role. To solve the problem, we propose a novel framework, Self-DETR, which utilizes cross-attention maps of the decoder to reactivate self-attention modules. We recover the relationship between encoder features by simple matrix multiplication of the cross-attention map and its transpose. Likewise, we also get the information within decoder queries. By guiding collapsed self-attention maps with the guidance map calculated, we settle down the temporal collapse of self-attention modules in the encoder and decoder. Our extensive experiments demonstrate that Self-DETR resolves the temporal collapse problem by keeping high diversity of attention over all layers.

Convolution is the most time-consuming operation in deep neural network operations, so its performance is critical to the overall performance of the neural network. The commonly used methods for convolution on GPU include the general matrix multiplication (GEMM)-based convolution and the direct convolution. GEMM-based convolution relies on the im2col algorithm, which results in a large memory footprint and reduced performance. Direct convolution does not have the large memory footprint problem, but the performance is not on par with GEMM-based approach because of the discontinuous memory access. This paper proposes a window-order-based convolution paradigm on GPU, called im2win, which not only reduces memory footprint but also offers continuous memory accesses, resulting in improved performance. Furthermore, we apply a range of optimization techniques on the convolution CUDA kernel, including shared memory, tiling, micro-kernel, double buffer, and prefetching. We compare our implementation with the direct convolution, and PyTorch's GEMM-based convolution with cuBLAS and six cuDNN-based convolution implementations, with twelve state-of-the-art DNN benchmarks. The experimental results show that our implementation 1) uses less memory footprint by 23.1% and achieves 3.5times TFLOPS compared with cuBLAS, 2) uses less memory footprint by 32.8% and achieves up to 1.8times TFLOPS compared with the best performant convolutions in cuDNN, and 3) achieves up to 155times TFLOPS compared with the direct convolution. We further perform an ablation study on the applied optimization techniques and find that the micro-kernel has the greatest positive impact on performance.

Tensor Cores have been an important unit to accelerate Fused Matrix Multiplication Accumulation (MMA) in all NVIDIA GPUs since Volta Architecture. To program Tensor Cores, users have to use either legacy wmma APIs or current mma APIs. Legacy wmma APIs are more easy-to-use but can only exploit limited features and power of Tensor Cores. Specifically, wmma APIs support fewer operand shapes and can not leverage the new sparse matrix multiplication feature of the newest Ampere Tensor Cores. However, the performance of current programming interface has not been well explored. Furthermore, the computation numeric behaviors of low-precision floating points (TF32, BF16, and FP16) supported by the newest Ampere Tensor Cores are also mysterious. In this paper, we explore the throughput and latency of current programming APIs. We also intuitively study the numeric behaviors of Tensor Cores MMA and profile the intermediate operations including multiplication, addition of inner product, and accumulation. All codes used in this work can be found in https://github.com/sunlex0717/DissectingTensorCores.

Transformers based on the attention mechanism have achieved impressive success in various areas. However, the attention mechanism has a quadratic complexity, significantly impeding Transformers from dealing with numerous tokens and scaling up to bigger models. Previous methods mainly utilize the similarity decomposition and the associativity of matrix multiplication to devise linear-time attention mechanisms. They avoid degeneration of attention to a trivial distribution by reintroducing inductive biases such as the locality, thereby at the expense of model generality and expressiveness. In this paper, we linearize Transformers free from specific inductive biases based on the flow network theory. We cast attention as the information flow aggregated from the sources (values) to the sinks (results) through the learned flow capacities (attentions). Within this framework, we apply the property of flow conservation into attention and propose the Flow-Attention mechanism of linear complexity. By respectively conserving the incoming flow of sinks for source competition and the outgoing flow of sources for sink allocation, Flow-Attention inherently generates informative attentions without using specific inductive biases. Empowered by the Flow-Attention, Flowformer yields strong performance in linear time for wide areas, including long sequence, time series, vision, natural language, and reinforcement learning. The code and settings are available at this repository: https://github.com/thuml/Flowformer.

As language models grow ever larger, so do their vocabularies. This has shifted the memory footprint of LLMs during training disproportionately to one single layer: the cross-entropy in the loss computation. Cross-entropy builds up a logit matrix with entries for each pair of input tokens and vocabulary items and, for small models, consumes an order of magnitude more memory than the rest of the LLM combined. We propose Cut Cross-Entropy (CCE), a method that computes the cross-entropy loss without materializing the logits for all tokens into global memory. Rather, CCE only computes the logit for the correct token and evaluates the log-sum-exp over all logits on the fly. We implement a custom kernel that performs the matrix multiplications and the log-sum-exp reduction over the vocabulary in flash memory, making global memory consumption for the cross-entropy computation negligible. This has a dramatic effect. Taking the Gemma 2 (2B) model as an example, CCE reduces the memory footprint of the loss computation from 24 GB to 1 MB, and the total training-time memory consumption of the classifier head from 28 GB to 1 GB. To improve the throughput of CCE, we leverage the inherent sparsity of softmax and propose to skip elements of the gradient computation that have a negligible (i.e., below numerical precision) contribution to the gradient. Experiments demonstrate that the dramatic reduction in memory consumption is accomplished without sacrificing training speed or convergence.

The efficiency of attention is critical because its time complexity grows quadratically with sequence length. SageAttention2 addresses this by utilizing quantization to accelerate matrix multiplications (Matmul) in attention. To further accelerate SageAttention2, we propose to utilize the faster instruction of FP8 Matmul accumulated in FP16. The instruction is 2x faster than the FP8 Matmul used in SageAttention2. Our experiments show that SageAttention2++ achieves a 3.9x speedup over FlashAttention while maintaining the same attention accuracy as SageAttention2. This means SageAttention2++ effectively accelerates various models, including those for language, image, and video generation, with negligible end-to-end metrics loss. The code will be available at https://github.com/thu-ml/SageAttention.

We introduce QuaRot, a new Quantization scheme based on Rotations, which is able to quantize LLMs end-to-end, including all weights, activations, and KV cache in 4 bits. QuaRot rotates LLMs in a way that removes outliers from the hidden state without changing the output, making quantization easier. This computational invariance is applied to the hidden state (residual) of the LLM, as well as to the activations of the feed-forward components, aspects of the attention mechanism and to the KV cache. The result is a quantized model where all matrix multiplications are performed in 4-bits, without any channels identified for retention in higher precision. Our quantized LLaMa2-70B model has losses of at most 0.29 WikiText-2 perplexity and retains 99% of the zero-shot performance. Code is available at: https://github.com/spcl/QuaRot.

Low precision (LP) datatypes such as MXFP4 can accelerate matrix multiplications (GEMMs) and reduce training costs. However, directly using MXFP4 instead of BF16 during training significantly degrades model quality. In this work, we present the first near-lossless training recipe that uses MXFP4 GEMMs, which are 2times faster than FP8 on supported hardware. Our key insight is to compute unbiased gradient estimates with stochastic rounding (SR), resulting in more accurate model updates. However, directly applying SR to MXFP4 can result in high variance from block-level outliers, harming convergence. To overcome this, we use the random Hadamard tranform to theoretically bound the variance of SR. We train GPT models up to 6.7B parameters and find that our method induces minimal degradation over mixed-precision BF16 training. Our recipe computes >1/2 the training FLOPs in MXFP4, enabling an estimated speedup of >1.3times over FP8 and >1.7times over BF16 during backpropagation.

Self-attention has become a defacto choice for capturing global context in various vision applications. However, its quadratic computational complexity with respect to image resolution limits its use in real-time applications, especially for deployment on resource-constrained mobile devices. Although hybrid approaches have been proposed to combine the advantages of convolutions and self-attention for a better speed-accuracy trade-off, the expensive matrix multiplication operations in self-attention remain a bottleneck. In this work, we introduce a novel efficient additive attention mechanism that effectively replaces the quadratic matrix multiplication operations with linear element-wise multiplications. Our design shows that the key-value interaction can be replaced with a linear layer without sacrificing any accuracy. Unlike previous state-of-the-art methods, our efficient formulation of self-attention enables its usage at all stages of the network. Using our proposed efficient additive attention, we build a series of models called "SwiftFormer" which achieves state-of-the-art performance in terms of both accuracy and mobile inference speed. Our small variant achieves 78.5% top-1 ImageNet-1K accuracy with only 0.8 ms latency on iPhone 14, which is more accurate and 2x faster compared to MobileViT-v2. Code: https://github.com/Amshaker/SwiftFormer

In modern neural networks like Transformers, linear layers require significant memory to store activations during backward pass. This study proposes a memory reduction approach to perform backpropagation through linear layers. Since the gradients of linear layers are computed by matrix multiplications, we consider methods for randomized matrix multiplications and demonstrate that they require less memory with a moderate decrease of the test accuracy. Also, we investigate the variance of the gradient estimate induced by the randomized matrix multiplication. We compare this variance with the variance coming from gradient estimation based on the batch of samples. We demonstrate the benefits of the proposed method on the fine-tuning of the pre-trained RoBERTa model on GLUE tasks.

This paper proposes a novel approach for rendering a pre-trained Neural Radiance Field (NeRF) in real-time on resource-constrained devices. We introduce Re-ReND, a method enabling Real-time Rendering of NeRFs across Devices. Re-ReND is designed to achieve real-time performance by converting the NeRF into a representation that can be efficiently processed by standard graphics pipelines. The proposed method distills the NeRF by extracting the learned density into a mesh, while the learned color information is factorized into a set of matrices that represent the scene's light field. Factorization implies the field is queried via inexpensive MLP-free matrix multiplications, while using a light field allows rendering a pixel by querying the field a single time-as opposed to hundreds of queries when employing a radiance field. Since the proposed representation can be implemented using a fragment shader, it can be directly integrated with standard rasterization frameworks. Our flexible implementation can render a NeRF in real-time with low memory requirements and on a wide range of resource-constrained devices, including mobiles and AR/VR headsets. Notably, we find that Re-ReND can achieve over a 2.6-fold increase in rendering speed versus the state-of-the-art without perceptible losses in quality.

The rise of on-chip accelerators signifies a major shift in computing, driven by the growing demands of artificial intelligence (AI) and specialized applications. These accelerators have gained popularity due to their ability to substantially boost performance, cut energy usage, lower total cost of ownership (TCO), and promote sustainability. Intel's Advanced Matrix Extensions (AMX) is one such on-chip accelerator, specifically designed for handling tasks involving large matrix multiplications commonly used in machine learning (ML) models, image processing, and other computational-heavy operations. In this paper, we introduce a novel value-dependent timing side-channel vulnerability in Intel AMX. By exploiting this weakness, we demonstrate a software-based, value-dependent timing side-channel attack capable of inferring the sparsity of neural network weights without requiring any knowledge of the confidence score, privileged access or physical proximity. Our attack method can fully recover the sparsity of weights assigned to 64 input elements within 50 minutes, which is 631% faster than the maximum leakage rate achieved in the Hertzbleed attack.

Large Language Models (LLMs) pose significant hardware challenges related to memory requirements and computational ability. There are two mainstream quantization schemes for LLMs: coarse-grained (e.g., channel-wise) quantization and fine-grained (e.g., group-wise) quantization. Fine-grained quantization has smaller quantization loss, consequently achieving superior performance. However, when applied to weight-activation quantization, it disrupts continuous integer matrix multiplication, leading to inefficient inference. In this paper, we introduce Dual Grained Quantization (DGQ), a novel A8W4 quantization for LLM that maintains superior performance while ensuring fast inference speed. DSQ dequantizes the fine-grained INT4 weight into coarse-grained INT8 representation and preform matrix multiplication using INT8 kernels. Besides, we develop a two-phase grid search algorithm to simplify the determination of fine-grained and coarse-grained quantization scales. We also devise a percentile clipping schema for smoothing the activation outliers without the need for complex optimization techniques. Experimental results demonstrate that DGQ consistently outperforms prior methods across various LLM architectures and a wide range of tasks. Remarkably, by our implemented efficient CUTLASS kernel, we achieve 1.12 times memory reduction and 3.24 times speed gains comparing A16W4 implementation. These advancements enable efficient deployment of A8W4 LLMs for real-world applications.

In this work, we demonstrate that a simple two-layer neural network with standard activation functions can learn an arbitrary word operation in any finite group, provided sufficient width is available and exhibits grokking while doing so. To explain the mechanism by which this is achieved, we reframe the problem as that of learning a particular 3-tensor, which we show is typically of low rank. A key insight is that low-rank implementations of this tensor can be obtained by decomposing it along triplets of basic self-conjugate representations of the group and leveraging the fusion structure to rule out many components. Focusing on a phenomenologically similar but more tractable surrogate model, we show that the network is able to find such low-rank implementations (or approximations thereof), thereby using limited width to approximate the word-tensor in a generalizable way. In the case of the simple multiplication word, we further elucidate the form of these low-rank implementations, showing that the network effectively implements efficient matrix multiplication in the sense of Strassen. Our work also sheds light on the mechanism by which a network reaches such a solution under gradient descent.

The Multi-Head Attention mechanism is central to LLM operation, and multiple works target its compute and memory efficiency during training. While most works focus on approximating the scaled dot product, the memory consumption of the linear projections that compute the Q, K, and V tensors from the input x is often overlooked. To address this, we propose Point-Approximate Matrix Multiplication (PAMM), a novel tensor compression technique that reduces memory consumption of the Q,K,V projections in attention layers by a factor of up to times 512, effectively erasing their memory footprint, while achieving similar or better final perplexity. PAMM is fully composable with efficient attention techniques such as FlashAttention, making it a practical and complementary method for memory-efficient LLM training.

As large language models become increasingly prevalent, memory bandwidth constraints significantly limit inference throughput, motivating post-training quantization (PTQ). In this paper, we propose FireQ, a co-designed PTQ framework and an INT4-FP8 matrix multiplication kernel that accelerates LLM inference across all linear layers. Specifically, FireQ quantizes linear layer weights and key-values to INT4, and activations and queries to FP8, significantly enhancing throughput. Additionally, we introduce a three-stage pipelining for the prefill phase, which modifies the FlashAttention-3 kernel, effectively reducing time-to-first-token in the prefill phase. To minimize accuracy loss from quantization, we develop novel outlier smoothing techniques tailored separately for linear and attention layers. In linear layers, we explicitly use per-tensor scaling to prevent underflow caused by the FP8 quantization scaling factor of INT4 quantization, and channel-wise scaling to compensate for coarse granularity of INT4. In attention layers, we address quantization challenges posed by rotary positional embeddings (RoPE) by combining pre-RoPE and post-RoPE scaling strategies. FireQ significantly outperforms state-of-the-art methods, achieving 1.68x faster inference in feed-forward network layers on Llama2-7B and 1.26x faster prefill phase performance on Llama3-8B compared to QServe, with negligible accuracy loss.

As large language model (LLM) inference demands ever-greater resources, there is a rapid growing trend of using low-bit weights to shrink memory usage and boost inference efficiency. However, these low-bit LLMs introduce the need for mixed-precision matrix multiplication (mpGEMM), which is a crucial yet under-explored operation that involves multiplying lower-precision weights with higher-precision activations. Unfortunately, current hardware does not natively support mpGEMM, resulting in indirect and inefficient dequantization-based implementations. To address the mpGEMM requirements in low-bit LLMs, we explored the lookup table (LUT)-based approach for mpGEMM. However, a conventional LUT implementation falls short of its potential. To fully harness the power of LUT-based mpGEMM, we introduce LUT Tensor Core, a software-hardware co-design optimized for low-bit LLM inference. Specifically, we introduce software-based operator fusion and table symmetrization techniques to optimize table precompute and table storage, respectively. Then, LUT Tensor Core proposes the hardware design featuring an elongated tiling shape design to enhance table reuse and a bit-serial design to support various precision combinations in mpGEMM. Moreover, we design an end-to-end compilation stack with new instructions for LUT-based mpGEMM, enabling efficient LLM compilation and optimizations. The evaluation on low-bit LLMs (e.g., BitNet, LLAMA) shows that LUT Tensor Core achieves more than a magnitude of improvements on both compute density and energy efficiency.

Large language models (LLMs) show excellent performance but are compute- and memory-intensive. Quantization can reduce memory and accelerate inference. However, existing methods cannot maintain accuracy and hardware efficiency at the same time. We propose SmoothQuant, a training-free, accuracy-preserving, and general-purpose post-training quantization (PTQ) solution to enable 8-bit weight, 8-bit activation (W8A8) quantization for LLMs. Based on the fact that weights are easy to quantize while activations are not, SmoothQuant smooths the activation outliers by offline migrating the quantization difficulty from activations to weights with a mathematically equivalent transformation. SmoothQuant enables an INT8 quantization of both weights and activations for all the matrix multiplications in LLMs, including OPT, BLOOM, GLM, MT-NLG, and LLaMA family. We demonstrate up to 1.56x speedup and 2x memory reduction for LLMs with negligible loss in accuracy. SmoothQuant enables serving 530B LLM within a single node. Our work offers a turn-key solution that reduces hardware costs and democratizes LLMs. Code is available at https://github.com/mit-han-lab/smoothquant.

Parameter-Efficient Fine-Tuning (PEFT) methods, such as LoRA, significantly reduce the number of trainable parameters by introducing low-rank decomposition matrices. However, existing methods perform extensive matrix multiplications in domain specialization tasks, resulting in computational inefficiency and sub-optimal fine-tuning performance. Hence, we propose LoSiA(Low-Resources Subnet Integration Adaptation), an innovative method that dynamically localizes and optimizes critical parameters during the training process. Specifically, it identifies a sub-network using gradient sparsity analysis and optimizes it as the trainable target. This design enables effective high-rank adaptation by updating only the sub-network parameters, reducing the additional matrix multiplication. We also present LoSiA-Pro, a faster implementation of LoSiA, which reduces the training latency by about 27% compared to LoRA. Extensive evaluations show that our method achieves minimal performance drop compared to full fine-tuning, while requiring the least training time across domain specialization and common-sense reasoning tasks. Further analysis shows that LoSiA also reduces forgetting during continued training.

In recent years, Transformer-based language models have become the standard approach for natural language processing tasks. However, stringent throughput and latency requirements in industrial applications are limiting their adoption. To mitigate the gap, model compression techniques such as structured pruning are being used to improve inference efficiency. However, most existing neural network inference runtimes lack adequate support for structured sparsity. In this paper, we propose an efficient sparse deep learning inference software stack for Transformer-based language models where the weights are pruned with constant block size. Our sparse software accelerator leverages Intel Deep Learning Boost to maximize the performance of sparse matrix - dense matrix multiplication (commonly abbreviated as SpMM) on CPUs. Our SpMM kernel outperforms the existing sparse libraries (oneMKL, TVM, and LIBXSMM) by an order of magnitude on a wide range of GEMM shapes under 5 representative sparsity ratios (70%, 75%, 80%, 85%, 90%). Moreover, our SpMM kernel shows up to 5x speedup over dense GEMM kernel of oneDNN, a well-optimized dense library widely used in industry. We apply our sparse accelerator on widely-used Transformer-based language models including Bert-Mini, DistilBERT, Bert-Base, and BERT-Large. Our sparse inference software shows up to 1.5x speedup over Neural Magic's Deepsparse under same configurations on Xeon on Amazon Web Services under proxy production latency constraints. We also compare our solution with two framework-based inference solutions, ONNX Runtime and PyTorch, and demonstrate up to 37x speedup over ONNX Runtime and 345x over PyTorch on Xeon under the latency constraints. All the source code is publicly available on Github: https://github.com/intel/intel-extension-for-transformers.

Post-training quantization (PTQ) plays a crucial role in the democratization of large language models (LLMs). However, existing low-bit quantization and sparsification techniques are difficult to balance accuracy and efficiency due to the limited hardware support. For example, W4A8 can only achieve the same peak TOPS as W8A8 whereas the GPU-supported sparse data format (2:4 semi-structure sparse) is seldomly adopted due to the loss of accuracy. To bridge this gap, in this paper, we propose the Sparse-Quantized Format (SQ-format), which is a unified data format for quantization and sparsification potentially easily supported by new hardware and existing GPUs. SQ-format makes use of the fact that sparse matrix can be accelerated in high-precision, and low-precision matrix multiplication can also be accelerated accordingly. As such, SQ-format is proposed to achieve Pareto improvement between performance and throughput. This format is particularly suitable for activations with outlier inequality status and makes their static compression possible. We show the state-of-the-art PTQ performance with SQ-format, propose the hardware required to support it, and further offer the design exploration and insights for the next-generation AI accelerators.

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.

Quantization has been widely used to compress and accelerate inference of large language models (LLMs). Existing methods focus on exploring the per-token dynamic calibration to ensure both inference acceleration and model accuracy under 4-bit quantization. However, in autoregressive generation inference of long sequences, the overhead of repeated dynamic quantization and dequantization steps becomes considerably expensive. In this work, we propose MergeQuant, an accurate and efficient per-channel static quantization framework. MergeQuant integrates the per-channel quantization steps with the corresponding scalings and linear mappings through a Quantization Step Migration (QSM) method, thereby eliminating the quantization overheads before and after matrix multiplication. Furthermore, in view of the significant differences between the different channel ranges, we propose dimensional reconstruction and adaptive clipping to address the non-uniformity of quantization scale factors and redistribute the channel variations to the subsequent modules to balance the parameter distribution under QSM. Within the static quantization setting of W4A4, MergeQuant reduces the accuracy gap on zero-shot tasks compared to FP16 baseline to 1.3 points on Llama-2-70B model. On Llama-2-7B model, MergeQuant achieves up to 1.77x speedup in decoding, and up to 2.06x speedup in end-to-end compared to FP16 baseline.

Training deep neural networks (DNNs) is costly. Fortunately, Nvidia Ampere and Hopper GPUs can accelerate matrix multiplications twice as fast as a dense equivalent by implementing 2:4 sparsity. However, previous STE-based 2:4 pre-training methods (e.g. STE with hard-thresholding, SR-STE) suffer from optimization difficulties because of discontinuous pruning function. In this study, we comprehensively analyse the bottleneck of traditional N:M sparse training and recognize three drawbacks with discontinuity: incorrect descending direction, inability to predict the amount of descent and sparse mask oscillation. In light of this, we propose S-STE, a simple yet powerful 2:4 training method that contains two parts: to continuously project weights to be 2:4 sparse, and to rescale sparse weights with a per-tensor fixed scaling factor. Besides, we adopt minimum-variance unbiased estimation for activation gradient and FP8 quantization for whole process. Results show that our method surpasses previous 2:4 pre-training recipes and is comparable even with full parameter models. Our toolkit is available at https://github.com/huyz2023/2by4-pretrain.

This paper proposes an efficient multi-camera to Bird's-Eye-View (BEV) view transformation method for 3D perception, dubbed MatrixVT. Existing view transformers either suffer from poor transformation efficiency or rely on device-specific operators, hindering the broad application of BEV models. In contrast, our method generates BEV features efficiently with only convolutions and matrix multiplications (MatMul). Specifically, we propose describing the BEV feature as the MatMul of image feature and a sparse Feature Transporting Matrix (FTM). A Prime Extraction module is then introduced to compress the dimension of image features and reduce FTM's sparsity. Moreover, we propose the Ring \& Ray Decomposition to replace the FTM with two matrices and reformulate our pipeline to reduce calculation further. Compared to existing methods, MatrixVT enjoys a faster speed and less memory footprint while remaining deploy-friendly. Extensive experiments on the nuScenes benchmark demonstrate that our method is highly efficient but obtains results on par with the SOTA method in object detection and map segmentation tasks

Post-training compression of large language models (LLMs) largely relies on low-rank weight approximation, which represents each column of a weight matrix in a shared low-dimensional subspace. While this is a computationally efficient strategy, the imposed structural constraint is rigid and can lead to a noticeable model accuracy drop. In this work, we propose CoSpaDi (Compression via Sparse Dictionary Learning), a novel training-free compression framework that replaces low-rank decomposition with a more flexible structured sparse factorization in which each weight matrix is represented with a dense dictionary and a column-sparse coefficient matrix. This formulation enables a union-of-subspaces representation: different columns of the original weight matrix are approximated in distinct subspaces spanned by adaptively selected dictionary atoms, offering greater expressiveness than a single invariant basis. Crucially, CoSpaDi leverages a small calibration dataset to optimize the factorization such that the output activations of compressed projection layers closely match those of the original ones, thereby minimizing functional reconstruction error rather than mere weight approximation. This data-aware strategy preserves better model fidelity without any fine-tuning under reasonable compression ratios. Moreover, the resulting structured sparsity allows efficient sparse-dense matrix multiplication and is compatible with post-training quantization for further memory and latency gains. We evaluate CoSpaDi across multiple Llama and Qwen models under per-layer and per-group settings at 20-50\% compression ratios, demonstrating consistent superiority over state-of-the-art data-aware low-rank methods both in accuracy and perplexity. Our results establish structured sparse dictionary learning as a powerful alternative to conventional low-rank approaches for efficient LLM deployment.

We present any4, a learned 4-bit weight quantization solution for large language models (LLMs) providing arbitrary numeric representations without requiring pre-processing of weights or activations. any4 yields higher accuracy compared to other related 4-bit numeric representation types: int4, fp4 and nf4, as evaluated on a range of model sizes, generations and families (Llama 2, Llama 3, Mistral and Mixtral). While any4 does not require preprocessing of weights or activations, it is also competitive with orthogonal techniques that require such preprocessing (e.g., AWQ and GPTQ). We also experiment with any3 and any2 and show competitiveness at lower bits. Additionally, we show that we can calibrate using a single curated diverse sample rather than hundreds of samples from a dataset as done in most quantization approaches. We also open source tinygemm, a latency optimized GPU matrix multiplication library for LLMs, that implements any4 using a GPU-efficient lookup table strategy along with other common quantization methods. We open source our code at https://github.com/facebookresearch/any4 .

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

Large Language Models (LLMs) stand out for their impressive performance in intricate language modeling tasks. However, their demanding computational and memory needs pose obstacles for broad use on edge devices. Quantization is then introduced to boost LLMs' on-device efficiency. Recent works show that 8-bit or lower weight quantization is feasible with minimal impact on end-to-end task performance, while the activation is still not quantized. On the other hand, mainstream commodity edge devices still struggle to execute these sub-8-bit quantized networks effectively. In this paper, we propose Agile-Quant, an activation-guided quantization framework for popular Large Language Models (LLMs), and implement an end-to-end accelerator on multiple edge devices for faster inference. Considering the hardware profiling and activation analysis, we first introduce a basic activation quantization strategy to balance the trade-off of task performance and real inference speed. Then we leverage the activation-aware token pruning technique to reduce the outliers and the adverse impact on attentivity. Ultimately, we utilize the SIMD-based 4-bit multiplier and our efficient TRIP matrix multiplication to implement the accelerator for LLMs on the edge. We apply our framework on different scales of LLMs including LLaMA, OPT, and BLOOM with 4-bit or 8-bit for the activation and 4-bit for the weight quantization. Experiments show that Agile-Quant achieves simultaneous quantization of model weights and activations while maintaining task performance comparable to existing weight-only quantization methods. Moreover, in the 8- and 4-bit scenario, Agile-Quant achieves an on-device speedup of up to 2.55x compared to its FP16 counterparts across multiple edge devices, marking a pioneering advancement in this domain.

Enabling On-Device Learning (ODL) for Ultra-Low-Power Micro-Controller Units (MCUs) is a key step for post-deployment adaptation and fine-tuning of Deep Neural Network (DNN) models in future TinyML applications. This paper tackles this challenge by introducing a novel reduced precision optimization technique for ODL primitives on MCU-class devices, leveraging the State-of-Art advancements in RISC-V RV32 architectures with support for vectorized 16-bit floating-point (FP16) Single-Instruction Multiple-Data (SIMD) operations. Our approach for the Forward and Backward steps of the Back-Propagation training algorithm is composed of specialized shape transform operators and Matrix Multiplication (MM) kernels, accelerated with parallelization and loop unrolling. When evaluated on a single training step of a 2D Convolution layer, the SIMD-optimized FP16 primitives result up to 1.72times faster than the FP32 baseline on a RISC-V-based 8+1-core MCU. An average computing efficiency of 3.11 Multiply and Accumulate operations per clock cycle (MAC/clk) and 0.81 MAC/clk is measured for the end-to-end training tasks of a ResNet8 and a DS-CNN for Image Classification and Keyword Spotting, respectively -- requiring 17.1 ms and 6.4 ms on the target platform to compute a training step on a single sample. Overall, our approach results more than two orders of magnitude faster than existing ODL software frameworks for single-core MCUs and outperforms by 1.6 times previous FP32 parallel implementations on a Continual Learning setup.

Transformer is a transformative framework that models sequential data and has achieved remarkable performance on a wide range of tasks, but with high computational and energy cost. To improve its efficiency, a popular choice is to compress the models via binarization which constrains the floating-point values into binary ones to save resource consumption owing to cheap bitwise operations significantly. However, existing binarization methods only aim at minimizing the information loss for the input distribution statistically, while ignoring the pairwise similarity modeling at the core of the attention. To this end, we propose a new binarization paradigm customized to high-dimensional softmax attention via kernelized hashing, called EcoFormer, to map the original queries and keys into low-dimensional binary codes in Hamming space. The kernelized hash functions are learned to match the ground-truth similarity relations extracted from the attention map in a self-supervised way. Based on the equivalence between the inner product of binary codes and the Hamming distance as well as the associative property of matrix multiplication, we can approximate the attention in linear complexity by expressing it as a dot-product of binary codes. Moreover, the compact binary representations of queries and keys enable us to replace most of the expensive multiply-accumulate operations in attention with simple accumulations to save considerable on-chip energy footprint on edge devices. Extensive experiments on both vision and language tasks show that EcoFormer consistently achieves comparable performance with standard attentions while consuming much fewer resources. For example, based on PVTv2-B0 and ImageNet-1K, Ecoformer achieves a 73% on-chip energy footprint reduction with only a 0.33% performance drop compared to the standard attention. Code is available at https://github.com/ziplab/EcoFormer.

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

Large Language Models (LLMs) demonstrate exceptional performance but entail significant memory and computational costs, restricting their practical deployment. While existing INT4/INT8 quantization reduces these costs, they often degrade accuracy or lack optimal efficiency. INT6 quantization offers a superior trade-off between model accuracy and inference efficiency, but lacks hardware support in modern GPUs, forcing emulation via higher-precision arithmetic units that limit acceleration. In this paper, we propose FlexQ, a novel post-training INT6 quantization framework combining algorithmic innovation with system-level optimizations. FlexQ employs uniform 6-bit weight quantization across all layers, with adaptive retention of 8-bit activations in layers identified through layer-wise sensitivity analysis. To maximize hardware efficiency, we develop a specialized high-performance GPU kernel supporting matrix multiplication for W6A6 and W6A8 representations via Binary Tensor Core (BTC) equivalents, effectively bypassing the lack of native INT6 tensor cores. Evaluations on LLaMA models show FlexQ maintains near-FP16 accuracy, with perplexity increases of no more than 0.05. The proposed kernel achieves an average 1.39times speedup over ABQ-LLM on LLaMA-2-70B linear layers. End-to-end, FlexQ delivers 1.33times inference acceleration and 1.21times memory savings over SmoothQuant. Code is released at https://github.com/FlyFoxPlayer/FlexQ.

Selective state space models (SSMs), such as Mamba, highly excel at capturing long-range dependencies in 1D sequential data, while their applications to 2D vision tasks still face challenges. Current visual SSMs often convert images into 1D sequences and employ various scanning patterns to incorporate local spatial dependencies. However, these methods are limited in effectively capturing the complex image spatial structures and the increased computational cost caused by the lengthened scanning paths. To address these limitations, we propose Spatial-Mamba, a novel approach that establishes neighborhood connectivity directly in the state space. Instead of relying solely on sequential state transitions, we introduce a structure-aware state fusion equation, which leverages dilated convolutions to capture image spatial structural dependencies, significantly enhancing the flow of visual contextual information. Spatial-Mamba proceeds in three stages: initial state computation in a unidirectional scan, spatial context acquisition through structure-aware state fusion, and final state computation using the observation equation. Our theoretical analysis shows that Spatial-Mamba unifies the original Mamba and linear attention under the same matrix multiplication framework, providing a deeper understanding of our method. Experimental results demonstrate that Spatial-Mamba, even with a single scan, attains or surpasses the state-of-the-art SSM-based models in image classification, detection and segmentation. Source codes and trained models can be found at https://github.com/EdwardChasel/Spatial-Mamba.

High-performance sparse matrix-matrix (SpMM) multiplication is paramount for science and industry, as the ever-increasing sizes of data prohibit using dense data structures. Yet, existing hardware, such as Tensor Cores (TC), is ill-suited for SpMM, as it imposes strict constraints on data structures that cannot be met by unstructured sparsity found in many applications. To address this, we introduce (S)parse (Ma)trix Matrix (T)ensor Core-accelerated (SMaT): a novel SpMM library that utilizes TCs for unstructured sparse matrices. Our block-sparse library leverages the low-level CUDA MMA (matrix-matrix-accumulate) API, maximizing the performance offered by modern GPUs. Algorithmic optimizations such as sparse matrix permutation further improve performance by minimizing the number of non-zero blocks. The evaluation on NVIDIA A100 shows that SMaT outperforms SotA libraries (DASP, cuSPARSE, and Magicube) by up to 125x (on average 2.6x). SMaT can be used to accelerate many workloads in scientific computing, large-model training, inference, and others.

We show that both an LSTM and a unitary-evolution recurrent neural network (URN) can achieve encouraging accuracy on two types of syntactic patterns: context-free long distance agreement, and mildly context-sensitive cross serial dependencies. This work extends recent experiments on deeply nested context-free long distance dependencies, with similar results. URNs differ from LSTMs in that they avoid non-linear activation functions, and they apply matrix multiplication to word embeddings encoded as unitary matrices. This permits them to retain all information in the processing of an input string over arbitrary distances. It also causes them to satisfy strict compositionality. URNs constitute a significant advance in the search for explainable models in deep learning applied to NLP.

Mobile vision transformers (MobileViT) can achieve state-of-the-art performance across several mobile vision tasks, including classification and detection. Though these models have fewer parameters, they have high latency as compared to convolutional neural network-based models. The main efficiency bottleneck in MobileViT is the multi-headed self-attention (MHA) in transformers, which requires O(k^2) time complexity with respect to the number of tokens (or patches) k. Moreover, MHA requires costly operations (e.g., batch-wise matrix multiplication) for computing self-attention, impacting latency on resource-constrained devices. This paper introduces a separable self-attention method with linear complexity, i.e. O(k). A simple yet effective characteristic of the proposed method is that it uses element-wise operations for computing self-attention, making it a good choice for resource-constrained devices. The improved model, MobileViTv2, is state-of-the-art on several mobile vision tasks, including ImageNet object classification and MS-COCO object detection. With about three million parameters, MobileViTv2 achieves a top-1 accuracy of 75.6% on the ImageNet dataset, outperforming MobileViT by about 1% while running 3.2times faster on a mobile device. Our source code is available at: https://github.com/apple/ml-cvnets

Deep learning frameworks commonly implement convolution operators with GEMM-based algorithms. In these algorithms, convolution is implemented on top of matrix-matrix multiplication (GEMM) functions, provided by highly optimized BLAS libraries. Convolutions with 1x1 kernels can be directly represented as a GEMM call, but convolutions with larger kernels require a special memory layout transformation - im2col or im2row - to fit into GEMM interface. The Indirect Convolution algorithm provides the efficiency of the GEMM primitive without the overhead of im2col transformation. In contrast to GEMM-based algorithms, the Indirect Convolution does not reshuffle the data to fit into the GEMM primitive but introduces an indirection buffer - a buffer of pointers to the start of each row of image pixels. This broadens the application of our modified GEMM function to convolutions with arbitrary kernel size, padding, stride, and dilation. The Indirect Convolution algorithm reduces memory overhead proportionally to the number of input channels and outperforms the GEMM-based algorithm by up to 62% on convolution parameters which involve im2col transformations in GEMM-based algorithms. This, however, comes at cost of minor performance reduction on 1x1 stride-1 convolutions.

Quantizing the activation, weight, and gradient to 4-bit is promising to accelerate neural network training. However, existing 4-bit training methods require custom numerical formats which are not supported by contemporary hardware. In this work, we propose a training method for transformers with all matrix multiplications implemented with the INT4 arithmetic. Training with an ultra-low INT4 precision is challenging. To achieve this, we carefully analyze the specific structures of activation and gradients in transformers to propose dedicated quantizers for them. For forward propagation, we identify the challenge of outliers and propose a Hadamard quantizer to suppress the outliers. For backpropagation, we leverage the structural sparsity of gradients by proposing bit splitting and leverage score sampling techniques to quantize gradients accurately. Our algorithm achieves competitive accuracy on a wide range of tasks including natural language understanding, machine translation, and image classification. Unlike previous 4-bit training methods, our algorithm can be implemented on the current generation of GPUs. Our prototypical linear operator implementation is up to 2.2 times faster than the FP16 counterparts and speeds up the training by up to 35.1%.

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.

With the fast growth of parameter size, it becomes increasingly challenging to deploy large generative models as they typically require large GPU memory consumption and massive computation. Unstructured model pruning has been a common approach to reduce both GPU memory footprint and the overall computation while retaining good model accuracy. However, the existing solutions do not provide a highly-efficient support for handling unstructured sparsity on modern GPUs, especially on the highly-structured Tensor Core hardware. Therefore, we propose Flash-LLM for enabling low-cost and highly-efficient large generative model inference with the sophisticated support of unstructured sparsity on high-performance but highly restrictive Tensor Cores. Based on our key observation that the main bottleneck of generative model inference is the several skinny matrix multiplications for which Tensor Cores would be significantly under-utilized due to low computational intensity, we propose a general Load-as-Sparse and Compute-as-Dense methodology for unstructured sparse matrix multiplication. The basic insight is to address the significant memory bandwidth bottleneck while tolerating redundant computations that are not critical for end-to-end performance on Tensor Cores. Based on this, we design an effective software framework for Tensor Core based unstructured SpMM, leveraging on-chip resources for efficient sparse data extraction and computation/memory-access overlapping. At SpMM kernel level, Flash-LLM significantly outperforms the state-of-the-art library, i.e., Sputnik and SparTA by an average of 2.9x and 1.5x, respectively. At end-to-end framework level on OPT-30B/66B/175B models, for tokens per GPU-second, Flash-LLM achieves up to 3.8x and 3.6x improvement over DeepSpeed and FasterTransformer, respectively, with significantly lower inference cost.

The lottery ticket hypothesis (LTH) has increased attention to pruning neural networks at initialization. We study this problem in the linear setting. We show that finding a sparse mask at initialization is equivalent to the sketching problem introduced for efficient matrix multiplication. This gives us tools to analyze the LTH problem and gain insights into it. Specifically, using the mask found at initialization, we bound the approximation error of the pruned linear model at the end of training. We theoretically justify previous empirical evidence that the search for sparse networks may be data independent. By using the sketching perspective, we suggest a generic improvement to existing algorithms for pruning at initialization, which we show to be beneficial in the data-independent case.

Automatic differentiation through digital signal processing algorithms for virtual analogue modelling has recently gained popularity. These algorithms are typically more computationally efficient than black-box neural networks that rely on dense matrix multiplications. Due to their differentiable nature, they can be integrated with neural networks and jointly trained using gradient descent algorithms, resulting in more efficient systems. Furthermore, signal processing algorithms have significantly fewer parameters than neural networks, allowing the application of the Newton-Raphson method. This method offers faster and more robust convergence than gradient descent at the cost of quadratic storage. This paper presents a method to emulate analogue levelling amplifiers using a feed-forward digital compressor with parameters optimised via the Newton-Raphson method. We demonstrate that a digital compressor can successfully approximate the behaviour of our target unit, the Teletronix LA-2A. Different strategies for computing the Hessian matrix are benchmarked. We leverage parallel algorithms for recursive filters to achieve efficient training on modern GPUs. The resulting model is made into a VST plugin and is open-sourced at https://github.com/aim-qmul/4a2a.

Despite the significant potential of FP8 data formats for large language model (LLM) pre-training, their adoption has been limited due to challenges in maintaining stability at scale. Existing approaches often rely on suboptimal fine-grained FP8 kernels or fall back to higher-precision matrix multiplications (GEMMs) in sensitive components, such as attention projections, compromising potential throughput gains. We introduce a new class of LLM architectures that, for the first time, support FP8 computation for all GEMMs within transformer blocks during both forward and backward passes. This enables unprecedented throughput gains, particularly at scale, while matching the downstream performance of standard BF16 training. Our architecture design reduces large outlier activations, promoting stable long-term FP8 training. In addition, we identify key metrics to monitor low-precision training and predict potential future divergences.

Activation sparsity can enable practical inference speedups in large language models (LLMs) by reducing the compute and memory-movement required for matrix multiplications during the forward pass. However, existing methods face limitations that inhibit widespread adoption. Some approaches are tailored towards older models with ReLU-based sparsity, while others require extensive continued pre-training on up to hundreds of billions of tokens. This paper describes TEAL, a simple training-free method that applies magnitude-based activation sparsity to hidden states throughout the entire model. TEAL achieves 40-50% model-wide sparsity with minimal performance degradation across Llama-2, Llama-3, and Mistral families, with sizes varying from 7B to 70B. We improve existing sparse kernels and demonstrate wall-clock decoding speed-ups of up to 1.53times and 1.8times at 40% and 50% model-wide sparsity. TEAL is compatible with weight quantization, enabling further efficiency gains.

In this work, we investigate a typical scenario in code generation where a developer edits existing code in real time and requests a code assistant, e.g., a large language model, to re-predict the next token or next line on the fly. Naively, the LLM needs to re-encode the entire KV cache to provide an accurate prediction. However, this process is computationally expensive, especially when the sequence length is long. Simply encoding the edited subsequence and integrating it to the original KV cache meets the temporal confusion problem, leading to significantly worse performance. We address this efficiency and accuracy trade-off by introducing \textbf{Positional \textbf{Integrity Encoding} (PIE). Building upon the rotary positional encoding, PIE first removes the rotary matrices in the Key cache that introduce temporal confusion and then reapplies the correct rotary matrices. This process ensures that positional relationships between tokens are correct and requires only a single round of matrix multiplication. We validate the effectiveness of PIE through extensive experiments on the RepoBench-C-8k dataset, utilizing DeepSeek-Coder models with 1.3B, 6.7B, and 33B parameters. Our evaluation includes three real-world coding tasks: code insertion, code deletion, and multi-place code editing. Results demonstrate that PIE reduces computational overhead by over 85% compared to the standard full recomputation approach across all model sizes and tasks while well approximating the model performance.

Modern transformer-based Large Language Models (LLMs) are constructed with a series of decoder blocks. Each block comprises three key components: (1) QKV generation, (2) multi-head attention, and (3) feed-forward networks. In batched processing, QKV generation and feed-forward networks involve compute-intensive matrix-matrix multiplications (GEMM), while multi-head attention requires bandwidth-heavy matrix-vector multiplications (GEMV). Machine learning accelerators like TPUs or NPUs are proficient in handling GEMM but are less efficient for GEMV computations. Conversely, Processing-in-Memory (PIM) technology is tailored for efficient GEMV computation, while it lacks the computational power to handle GEMM effectively. Inspired by this insight, we propose NeuPIMs, a heterogeneous acceleration system that jointly exploits a conventional GEMM-focused NPU and GEMV-optimized PIM devices. The main challenge in efficiently integrating NPU and PIM lies in enabling concurrent operations on both platforms, each addressing a specific kernel type. First, existing PIMs typically operate in a "blocked" mode, allowing only either NPU or PIM to be active at any given time. Second, the inherent dependencies between GEMM and GEMV in LLMs restrict their parallel processing. To tackle these challenges, NeuPIMs is equipped with dual row buffers in each bank, facilitating the simultaneous management of memory read/write operations and PIM commands. Further, NeuPIMs employs a runtime sub-batch interleaving technique to maximize concurrent execution, leveraging batch parallelism to allow two independent sub-batches to be pipelined within a single NeuPIMs device. Our evaluation demonstrates that compared to GPU-only, NPU-only, and a na\"ive NPU+PIM integrated acceleration approaches, NeuPIMs achieves 3times, 2.4times and 1.6times throughput improvement, respectively.

We discuss an unfortunate mistake, for a Dirac free particle, in the last Fermi lecture notes on quantum mechanics, in a course given at the University of Chicago in winter and spring of 1954. As is demonstrated, the correct result can be obtained by a simple matrix multiplication. An attempt to collect a relevant bibliography is made.

Very recently, a variety of vision transformer architectures for dense prediction tasks have been proposed and they show that the design of spatial attention is critical to their success in these tasks. In this work, we revisit the design of the spatial attention and demonstrate that a carefully-devised yet simple spatial attention mechanism performs favourably against the state-of-the-art schemes. As a result, we propose two vision transformer architectures, namely, Twins-PCPVT and Twins-SVT. Our proposed architectures are highly-efficient and easy to implement, only involving matrix multiplications that are highly optimized in modern deep learning frameworks. More importantly, the proposed architectures achieve excellent performance on a wide range of visual tasks, including image level classification as well as dense detection and segmentation. The simplicity and strong performance suggest that our proposed architectures may serve as stronger backbones for many vision tasks. Our code is released at https://github.com/Meituan-AutoML/Twins .

Transformer-based large language models (e.g., BERT and GPT) achieve great success, and fine-tuning, which tunes a pre-trained model on a task-specific dataset, is the standard practice to utilize these models for downstream tasks. However, Transformer fine-tuning has long running time and high memory consumption due to the large size of the models. We propose the SPT system to fine-tune Transformer-based models efficiently by introducing sparsity. We observe that the memory consumption of Transformer mainly comes from storing attention weights for multi-head attention (MHA), and the majority of running time is spent on feed-forward network (FFN). Thus, we design the sparse MHA module, which computes and stores only large attention weights to reduce memory consumption, and the routed FFN module, which dynamically activates a subset of model parameters for each token to reduce computation cost. We implement SPT on PyTorch and customize CUDA kernels to run sparse MHA and routed FFN efficiently. Specifically, we use product quantization to identify the large attention weights and compute attention via sparse matrix multiplication for sparse MHA. For routed FFN, we batch the tokens according to their activated model parameters for efficient computation. We conduct extensive experiments to evaluate SPT on various model configurations. The results show that SPT consistently outperforms well-optimized baselines, reducing the peak memory consumption by up to 50% and accelerating fine-tuning by up to 2.2x.

We introduce Evolution Guided General Optimization via Low-rank Learning (EGGROLL), an evolution strategies (ES) algorithm designed to scale backprop-free optimization to large population sizes for modern large neural network architectures with billions of parameters. ES is a set of powerful blackbox optimisation methods that can handle non-differentiable or noisy objectives with excellent scaling potential through parallelisation. Na{ï}ve ES becomes prohibitively expensive at scale due to the computational and memory costs associated with generating matrix perturbations EinR^{mtimes n} and the batched matrix multiplications needed to compute per-member forward passes. EGGROLL overcomes these bottlenecks by generating random matrices Ain R^{mtimes r}, Bin R^{ntimes r} with rll min(m,n) to form a low-rank matrix perturbation A B^top that are used in place of the full-rank perturbation E. As the overall update is an average across a population of N workers, this still results in a high-rank update but with significant memory and computation savings, reducing the auxiliary storage from mn to r(m+n) per layer and the cost of a forward pass from O(mn) to O(r(m+n)) when compared to full-rank ES. A theoretical analysis reveals our low-rank update converges to the full-rank update at a fast Oleft(1{r}right) rate. Our experiments show that (1) EGGROLL does not compromise the performance of ES in tabula-rasa RL settings, despite being faster, (2) it is competitive with GRPO as a technique for improving LLM reasoning, and (3) EGGROLL enables stable pre-training of nonlinear recurrent language models that operate purely in integer datatypes.

Low-Rank Adaptation (LoRA), which introduces a product of two trainable low-rank matrices into frozen pre-trained weights, is widely used for efficient fine-tuning of language models in federated learning (FL). However, when combined with differentially private stochastic gradient descent (DP-SGD), LoRA faces substantial noise amplification: DP-SGD perturbs per-sample gradients, and the matrix multiplication of the LoRA update (BA) intensifies this effect. Freezing one matrix (e.g., A) reduces the noise but restricts model expressiveness, often resulting in suboptimal adaptation. To address this, we propose FedSVD, a simple yet effective method that introduces a global reparameterization based on singular value decomposition (SVD). In our approach, each client optimizes only the B matrix and transmits it to the server. The server aggregates the B matrices, computes the product BA using the previous A, and refactorizes the result via SVD. This yields a new adaptive A composed of the orthonormal right singular vectors of BA, and an updated B containing the remaining SVD components. This reparameterization avoids quadratic noise amplification, while allowing A to better capture the principal directions of the aggregate updates. Moreover, the orthonormal structure of A bounds the gradient norms of B and preserves more signal under DP-SGD, as confirmed by our theoretical analysis. As a result, FedSVD consistently improves stability and performance across a variety of privacy settings and benchmarks, outperforming relevant baselines under both private and non-private regimes.

Cloud-based services are making the outsourcing of sensitive client data increasingly common. Although homomorphic encryption (HE) offers strong privacy guarantee, it requires substantially more resources than computing on plaintext, often leading to unacceptably large latencies in getting the results. HE accelerators have emerged to mitigate this latency issue, but with the high cost of ASICs. In this paper we show that HE primitives can be converted to AI operators and accelerated on existing ASIC AI accelerators, like TPUs, which are already widely deployed in the cloud. Adapting such accelerators for HE requires (1) supporting modular multiplication, (2) high-precision arithmetic in software, and (3) efficient mapping on matrix engines. We introduce the CROSS compiler (1) to adopt Barrett reduction to provide modular reduction support using multiplier and adder, (2) Basis Aligned Transformation (BAT) to convert high-precision multiplication as low-precision matrix-vector multiplication, (3) Matrix Aligned Transformation (MAT) to covert vectorized modular operation with reduction into matrix multiplication that can be efficiently processed on 2D spatial matrix engine. Our evaluation of CROSS on a Google TPUv4 demonstrates significant performance improvements, with up to 161x and 5x speedup compared to the previous work on many-core CPUs and V100. The kernel-level codes are open-sourced at https://github.com/google/jaxite/tree/main/jaxite_word.

We introduce three new attention mechanisms that outperform standard multi-head attention in terms of efficiency and learning capabilities, thereby improving the performance and broader deployability of Transformer models. Our first contribution is Optimised Attention, which performs similarly to standard attention, but has 3/4 as many parameters and one matrix multiplication fewer per head. Next, we introduce Efficient Attention, which performs on par with standard attention with only 1/2 as many parameters as many parameters and two matrix multiplications fewer per head and is up to twice as fast as standard attention. Lastly, we introduce Super Attention, which surpasses standard attention by a significant margin in both vision and natural language processing tasks while having fewer parameters and matrix multiplications. In addition to providing rigorous mathematical comparisons, we evaluate the presented attention mechanisms on MNIST, CIFAR100, IMDB Movie Reviews, and Amazon Reviews datasets.

Sliding Window Sum algorithms have been successfully used for training and inference of Deep Neural Networks. We have shown before how both pooling and convolution 1-D primitives could be expressed as sliding sums and evaluated by the compute kernels with a shared structure. In this paper, we present an extensive study of the Sliding Window convolution technique as a more efficient alternative to the commonly used General Matrix Multiplication (GEMM) based convolution in Deep Neural Networks (DNNs). The Sliding Window technique addresses the memory bloating problem and demonstrates a significant speedup in 2-D convolution. We explore the performance of this technique on a range of implementations, including custom kernels for specific filter sizes. Our results suggest that the Sliding Window computation kernels can outperform GEMM-based convolution on a CPU and even on dedicated hardware accelerators. This could promote a wider adoption of AI on low-power and low-memory devices without the need for specialized hardware. We also discuss the compatibility of model compression methods and optimized network architectures with the Sliding Window technique, encouraging further research in these areas.

Resource scheduling and allocation is a critical component of many high impact systems ranging from congestion control to cloud computing. Finding more optimal solutions to these problems often has significant impact on resource and time savings, reducing device wear-and-tear, and even potentially improving carbon emissions. In this paper, we focus on a specific instance of a scheduling problem, namely the memory mapping problem that occurs during compilation of machine learning programs: That is, mapping tensors to different memory layers to optimize execution time. We introduce an approach for solving the memory mapping problem using Reinforcement Learning. RL is a solution paradigm well-suited for sequential decision making problems that are amenable to planning, and combinatorial search spaces with high-dimensional data inputs. We formulate the problem as a single-player game, which we call the mallocGame, such that high-reward trajectories of the game correspond to efficient memory mappings on the target hardware. We also introduce a Reinforcement Learning agent, mallocMuZero, and show that it is capable of playing this game to discover new and improved memory mapping solutions that lead to faster execution times on real ML workloads on ML accelerators. We compare the performance of mallocMuZero to the default solver used by the Accelerated Linear Algebra (XLA) compiler on a benchmark of realistic ML workloads. In addition, we show that mallocMuZero is capable of improving the execution time of the recently published AlphaTensor matrix multiplication model.

Transformer-based models, exemplified by GPT-3, ChatGPT, and GPT-4, have recently garnered considerable attention in both academia and industry due to their promising performance in general language tasks. Nevertheless, these models typically involve computationally encoding processes, and in some cases, decoding processes as well, both of which are fundamentally large-scale matrix multiplication. These operations bring the inevitable challenges of massive computation resources and huge memory footprint, usually requiring at least 10^23 FLOPs and hundreds of gigabytes, respectively. A common method to address this issue is to reduce the computational and memory requirements by applying layerwise quantization to the transformer, replacing the usual fp32 data type with a low-bit equivalent. Unfortunately, this method often leads to decreased model accuracy and necessitates time-consuming retraining. Such retraining not only requires fine-tuning skills but also substantial computational resources, posing challenges for users. To specifically tackle these issues, we propose BCT, a framework of blockwise compression for transformers without retraining, aiming to facilitate model deployment. Unlike layerwise compression methods, BCT achieves finer compression of the entire transformer by operating blockwise. This method mitigates data distribution deviation caused by quantization, eliminating the requirement for retraining. BCT effectively compresses all components of the model, including but not limited to the embedding, matrix multiplication, GELU, Softmax, layer normalization, and intermediate results. In a case study, an efficient model is compressed by BCT achieving up to 7.988x compression. Subsequently, we also evaluate it on several General Language Understanding Evaluation (GLUE) datasets.

Recent advances in artificial intelligence (AI), particularly deep learning, have led to widespread adoption across various applications. Yet, a fundamental challenge persists: how can we verify the correctness of AI model inference when model owners cannot (or will not) reveal their parameters? These parameters represent enormous training costs and valuable intellectual property, making transparent verification difficult. In this paper, we introduce a zero-knowledge framework capable of verifying deep learning inference without exposing model internal parameters. Built on recursively composed zero-knowledge proofs and requiring no trusted setup, our framework supports both linear and nonlinear neural network layers, including matrix multiplication, normalization, softmax, and SiLU. Leveraging the Fiat-Shamir heuristic, we obtain a succinct non-interactive argument of knowledge (zkSNARK) with constant-size proofs. To demonstrate the practicality of our approach, we translate the DeepSeek model into a fully SNARK-verifiable version named ZK-DeepSeek and show experimentally that our framework delivers both efficiency and flexibility in real-world AI verification workloads.

Linear Transformers and State Space Models have emerged as efficient alternatives to softmax Transformers for causal sequence modeling, enabling parallel training via matrix multiplication and efficient RNN-style inference. However, despite their success in causal tasks, no unified framework exists for applying Linear Transformers to bidirectional sequence modeling. We introduce LION, the first framework to systematically extend Linear Transformers to the bidirectional setting. LION generalizes three core representations commonly used in the causal case - full Linear Attention , bidirectional RNN, and chunkwise parallel form - to the bidirectional setting. These forms are theoretically equivalent and enable models to exploit the strengths of each during training and inference. We prove that a broad class of Linear Transformers can be extended using LION and validate our framework via three core examples based on the choice of decay type: LION-LIT, the bidirectional extension of arXiv:2006.16236; LION-D, based on arXiv:2307.08621; and LION-S, a variant using selective decay arXiv:2103.02143, arXiv:2312.0075. Across standard bidirectional tasks, LION enables models to match or exceed the performance of softmax Transformers, while offering significantly faster training and more efficient inference than existing State Space Models.

Large Language Models (LLMs) face significant deployment challenges due to their substantial resource requirements. While low-bit quantized weights can reduce memory usage and improve inference efficiency, current hardware lacks native support for mixed-precision General Matrix Multiplication (mpGEMM), resulting in inefficient dequantization-based implementations. Moreover, uniform quantization methods often fail to capture weight distributions adequately, leading to performance degradation. We propose GANQ (GPU-Adaptive Non-Uniform Quantization), a layer-wise post-training non-uniform quantization framework optimized for hardware-efficient lookup table-based mpGEMM. GANQ achieves superior quantization performance by utilizing a training-free, GPU-adaptive optimization algorithm to efficiently reduce layer-wise quantization errors. Extensive experiments demonstrate GANQ's ability to reduce the perplexity gap from the FP16 baseline compared to state-of-the-art methods for both 3-bit and 4-bit quantization. Furthermore, when deployed on a single NVIDIA RTX 4090 GPU, GANQ's quantized models achieve up to 2.57times speedup over the baseline, advancing memory and inference efficiency in LLM deployment.

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.

The demand for efficient processing of deep neural networks (DNNs) on embedded devices is a significant challenge limiting their deployment. Exploiting sparsity in the network's feature maps is one of the ways to reduce its inference latency. It is known that unstructured sparsity results in lower accuracy degradation with respect to structured sparsity but the former needs extensive inference engine changes to get latency benefits. To tackle this challenge, we propose a solution to induce semi-structured activation sparsity exploitable through minor runtime modifications. To attain high speedup levels at inference time, we design a sparse training procedure with awareness of the final position of the activations while computing the General Matrix Multiplication (GEMM). We extensively evaluate the proposed solution across various models for image classification and object detection tasks. Remarkably, our approach yields a speed improvement of 1.25 times with a minimal accuracy drop of 1.1% for the ResNet18 model on the ImageNet dataset. Furthermore, when combined with a state-of-the-art structured pruning method, the resulting models provide a good latency-accuracy trade-off, outperforming models that solely employ structured pruning techniques.

Secure multi-party computation (MPC) enables computation directly on encrypted data and protects both data and model privacy in deep learning inference. However, existing neural network architectures, including Vision Transformers (ViTs), are not designed or optimized for MPC and incur significant latency overhead. We observe Softmax accounts for the major latency bottleneck due to a high communication complexity, but can be selectively replaced or linearized without compromising the model accuracy. Hence, in this paper, we propose an MPC-friendly ViT, dubbed MPCViT, to enable accurate yet efficient ViT inference in MPC. Based on a systematic latency and accuracy evaluation of the Softmax attention and other attention variants, we propose a heterogeneous attention optimization space. We also develop a simple yet effective MPC-aware neural architecture search algorithm for fast Pareto optimization. To further boost the inference efficiency, we propose MPCViT+, to jointly optimize the Softmax attention and other network components, including GeLU, matrix multiplication, etc. With extensive experiments, we demonstrate that MPCViT achieves 1.9%, 1.3% and 3.6% higher accuracy with 6.2x, 2.9x and 1.9x latency reduction compared with baseline ViT, MPCFormer and THE-X on the Tiny-ImageNet dataset, respectively. MPCViT+ further achieves a better Pareto front compared with MPCViT. The code and models for evaluation are available at https://github.com/PKU-SEC-Lab/mpcvit.

Large language models have led to state-of-the-art accuracies across a range of tasks. However,training large language model needs massive computing resource, as more and more open source pre-training models are available, it is worthy to study how to take full advantage of available model. We find a method to save training time and resource cost by changing the small well-trained model to large model. We initialize a larger target model from a smaller source model by copy weight values from source model and padding with zeros or small initialization values on it to make the source and target model have approximate outputs, which is valid due to block matrix multiplication and residual connection in transformer structure. We test the target model on several data sets and find it is still comparable with the source model. When we continue training the target model, the training loss can start from a smaller value.

Deep representation learning has become one of the most widely adopted approaches for visual search, recommendation, and identification. Retrieval of such representations from a large database is however computationally challenging. Approximate methods based on learning compact representations, have been widely explored for this problem, such as locality sensitive hashing, product quantization, and PCA. In this work, in contrast to learning compact representations, we propose to learn high dimensional and sparse representations that have similar representational capacity as dense embeddings while being more efficient due to sparse matrix multiplication operations which can be much faster than dense multiplication. Following the key insight that the number of operations decreases quadratically with the sparsity of embeddings provided the non-zero entries are distributed uniformly across dimensions, we propose a novel approach to learn such distributed sparse embeddings via the use of a carefully constructed regularization function that directly minimizes a continuous relaxation of the number of floating-point operations (FLOPs) incurred during retrieval. Our experiments show that our approach is competitive to the other baselines and yields a similar or better speed-vs-accuracy tradeoff on practical datasets.

An efficient attention implementation is essential for large models due to its quadratic time complexity. Fortunately, attention commonly exhibits sparsity, i.e., many values in the attention map are near zero, allowing for the omission of corresponding computations. Many studies have utilized the sparse pattern to accelerate attention. However, most existing works focus on optimizing attention within specific models by exploiting certain sparse patterns of the attention map. A universal sparse attention that guarantees both the speedup and end-to-end performance of diverse models remains elusive. In this paper, we propose SpargeAttn, a universal sparse and quantized attention for any model. Our method uses a two-stage online filter: in the first stage, we rapidly and accurately predict the attention map, enabling the skip of some matrix multiplications in attention. In the second stage, we design an online softmax-aware filter that incurs no extra overhead and further skips some matrix multiplications. Experiments show that our method significantly accelerates diverse models, including language, image, and video generation, without sacrificing end-to-end metrics. The codes are available at https://github.com/thu-ml/SpargeAttn.

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

Modern automatic speech recognition (ASR) models, such as OpenAI's Whisper, rely on deep encoder-decoder architectures, and their encoders are a critical bottleneck for efficient deployment due to high computational intensity. We introduce LiteASR, a low-rank compression scheme for ASR encoders that significantly reduces inference costs while maintaining transcription accuracy. Our approach leverages the strong low-rank properties observed in intermediate activations: by applying principal component analysis (PCA) with a small calibration dataset, we approximate linear transformations with a chain of low-rank matrix multiplications, and further optimize self-attention to work in the reduced dimension. Evaluation results show that our method can compress Whisper large-v3's encoder size by over 50%, matching Whisper medium's size with better transcription accuracy, thereby establishing a new Pareto-optimal frontier of efficiency and performance. The code of LiteASR is available at https://github.com/efeslab/LiteASR.

Orthogonal finetuning (OFT) offers highly parameter-efficient adaptation while preventing catastrophic forgetting, but its high runtime and memory demands limit practical deployment. We identify the core computational bottleneck in OFT as its weight-centric implementation, which relies on costly matrix-matrix multiplications with cubic complexity. To overcome this, we propose OFTv2, an input-centric reformulation that instead uses matrix-vector multiplications (i.e., matrix-free computation), reducing the computational cost to quadratic. We further introduce the Cayley-Neumann parameterization, an efficient orthogonal parameterization that approximates the matrix inversion in Cayley transform via a truncated Neumann series. These modifications allow OFTv2 to achieve up to 10x faster training and 3x lower GPU memory usage without compromising performance. In addition, we extend OFTv2 to support finetuning quantized foundation models and show that it outperforms the popular QLoRA in training stability, efficiency, and memory usage.

Transformers with linear attention allow for efficient parallel training but can simultaneously be formulated as an RNN with 2D (matrix-valued) hidden states, thus enjoying linear (with respect to output length) inference complexity. Recent works such as RetNet (Sun et al., 2023) and TransNormerLLM (Qin et al., 2023a) observe that adding a global decay term to the additive RNN update rule greatly improves performance, sometimes outperforming standard Transformers with softmax attention when trained at scale. In this work we show that adding a data-dependent gating mechanism further improves performance. We derive a parallel form of this gated linear attention layer that enables efficient training. However, a straightforward, numerically stable implementation of this parallel form requires generalized matrix multiplications in log-space for numerical stability, and thus cannot take advantage of tensor cores on modern GPUs which are optimized for standard matrix multiplications. We develop a hardware-efficient version of the parallel form that can still make use of tensor cores through block-parallel computations over sequence chunks. Experiments on moderate-scale language modeling (340M-parameter models trained on 15B tokens, 1.3B-parameter models trained on 100B tokens) show that gated linear attention (GLA) Transformers perform competitively against a strong LLaMA-architecture Transformer baseline (Touvron et al., 2023) as well as Mamba (Gu & Dao, 2023), a recently introduced state-space model with a data-dependent state transition mechanism. For training speed, our Triton-based implementation performs comparably to CUDA-optimized FlashAttention-2 (Dao, 2023) under the regular 2048 training length setting, while outperforming FlashAttention-2 when training on longer sequences beyond 4096.

Large Language Models (LLMs) are increasingly used in applications requiring long context lengths, but the key-value (KV) cache often becomes a memory bottleneck on GPUs as context grows. To address this, we propose Commutative Vector Quantization (CommVQ) to significantly reduce memory usage for long-context LLM inference. We first introduce additive quantization with a lightweight encoder and codebook to compress the KV cache, which can be decoded via simple matrix multiplication. To further reduce computational costs during decoding, we design the codebook to be commutative with Rotary Position Embedding (RoPE) and train it using an Expectation-Maximization (EM) algorithm. This enables efficient integration of decoding into the self-attention mechanism. Our approach achieves high accuracy with additive quantization and low overhead via the RoPE-commutative codebook. Experiments on long-context benchmarks and GSM8K show that our method reduces FP16 KV cache size by 87.5% with 2-bit quantization, while outperforming state-of-the-art KV cache quantization methods. Notably, it enables 1-bit KV cache quantization with minimal accuracy loss, allowing a LLaMA-3.1 8B model to run with a 128K context length on a single RTX 4090 GPU. The source code is available at: https://github.com/UMass-Embodied-AGI/CommVQ.

Text-attributed graphs are widely used across domains, offering rich opportunities for zero-shot learning via graph-text alignment. However, existing methods struggle with tasks requiring fine-grained pattern recognition, particularly on heterophilic graphs. Through empirical and theoretical analysis, we identify an over-abstraction problem: current approaches operate at excessively large hyperbolic radii, compressing multi-scale structural information into uniform high-level abstractions. This abstraction-induced information loss obscures critical local patterns essential for accurate predictions. By analyzing embeddings in hyperbolic space, we demonstrate that optimal graph learning requires faithful preservation of fine-grained structural details, better retained by representations positioned closer to the origin. To address this, we propose H4G, a framework that systematically reduces embedding radii using learnable block-diagonal scaling matrices and M\"obius matrix multiplication. This approach restores access to fine-grained patterns while maintaining global receptive ability with minimal computational overhead. Experiments show H4G achieves state-of-the-art zero-shot performance with 12.8\% improvement on heterophilic graphs and 8.4\% on homophilic graphs, confirming that radius reduction enables faithful multi-scale representation for advancing zero-shot graph learning.

Mamba is an efficient sequence model that rivals Transformers and demonstrates significant potential as a foundational architecture for various tasks. Quantization is commonly used in neural networks to reduce model size and computational latency. However, applying quantization to Mamba remains underexplored, and existing quantization methods, which have been effective for CNN and Transformer models, appear inadequate for Mamba models (e.g., Quarot suffers a 21% accuracy drop on Vim-T^dagger even under W8A8). We have pioneered the exploration of this issue and identified several key challenges. First, significant outliers are present in gate projections, output projections, and matrix multiplications. Second, Mamba's unique parallel scan further amplifies these outliers, leading to uneven and heavy-tailed data distributions. Third, even with the application of the Hadamard transform, the variance across channels in weights and activations still remains inconsistent. To these ends, we propose MambaQuant, a post-training quantization (PTQ) framework consisting of: 1) Karhunen-Loeve Transformation (KLT) enhanced rotation, rendering the rotation matrix adaptable to diverse channel distributions. 2) Smooth-Fused rotation, which equalizes channel variances and can merge additional parameters into model weights. Experiments show that MambaQuant can quantize both weights and activations into 8-bit with less than 1% accuracy loss for Mamba-based vision and language tasks. To the best of our knowledge, MambaQuant is the first comprehensive PTQ design for the Mamba family, paving the way for further advancements in its application.

Vision transformers have emerged as a promising alternative to convolutional neural networks for various image analysis tasks, offering comparable or superior performance. However, one significant drawback of ViTs is their resource-intensive nature, leading to increased memory footprint, computation complexity, and power consumption. To democratize this high-performance technology and make it more environmentally friendly, it is essential to compress ViT models, reducing their resource requirements while maintaining high performance. In this paper, we introduce a new block-structured pruning to address the resource-intensive issue for ViTs, offering a balanced trade-off between accuracy and hardware acceleration. Unlike unstructured pruning or channel-wise structured pruning, block pruning leverages the block-wise structure of linear layers, resulting in more efficient matrix multiplications. To optimize this pruning scheme, our paper proposes a novel hardware-aware learning objective that simultaneously maximizes speedup and minimizes power consumption during inference, tailored to the block sparsity structure. This objective eliminates the need for empirical look-up tables and focuses solely on reducing parametrized layer connections. Moreover, our paper provides a lightweight algorithm to achieve post-training pruning for ViTs, utilizing second-order Taylor approximation and empirical optimization to solve the proposed hardware-aware objective. Extensive experiments on ImageNet are conducted across various ViT architectures, including DeiT-B and DeiT-S, demonstrating competitive performance with other pruning methods and achieving a remarkable balance between accuracy preservation and power savings. Especially, we achieve up to 3.93x and 1.79x speedups on dedicated hardware and GPUs respectively for DeiT-B, and also observe an inference power reduction by 1.4x on real-world GPUs.

Current high-performance semantic segmentation models are purely data-driven sub-symbolic approaches and blind to the structured nature of the visual world. This is in stark contrast to human cognition which abstracts visual perceptions at multiple levels and conducts symbolic reasoning with such structured abstraction. To fill these fundamental gaps, we devise LOGICSEG, a holistic visual semantic parser that integrates neural inductive learning and logic reasoning with both rich data and symbolic knowledge. In particular, the semantic concepts of interest are structured as a hierarchy, from which a set of constraints are derived for describing the symbolic relations and formalized as first-order logic rules. After fuzzy logic-based continuous relaxation, logical formulae are grounded onto data and neural computational graphs, hence enabling logic-induced network training. During inference, logical constraints are packaged into an iterative process and injected into the network in a form of several matrix multiplications, so as to achieve hierarchy-coherent prediction with logic reasoning. These designs together make LOGICSEG a general and compact neural-logic machine that is readily integrated into existing segmentation models. Extensive experiments over four datasets with various segmentation models and backbones verify the effectiveness and generality of LOGICSEG. We believe this study opens a new avenue for visual semantic parsing.

As the core building block of vision transformers, attention is a powerful tool to capture long-range dependency. However, such power comes at a cost: it incurs a huge computation burden and heavy memory footprint as pairwise token interaction across all spatial locations is computed. A series of works attempt to alleviate this problem by introducing handcrafted and content-agnostic sparsity into attention, such as restricting the attention operation to be inside local windows, axial stripes, or dilated windows. In contrast to these approaches, we propose a novel dynamic sparse attention via bi-level routing to enable a more flexible allocation of computations with content awareness. Specifically, for a query, irrelevant key-value pairs are first filtered out at a coarse region level, and then fine-grained token-to-token attention is applied in the union of remaining candidate regions (\ie, routed regions). We provide a simple yet effective implementation of the proposed bi-level routing attention, which utilizes the sparsity to save both computation and memory while involving only GPU-friendly dense matrix multiplications. Built with the proposed bi-level routing attention, a new general vision transformer, named BiFormer, is then presented. As BiFormer attends to a small subset of relevant tokens in a query adaptive manner without distraction from other irrelevant ones, it enjoys both good performance and high computational efficiency, especially in dense prediction tasks. Empirical results across several computer vision tasks such as image classification, object detection, and semantic segmentation verify the effectiveness of our design. Code is available at https://github.com/rayleizhu/BiFormer.

We consider learning representations of entities and relations in KBs using the neural-embedding approach. We show that most existing models, including NTN (Socher et al., 2013) and TransE (Bordes et al., 2013b), can be generalized under a unified learning framework, where entities are low-dimensional vectors learned from a neural network and relations are bilinear and/or linear mapping functions. Under this framework, we compare a variety of embedding models on the link prediction task. We show that a simple bilinear formulation achieves new state-of-the-art results for the task (achieving a top-10 accuracy of 73.2% vs. 54.7% by TransE on Freebase). Furthermore, we introduce a novel approach that utilizes the learned relation embeddings to mine logical rules such as "BornInCity(a,b) and CityInCountry(b,c) => Nationality(a,c)". We find that embeddings learned from the bilinear objective are particularly good at capturing relational semantics and that the composition of relations is characterized by matrix multiplication. More interestingly, we demonstrate that our embedding-based rule extraction approach successfully outperforms a state-of-the-art confidence-based rule mining approach in mining Horn rules that involve compositional reasoning.

Parameter-efficient finetuning (PEFT) methods effectively adapt large language models (LLMs) to diverse downstream tasks, reducing storage and GPU memory demands. Despite these advantages, several applications pose new challenges to PEFT beyond mere parameter efficiency. One notable challenge involves the efficient deployment of LLMs equipped with multiple task- or user-specific adapters, particularly when different adapters are needed for distinct requests within the same batch. Another challenge is the interpretability of LLMs, which is crucial for understanding how LLMs function. Previous studies introduced various approaches to address different challenges. In this paper, we introduce a novel method, RoAd, which employs a straightforward 2D rotation to adapt LLMs and addresses all the above challenges: (1) RoAd is remarkably parameter-efficient, delivering optimal performance on GLUE, eight commonsense reasoning tasks and four arithmetic reasoning tasks with <0.1% trainable parameters; (2) RoAd facilitates the efficient serving of requests requiring different adapters within a batch, with an overhead comparable to element-wise multiplication instead of batch matrix multiplication; (3) RoAd enhances LLM's interpretability through integration within a framework of distributed interchange intervention, demonstrated via composition experiments.

Target similarity tuning (TST) is a method of selecting relevant examples in natural language (NL) to code generation through large language models (LLMs) to improve performance. Its goal is to adapt a sentence embedding model to have the similarity between two NL inputs match the similarity between their associated code outputs. In this paper, we propose different methods to apply and improve TST in the real world. First, we replace the sentence transformer with embeddings from a larger model, which reduces sensitivity to the language distribution and thus provides more flexibility in synthetic generation of examples, and we train a tiny model that transforms these embeddings to a space where embedding similarity matches code similarity, which allows the model to remain a black box and only requires a few matrix multiplications at inference time. Second, we show how to efficiently select a smaller number of training examples to train the TST model. Third, we introduce a ranking-based evaluation for TST that does not require end-to-end code generation experiments, which can be expensive to perform.

Motivated by the problem of fast processing of attention matrices, we study fast algorithms for computing matrix-vector products for asymmetric Gaussian Kernel matrices Kin R^{ntimes n}. K's columns are indexed by a set of n keys k_1,k_2ldots, k_nin R^d, rows by a set of n queries q_1,q_2,ldots,q_nin R^d , and its i,j entry is K_{ij} = e^{-|q_i-k_j|_2^2/2sigma^2} for some bandwidth parameter sigma>0. Given a vector xin R^n and error parameter epsilon>0, our task is to output a yin R^n such that |Kx-y|_2leq epsilon |x|_2 in time subquadratic in n and linear in d. Our algorithms rely on the following modelling assumption about the matrices K: the sum of the entries of K scales linearly in n, as opposed to worst case quadratic growth. We validate this assumption experimentally, for Gaussian kernel matrices encountered in various settings such as fast attention computation in LLMs. We obtain the first subquadratic-time algorithm that works under this assumption, for unrestricted vectors.

The training process of neural networks usually optimize weights and bias parameters of linear transformations, while nonlinear activation functions are pre-specified and fixed. This work develops a systematic approach to constructing matrix activation functions whose entries are generalized from ReLU. The activation is based on matrix-vector multiplications using only scalar multiplications and comparisons. The proposed activation functions depend on parameters that are trained along with the weights and bias vectors. Neural networks based on this approach are simple and efficient and are shown to be robust in numerical experiments.

Large Language Models (LLMs) have demonstrated remarkable capabilities but typically require extensive computational resources and memory for inference. Post-training quantization (PTQ) can effectively reduce these demands by storing weights in lower bit-width formats. However, standard uniform quantization often leads to notable performance degradation, particularly in low-bit scenarios. In this work, we introduce a Grouped Lattice Vector Quantization (GLVQ) framework that assigns each group of weights a customized lattice codebook, defined by a learnable generation matrix. To address the non-differentiability of the quantization process, we adopt Babai rounding to approximate nearest-lattice-point search during training, which enables stable optimization of the generation matrices. Once trained, decoding reduces to a simple matrix-vector multiplication, yielding an efficient and practical quantization pipeline. Experiments on multiple benchmarks show that our approach achieves a better trade-off between model size and accuracy compared to existing post-training quantization baselines, highlighting its effectiveness in deploying large models under stringent resource constraints. Our source code is available on GitHub repository: https://github.com/xzhang9308/GLVQ.

The energy consumption of large-scale ML models is dominated by data movement - shuffling billions of parameters across memory hierarchies and data centers. Effective sparsification to prune redundant parameters is still challenging: existing methods incur significant accuracy degradation, performance overhead, or both. We introduce (Bl)ock (a)nd (S)parse (T)ransformers (BLaST), a general, robust, and reliable sparsification method applicable to linear layers in all settings. Our method iteratively sparsifies weight matrices into a block sparsity pattern suitable for efficient sparse matrix-matrix (SpMM) multiplication. BLaST achieves up to 95% sparsity in MLP weights with negligible accuracy loss. Our fused, highly optimized Sparse MLP kernel delivers up to 16.7x speedup over dense MLPs across 9 architectures and 8 datasets, resulting in up to 1.6x inference speedup, 1.11x pretraining speedup and up to 3.12x inference memory usage reduction. BLaST enables the next generation of large-scale AI systems by reducing energy use, memory footprint, and latency.