Dan Alistarh

Elias Frantar, Saleh Ashkboos, Torsten Hoefler et al.

Generative Pre-trained Transformer models, known as GPT or OPT, set themselves apart through breakthrough performance across complex language modelling tasks, but also by their extremely high computational and storage costs. Specifically, due to their massive size, even inference for large, highly-accurate GPT models may require multiple performant GPUs, which limits the usability of such models. While there is emerging work on relieving this pressure via model compression, the applicability and performance of existing compression techniques is limited by the scale and complexity of GPT models. In this paper, we address this challenge, and propose GPTQ, a new one-shot weight quantization method based on approximate second-order information, that is both highly-accurate and highly-efficient. Specifically, GPTQ can quantize GPT models with 175 billion parameters in approximately four GPU hours, reducing the bitwidth down to 3 or 4 bits per weight, with negligible accuracy degradation relative to the uncompressed baseline. Our method more than doubles the compression gains relative to previously-proposed one-shot quantization methods, preserving accuracy, allowing us for the first time to execute an 175 billion-parameter model inside a single GPU for generative inference. Moreover, we also show that our method can still provide reasonable accuracy in the extreme quantization regime, in which weights are quantized to 2-bit or even ternary quantization levels. We show experimentally that these improvements can be leveraged for end-to-end inference speedups over FP16, of around 3.25x when using high-end GPUs (NVIDIA A100) and 4.5x when using more cost-effective ones (NVIDIA A6000). The implementation is available at https://github.com/IST-DASLab/gptq.

Elias Frantar, Dan Alistarh

We show for the first time that large-scale generative pretrained transformer (GPT) family models can be pruned to at least 50% sparsity in one-shot, without any retraining, at minimal loss of accuracy. This is achieved via a new pruning method called SparseGPT, specifically designed to work efficiently and accurately on massive GPT-family models. We can execute SparseGPT on the largest available open-source models, OPT-175B and BLOOM-176B, in under 4.5 hours, and can reach 60% unstructured sparsity with negligible increase in perplexity: remarkably, more than 100 billion weights from these models can be ignored at inference time. SparseGPT generalizes to semi-structured (2:4 and 4:8) patterns, and is compatible with weight quantization approaches. The code is available at: https://github.com/IST-DASLab/sparsegpt.

Eldar Kurtic, Daniel Campos, Tuan Nguyen et al.

Transformer-based language models have become a key building block for natural language processing. While these models are extremely accurate, they can be too large and computationally intensive to run on standard deployments. A variety of compression methods, including distillation, quantization, structured and unstructured pruning are known to decrease model size and increase inference speed, with low accuracy loss. In this context, this paper's contributions are two-fold. We perform an in-depth study of the accuracy-compression trade-off for unstructured weight pruning of BERT models. We introduce Optimal BERT Surgeon (oBERT), an efficient and accurate weight pruning method based on approximate second-order information, which we show to yield state-of-the-art results in both stages of language tasks: pre-training and fine-tuning. Specifically, oBERT extends existing work on unstructured second-order pruning by allowing for pruning blocks of weights, and by being applicable at the BERT scale. Second, we investigate the impact of this pruning method when compounding compression approaches to obtain highly compressed but accurate models for deployment on edge devices. These models significantly push boundaries of the current state-of-the-art sparse BERT models with respect to all metrics: model size, inference speed and task accuracy. For example, relative to the dense BERT-base, we obtain 10x model size compression (in MB) with < 1% accuracy drop, 10x CPU-inference speedup with < 2% accuracy drop, and 29x CPU-inference speedup with < 7.5% accuracy drop. Our code, fully integrated with Transformers and SparseML, is available at https://github.com/neuralmagic/sparseml/tree/main/research/optimal_BERT_surgeon_oBERT.

Eldar Kurtic, Elias Frantar, Dan Alistarh

The breakthrough performance of large language models (LLMs) comes with major computational footprints and high deployment costs. In this paper, we progress towards resolving this problem by proposing a novel structured compression approach for LLMs, called ZipLM. ZipLM achieves state-of-the-art accuracy-vs-speedup, while matching a set of desired target runtime speedups in any given inference environment. Specifically, given a model, a dataset, an inference environment, as well as a set of speedup targets, ZipLM iteratively identifies and removes components with the worst loss-runtime trade-off. Unlike prior methods that specialize in either the post-training/one-shot or the gradual compression setting, and only for specific families of models such as BERT (encoder) or GPT (decoder), ZipLM produces state-of-the-art compressed models across all these settings. Furthermore, ZipLM achieves superior results for a fraction of the computational cost relative to prior distillation and pruning techniques, making it a cost-effective approach for generating an entire family of smaller, faster, and highly accurate models, guaranteed to meet the desired inference specifications. In particular, ZipLM outperforms all prior BERT-base distillation and pruning techniques, such as CoFi, MiniLM, and TinyBERT. Moreover, it matches the performance of the heavily optimized MobileBERT model, obtained via extensive architecture search, by simply pruning the baseline BERT-large model. When compressing GPT2, ZipLM outperforms DistilGPT2 while being 60% smaller and 30% faster. Our code is available at: https://github.com/IST-DASLab/ZipLM.

Elias Frantar, Dan Alistarh

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

Saleh Ashkboos, Ilia Markov, Elias Frantar et al.

Large Language Models (LLMs) from the GPT family have become extremely popular, leading to a race towards reducing their inference costs to allow for efficient local computation. Yet, the vast majority of existing work focuses on weight-only quantization, which can reduce runtime costs in the memory-bound one-token-at-a-time generative setting, but does not address them in compute-bound scenarios, such as batched inference or prompt processing. In this paper, we address the general quantization problem, where both weights and activations should be quantized. We show, for the first time, that the majority of inference computations for large generative models such as LLaMA, OPT, and Falcon can be performed with both weights and activations being cast to 4 bits, in a way that leads to practical speedups, while at the same time maintaining good accuracy. We achieve this via a hybrid quantization strategy called QUIK, which compresses most of the weights and activations to 4-bit, while keeping some outlier weights and activations in higher-precision. The key feature of our scheme is that it is designed with computational efficiency in mind: we provide GPU kernels matching the QUIK format with highly-efficient layer-wise runtimes, which lead to practical end-to-end throughput improvements of up to 3.4x relative to FP16 execution. We provide detailed studies for models from the OPT, LLaMA-2 and Falcon families, as well as a first instance of accurate inference using quantization plus 2:4 sparsity. Code is available at: https://github.com/IST-DASLab/QUIK.

Alexandra Peste, Adrian Vladu, Eldar Kurtic et al.

Deep neural networks (DNNs) often have to be compressed, via pruning and/or quantization, before they can be deployed in practical settings. In this work we propose a new compression-aware minimizer dubbed CrAM that modifies the optimization step in a principled way, in order to produce models whose local loss behavior is stable under compression operations such as pruning. Thus, dense models trained via CrAM should be compressible post-training, in a single step, without significant accuracy loss. Experimental results on standard benchmarks, such as residual networks for ImageNet classification and BERT models for language modelling, show that CrAM produces dense models that can be more accurate than the standard SGD/Adam-based baselines, but which are stable under weight pruning: specifically, we can prune models in one-shot to 70-80% sparsity with almost no accuracy loss, and to 90% with reasonable ($\sim 1\%$) accuracy loss, which is competitive with gradual compression methods. Additionally, CrAM can produce sparse models which perform well for transfer learning, and it also works for semi-structured 2:4 pruning patterns supported by GPU hardware. The code for reproducing the results is available at https://github.com/IST-DASLab/CrAM .

Tim Dettmers, Ruslan Svirschevski, Vage Egiazarian et al.

Recent advances in large language model (LLM) pretraining have led to high-quality LLMs with impressive abilities. By compressing such LLMs via quantization to 3-4 bits per parameter, they can fit into memory-limited devices such as laptops and mobile phones, enabling personalized use. However, quantization down to 3-4 bits per parameter usually leads to moderate-to-high accuracy losses, especially for smaller models in the 1-10B parameter range, which are well-suited for edge deployments. To address this accuracy issue, we introduce the Sparse-Quantized Representation (SpQR), a new compressed format and quantization technique which enables for the first time near-lossless compression of LLMs across model scales, while reaching similar compression levels to previous methods. SpQR works by identifying and isolating outlier weights, which cause particularly-large quantization errors, and storing them in higher precision, while compressing all other weights to 3-4 bits, and achieves relative accuracy losses of less than 1% in perplexity for highly-accurate LLaMA and Falcon LLMs. This makes it possible to run 33B parameter LLM on a single 24 GB consumer GPU without any performance degradation at 15% speedup thus making powerful LLMs available to consumer without any downsides. SpQR comes with efficient algorithms for both encoding weights into its format, as well as decoding them efficiently at runtime. Specifically, we provide an efficient GPU inference algorithm for SpQR which yields faster inference than 16-bit baselines at similar accuracy, while enabling memory compression gains of more than 4x.

Tommaso Pegolotti, Elias Frantar, Dan Alistarh et al.

We present ongoing work on a new automatic code generation approach for supporting quantized generative inference on LLMs such as LLaMA or OPT on off-the-shelf CPUs. Our approach is informed by the target architecture and a performance model, including both hardware characteristics and method-specific accuracy constraints. Results on CPU-based inference for LLaMA models show that our approach can lead to high performance and high accuracy, comparing favorably to the best existing open-source solution. A preliminary implementation is available at https://github.com/IST-DASLab/QIGen.

Jiale Chen, Dingling Yao, Adeel Pervez et al.

We propose Scalable Mechanistic Neural Network (S-MNN), an enhanced neural network framework designed for scientific machine learning applications involving long temporal sequences. By reformulating the original Mechanistic Neural Network (MNN) (Pervez et al., 2024), we reduce the computational time and space complexities from cubic and quadratic with respect to the sequence length, respectively, to linear. This significant improvement enables efficient modeling of long-term dynamics without sacrificing accuracy or interpretability. Extensive experiments demonstrate that S-MNN matches the original MNN in precision while substantially reducing computational resources. Consequently, S-MNN can drop-in replace the original MNN in applications, providing a practical and efficient tool for integrating mechanistic bottlenecks into neural network models of complex dynamical systems. Source code is available at https://github.com/IST-DASLab/ScalableMNN.

Denis Kuznedelev, Soroush Tabesh, Kimia Noorbakhsh et al. · mit

Recent vision architectures and self-supervised training methods enable vision models that are extremely accurate and general, but come with massive parameter and computational costs. In practical settings, such as camera traps, users have limited resources, and may fine-tune a pretrained model on (often limited) data from a small set of specific categories of interest. These users may wish to make use of modern, highly-accurate models, but are often computationally constrained. To address this, we ask: can we quickly compress large generalist models into accurate and efficient specialists? For this, we propose a simple and versatile technique called Few-Shot Task-Aware Compression (TACO). Given a large vision model that is pretrained to be accurate on a broad task, such as classification over ImageNet-22K, TACO produces a smaller model that is accurate on specialized tasks, such as classification across vehicle types or animal species. Crucially, TACO works in few-shot fashion, i.e. only a few task-specific samples are used, and the procedure has low computational overheads. We validate TACO on highly-accurate ResNet, ViT/DeiT, and ConvNeXt models, originally trained on ImageNet, LAION, or iNaturalist, which we specialize and compress to a diverse set of "downstream" subtasks. TACO can reduce the number of non-zero parameters in existing models by up to 20x relative to the original models, leading to inference speedups of up to 3$\times$, while remaining accuracy-competitive with the uncompressed models on the specialized tasks.

Bapi Chatterjee, Vyacheslav Kungurtsev, Dan Alistarh

Powered by the simplicity of lock-free asynchrony, Hogwilld! is a go-to approach to parallelize SGD over a shared-memory setting. Despite its popularity and concomitant extensions, such as PASSM+ wherein concurrent processes update a shared model with partitioned gradients, scaling it to decentralized workers has surprisingly been relatively unexplored. To our knowledge, there is no convergence theory of such methods, nor systematic numerical comparisons evaluating speed-up. In this paper, we propose an algorithm incorporating decentralized distributed memory computing architecture with each node running multiprocessing parallel shared-memory SGD itself. Our scheme is based on the following algorithmic tools and features: (a) asynchronous local gradient updates on the shared-memory of workers, (b) partial backpropagation, and (c) non-blocking in-place averaging of the local models. We prove that our method guarantees ergodic convergence rates for non-convex objectives. On the practical side, we show that the proposed method exhibits improved throughput and competitive accuracy for standard image classification benchmarks on the CIFAR-10, CIFAR-100, and Imagenet datasets. Our code is available at https://github.com/bapi/LPP-SGD.

Hossein Zakerinia, Shayan Talaei, Giorgi Nadiradze et al. · eth-zurich

Federated Learning (FL) enables large-scale distributed training of machine learning models, while still allowing individual nodes to maintain data locally. However, executing FL at scale comes with inherent practical challenges: 1) heterogeneity of the local node data distributions, 2) heterogeneity of node computational speeds (asynchrony), but also 3) constraints in the amount of communication between the clients and the server. In this work, we present the first variant of the classic federated averaging (FedAvg) algorithm which, at the same time, supports data heterogeneity, partial client asynchrony, and communication compression. Our algorithm comes with a novel, rigorous analysis showing that, in spite of these system relaxations, it can provide similar convergence to FedAvg in interesting parameter regimes. Experimental results in the rigorous LEAF benchmark on setups of up to 300 nodes show that our algorithm ensures fast convergence for standard federated tasks, improving upon prior quantized and asynchronous approaches.

Elias Frantar, Carlos Riquelme, Neil Houlsby et al.

We explore the impact of parameter sparsity on the scaling behavior of Transformers trained on massive datasets (i.e., "foundation models"), in both vision and language domains. In this setting, we identify the first scaling law describing the relationship between weight sparsity, number of non-zero parameters, and amount of training data, which we validate empirically across model and data scales; on ViT/JFT-4B and T5/C4. These results allow us to characterize the "optimal sparsity", the sparsity level which yields the best performance for a given effective model size and training budget. For a fixed number of non-zero parameters, we identify that the optimal sparsity increases with the amount of data used for training. We also extend our study to different sparsity structures (such as the hardware-friendly n:m pattern) and strategies (such as starting from a pretrained dense model). Our findings shed light on the power and limitations of weight sparsity across various parameter and computational settings, offering both theoretical understanding and practical implications for leveraging sparsity towards computational efficiency improvements.

Elias Frantar, Roberto L. Castro, Jiale Chen et al.

As inference on Large Language Models (LLMs) emerges as an important workload in machine learning applications, weight quantization has become a standard technique for efficient GPU deployment. Quantization not only reduces model size, but has also been shown to yield substantial speedups for single-user inference, due to reduced memory movement, with low accuracy impact. Yet, it remains open whether speedups are achievable also in \emph{batched} settings with multiple parallel clients, which are highly relevant for practical serving. It is unclear whether GPU kernels can be designed to remain practically memory-bound, while supporting the substantially increased compute requirements of batched workloads. This paper resolves this question positively by describing the design of Mixed-precision Auto-Regressive LINear kernels, called MARLIN. Concretely, given a model whose weights are compressed via quantization to, e.g., 4 bits per element, MARLIN shows that batchsizes up to 16-32 can be supported with close to maximum ($4\times$) quantization speedup, and larger batchsizes up to 64-128 with gradually decreasing, but still significant, acceleration. MARLIN accomplishes this via a combination of techniques, such as asynchronous memory access, complex task scheduling and pipelining, and bespoke quantization support. Our experiments show that MARLIN's near-optimal performance on individual LLM layers across different scenarios can also lead to end-to-end LLM inference speedups (of up to $2.8\times$) when integrated with the popular vLLM serving engine. Finally, MARLIN is extensible to further compression techniques, like NVIDIA 2:4 sparsity, leading to additional speedups.

Matin Ansaripour, Shayan Talaei, Giorgi Nadiradze et al. · eth-zurich

Distributed optimization is the standard way of speeding up machine learning training, and most of the research in the area focuses on distributed first-order, gradient-based methods. Yet, there are settings where some computationally-bounded nodes may not be able to implement first-order, gradient-based optimization, while they could still contribute to joint optimization tasks. In this paper, we initiate the study of hybrid decentralized optimization, studying settings where nodes with zeroth-order and first-order optimization capabilities co-exist in a distributed system, and attempt to jointly solve an optimization task over some data distribution. We essentially show that, under reasonable parameter settings, such a system can not only withstand noisier zeroth-order agents but can even benefit from integrating such agents into the optimization process, rather than ignoring their information. At the core of our approach is a new analysis of distributed optimization with noisy and possibly-biased gradient estimators, which may be of independent interest. Our results hold for both convex and non-convex objectives. Experimental results on standard optimization tasks confirm our analysis, showing that hybrid first-zeroth order optimization can be practical, even when training deep neural networks.

Rustem Islamov, Mher Safaryan, Dan Alistarh

We analyze asynchronous-type algorithms for distributed SGD in the heterogeneous setting, where each worker has its own computation and communication speeds, as well as data distribution. In these algorithms, workers compute possibly stale and stochastic gradients associated with their local data at some iteration back in history and then return those gradients to the server without synchronizing with other workers. We present a unified convergence theory for non-convex smooth functions in the heterogeneous regime. The proposed analysis provides convergence for pure asynchronous SGD and its various modifications. Moreover, our theory explains what affects the convergence rate and what can be done to improve the performance of asynchronous algorithms. In particular, we introduce a novel asynchronous method based on worker shuffling. As a by-product of our analysis, we also demonstrate convergence guarantees for gradient-type algorithms such as SGD with random reshuffling and shuffle-once mini-batch SGD. The derived rates match the best-known results for those algorithms, highlighting the tightness of our approach. Finally, our numerical evaluations support theoretical findings and show the good practical performance of our method.

Ilia Markov, Adrian Vladu, Qi Guo et al.

Communication-reduction techniques are a popular way to improve scalability in data-parallel training of deep neural networks (DNNs). The recent emergence of large language models such as GPT has created the need for new approaches to exploit data-parallelism. Among these, fully-sharded data parallel (FSDP) training is highly popular, yet it still encounters scalability bottlenecks. One reason is that applying compression techniques to FSDP is challenging: as the vast majority of the communication involves the model's weights, direct compression alters convergence and leads to accuracy loss. We present QSDP, a variant of FSDP which supports both gradient and weight quantization with theoretical guarantees, is simple to implement and has essentially no overheads. To derive QSDP we prove that a natural modification of SGD achieves convergence even when we only maintain quantized weights, and thus the domain over which we train consists of quantized points and is, therefore, highly non-convex. We validate this approach by training GPT-family models with up to 1.3 billion parameters on a multi-node cluster. Experiments show that QSDP preserves model accuracy, while completely removing the communication bottlenecks of FSDP, providing end-to-end speedups of up to 2.2x.

Denis Kuznedelev, Eldar Kurtic, Elias Frantar et al.

Driven by significant improvements in architectural design and training pipelines, computer vision has recently experienced dramatic progress in terms of accuracy on classic benchmarks such as ImageNet. These highly-accurate models are challenging to deploy, as they appear harder to compress using standard techniques such as pruning. We address this issue by introducing the Correlation Aware Pruner (CAP), a new unstructured pruning framework which significantly pushes the compressibility limits for state-of-the-art architectures. Our method is based on two technical advancements: a new theoretically-justified pruner, which can handle complex weight correlations accurately and efficiently during the pruning process itself, and an efficient finetuning procedure for post-compression recovery. We validate our approach via extensive experiments on several modern vision models such as Vision Transformers (ViT), modern CNNs, and ViT-CNN hybrids, showing for the first time that these can be pruned to high sparsity levels (e.g. $\geq 75$%) with low impact on accuracy ($\leq 1$% relative drop). Our approach is also compatible with structured pruning and quantization, and can lead to practical speedups of 1.5 to 2.4x without accuracy loss. To further showcase CAP's accuracy and scalability, we use it to show for the first time that extremely-accurate large vision models, trained via self-supervised techniques, can also be pruned to moderate sparsities, with negligible accuracy loss.

Eldar Kurtic, Denis Kuznedelev, Elias Frantar et al.

We consider the problem of accurate sparse fine-tuning of large language models (LLMs), that is, fine-tuning pretrained LLMs on specialized tasks, while inducing sparsity in their weights. On the accuracy side, we observe that standard loss-based fine-tuning may fail to recover accuracy, especially at high sparsities. To address this, we perform a detailed study of distillation-type losses, determining an L2-based distillation approach we term SquareHead which enables accurate recovery even at higher sparsities, across all model types. On the practical efficiency side, we show that sparse LLMs can be executed with speedups by taking advantage of sparsity, for both CPU and GPU runtimes. While the standard approach is to leverage sparsity for computational reduction, we observe that in the case of memory-bound LLMs sparsity can also be leveraged for reducing memory bandwidth. We exhibit end-to-end results showing speedups due to sparsity, while recovering accuracy, on T5 (language translation), Whisper (speech translation), and open GPT-type (MPT for text generation). For MPT text generation, we show for the first time that sparse fine-tuning can reach 75% sparsity without accuracy drops, provide notable end-to-end speedups for both CPU and GPU inference, and highlight that sparsity is also compatible with quantization approaches. Models and software for reproducing our results are provided in Section 6.

Eldar Kurtic, Dan Alistarh

We revisit the performance of the classic gradual magnitude pruning (GMP) baseline for large language models, focusing on the classic BERT benchmark on various popular tasks. Despite existing evidence in the literature that GMP performs poorly, we show that a simple and general variant, which we call GMP*, can match and sometimes outperform more complex state-of-the-art methods. Our results provide a simple yet strong baseline for future work, highlight the importance of parameter tuning for baselines, and even improve the performance of the state-of-the-art second-order pruning method in this setting.

Mahdi Nikdan, Tommaso Pegolotti, Eugenia Iofinova et al.

We provide a new efficient version of the backpropagation algorithm, specialized to the case where the weights of the neural network being trained are sparse. Our algorithm is general, as it applies to arbitrary (unstructured) sparsity and common layer types (e.g., convolutional or linear). We provide a fast vectorized implementation on commodity CPUs, and show that it can yield speedups in end-to-end runtime experiments, both in transfer learning using already-sparsified networks, and in training sparse networks from scratch. Thus, our results provide the first support for sparse training on commodity hardware.

Denis Kuznedelev, Eldar Kurtic, Eugenia Iofinova et al.

Obtaining versions of deep neural networks that are both highly-accurate and highly-sparse is one of the main challenges in the area of model compression, and several high-performance pruning techniques have been investigated by the community. Yet, much less is known about the interaction between sparsity and the standard stochastic optimization techniques used for training sparse networks, and most existing work uses standard dense schedules and hyperparameters for training sparse networks. In this work, we examine the impact of high sparsity on model training using the standard computer vision and natural language processing sparsity benchmarks. We begin by showing that using standard dense training recipes for sparse training is suboptimal, and results in under-training. We provide new approaches for mitigating this issue for both sparse pre-training of vision models (e.g. ResNet50/ImageNet) and sparse fine-tuning of language models (e.g. BERT/GLUE), achieving state-of-the-art results in both settings in the high-sparsity regime, and providing detailed analyses for the difficulty of sparse training in both scenarios. Our work sets a new threshold in terms of the accuracies that can be achieved under high sparsity, and should inspire further research into improving sparse model training, to reach higher accuracies under high sparsity, but also to do so efficiently.

Eugenia Iofinova, Alexandra Peste, Dan Alistarh

Pruning - that is, setting a significant subset of the parameters of a neural network to zero - is one of the most popular methods of model compression. Yet, several recent works have raised the issue that pruning may induce or exacerbate bias in the output of the compressed model. Despite existing evidence for this phenomenon, the relationship between neural network pruning and induced bias is not well-understood. In this work, we systematically investigate and characterize this phenomenon in Convolutional Neural Networks for computer vision. First, we show that it is in fact possible to obtain highly-sparse models, e.g. with less than 10% remaining weights, which do not decrease in accuracy nor substantially increase in bias when compared to dense models. At the same time, we also find that, at higher sparsities, pruned models exhibit higher uncertainty in their outputs, as well as increased correlations, which we directly link to increased bias. We propose easy-to-use criteria which, based only on the uncompressed model, establish whether bias will increase with pruning, and identify the samples most susceptible to biased predictions post-compression.

Diyuan Wu, Ionut-Vlad Modoranu, Mher Safaryan et al.

The rising footprint of machine learning has led to a focus on imposing \emph{model sparsity} as a means of reducing computational and memory costs. For deep neural networks (DNNs), the state-of-the-art accuracy-vs-sparsity is achieved by heuristics inspired by the classical Optimal Brain Surgeon (OBS) framework~\citep{lecun90brain, hassibi1992second, hassibi1993optimal}, which leverages loss curvature information to make better pruning decisions. Yet, these results still lack a solid theoretical understanding, and it is unclear whether they can be improved by leveraging connections to the wealth of work on sparse recovery algorithms. In this paper, we draw new connections between these two areas and present new sparse recovery algorithms inspired by the OBS framework that comes with theoretical guarantees under reasonable assumptions and have strong practical performance. Specifically, our work starts from the observation that we can leverage curvature information in OBS-like fashion upon the projection step of classic iterative sparse recovery algorithms such as IHT. We show for the first time that this leads both to improved convergence bounds under standard assumptions. Furthermore, we present extensions of this approach to the practical task of obtaining accurate sparse DNNs, and validate it experimentally at scale for Transformer-based models on vision and language tasks.

Mohammadreza Alimohammadi, Ilia Markov, Elias Frantar et al.

Data-parallel distributed training of deep neural networks (DNN) has gained very widespread adoption, but can still experience communication bottlenecks. To address this issue, entire families of compression mechanisms have been developed, including quantization, sparsification, and low-rank approximation, some of which are seeing significant practical adoption. Despite this progress, almost all known compression schemes apply compression uniformly across DNN layers, although layers are heterogeneous in terms of parameter count and their impact on model accuracy. In this work, we provide a general framework for adapting the degree of compression across the model's layers dynamically during training, improving the overall compression, while leading to substantial speedups, without sacrificing accuracy. Our framework, called L-GreCo, is based on an adaptive algorithm, which automatically picks the optimal compression parameters for model layers guaranteeing the best compression ratio while satisfying an error constraint. Extensive experiments over image classification and language modeling tasks shows that L-GreCo is effective across all existing families of compression methods, and achieves up to 2.5$\times$ training speedup and up to 5$\times$ compression improvement over efficient implementations of existing approaches, while recovering full accuracy. Moreover, L-GreCo is complementary to existing adaptive algorithms, improving their compression ratio by 50% and practical throughput by 66%.

Vage Egiazarian, Denis Kuznedelev, Anton Voronov et al.

Text-to-image diffusion models have emerged as a powerful framework for high-quality image generation given textual prompts. Their success has driven the rapid development of production-grade diffusion models that consistently increase in size and already contain billions of parameters. As a result, state-of-the-art text-to-image models are becoming less accessible in practice, especially in resource-limited environments. Post-training quantization (PTQ) tackles this issue by compressing the pretrained model weights into lower-bit representations. Recent diffusion quantization techniques primarily rely on uniform scalar quantization, providing decent performance for the models compressed to 4 bits. This work demonstrates that more versatile vector quantization (VQ) may achieve higher compression rates for large-scale text-to-image diffusion models. Specifically, we tailor vector-based PTQ methods to recent billion-scale text-to-image models (SDXL and SDXL-Turbo), and show that the diffusion models of 2B+ parameters compressed to around 3 bits using VQ exhibit the similar image quality and textual alignment as previous 4-bit compression techniques.

Saleh Ashkboos, Amirkeivan Mohtashami, Maximilian L. Croci et al.

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 4-bit quantized LLaMa2-70B model has losses of at most 0.47 WikiText-2 perplexity and retains 99% of the zero-shot performance. We also show that QuaRot can provide lossless 6 and 8 bit LLaMa2 models without any calibration data using round-to-nearest quantization. Code is available at: https://github.com/spcl/QuaRot.

Jiale Chen, Vage Egiazarian, Torsten Hoefler et al.

Quantization to low bitwidth is a standard approach for deploying large language models, however, a few extreme weights and activations stretch the dynamic range and reduce the effective resolution of the quantizer. A common mitigation approach is to apply some fixed orthogonal transforms, such as Hadamard matrices, before quantization, which typically reduces the dynamic range. Yet, these transforms ignore the statistics of the data, and their optimality is currently not understood. In this work, we derive, for the first time, closed-form optimal linear blockwise transforms for joint weight-activation quantization using standard data-free quantizers for common numerical formats. Specifically, we provide derivations of the optimal adaptive (data-aware) transforms for round-to-nearest (RTN), AbsMax-scaled block quantizers for both integer and floating-point formats. The resulting construction, which we call WUSH, combines a Hadamard backbone with a data-dependent component based on second-order moments, yielding a non-orthogonal transform that is provably optimal under mild assumptions and remains structured for efficient implementation. Preliminary experimental results show that our approach consistently improves upon the Hadamard transform for common formats.

John Lazarsfeld, Dan Alistarh

We study a cooperative multi-agent bandit setting in the distributed GOSSIP model: in every round, each of $n$ agents chooses an action from a common set, observes the action's corresponding reward, and subsequently exchanges information with a single randomly chosen neighbor, which may inform its choice in the next round. We introduce and analyze families of memoryless and time-independent protocols for this setting, inspired by opinion dynamics that are well-studied for other algorithmic tasks in the GOSSIP model. For stationary reward settings, we prove for the first time that these simple protocols exhibit best-of-both-worlds behavior, simultaneously obtaining constant cumulative regret scaling like $R(T)/T = \widetilde O(1/T)$, and also reaching consensus on the highest-mean action within $\widetilde O(\sqrt{n})$ rounds. We obtain these results by showing a new connection between the global evolution of these decentralized protocols and a class of zero-sum multiplicative weights update} processes. Using this connection, we establish a general framework for analyzing the population-level regret and other properties of our protocols. Finally, we show our protocols are also surprisingly robust to adversarial rewards, and in this regime we obtain sublinear regret scaling like $R(T)/T = \widetilde O(1/\sqrt{T})$ as long as the number of rounds does not grow too fast as a function of $n$.

Ionut-Vlad Modoranu, Mher Safaryan, Grigory Malinovsky et al.

We propose a new variant of the Adam optimizer called MicroAdam that specifically minimizes memory overheads, while maintaining theoretical convergence guarantees. We achieve this by compressing the gradient information before it is fed into the optimizer state, thereby reducing its memory footprint significantly. We control the resulting compression error via a novel instance of the classical \emph{error feedback} mechanism from distributed optimization in which *the error correction information is itself compressed* to allow for practical memory gains. We prove that the resulting approach maintains theoretical convergence guarantees competitive to those of AMSGrad, while providing good practical performance. Specifically, we show that MicroAdam can be implemented efficiently on GPUs: on both million-scale (BERT) and billion-scale (LLaMA) models, MicroAdam provides practical convergence competitive to that of the uncompressed Adam baseline, with lower memory usage and similar running time. Our code is available at https://github.com/IST-DASLab/MicroAdam.

Thomas Robert, Mher Safaryan, Ionut-Vlad Modoranu et al.

We introduce LDAdam, a memory-efficient optimizer for training large models, that performs adaptive optimization steps within lower dimensional subspaces, while consistently exploring the full parameter space during training. This strategy keeps the optimizer's memory footprint to a fraction of the model size. LDAdam relies on a new projection-aware update rule for the optimizer states that allows for transitioning between subspaces, i.e., estimation of the statistics of the projected gradients. To mitigate the errors due to low-rank projection, LDAdam integrates a new generalized error feedback mechanism, which explicitly accounts for both gradient and optimizer state compression. We prove the convergence of LDAdam under standard assumptions, and show that LDAdam allows for accurate and efficient fine-tuning and pre-training of language models. Code is available at https://github.com/IST-DASLab/LDAdam

Andrei Panferov, Jiale Chen, Soroush Tabesh et al.

One approach to reducing the massive costs of large language models (LLMs) is the use of quantized or sparse representations for training or deployment. While post-training compression methods are very popular, the question of obtaining even more accurate compressed models by directly training over such representations, i.e., Quantization-Aware Training (QAT), is still open: for example, a recent study (arXiv:2411.04330) put the "optimal" bit-width at which models can be trained using QAT, while staying accuracy-competitive with standard FP16/BF16 precision, at 8-bits weights and activations. We advance this state-of-the-art via a new method called QuEST, for which we demonstrate optimality at 4-bits and stable convergence as low as 1-bit weights and activations. QuEST achieves this by improving two key aspects of QAT methods: (1) accurate and fast quantization of the (continuous) distributions of weights and activations via Hadamard normalization and MSE-optimal fitting; (2) a new trust gradient estimator based on the idea of explicitly minimizing the error between the noisy gradient computed over quantized states and the "true" (but unknown) full-precision gradient. Experiments on Llama-type architectures show that QuEST induces stable scaling laws across the entire range of hardware-supported precisions, and can be extended to sparse representations. We provide GPU kernel support showing that models produced by QuEST can be executed efficiently. Our code is available at https://github.com/IST-DASLab/QuEST.

Shengkun Tang, Oliver Sieberling, Eldar Kurtic et al.

Large Language Models (LLMs) have achieved significant success across various NLP tasks. However, their massive computational costs limit their widespread use, particularly in real-time applications. Structured pruning offers an effective solution by compressing models and directly providing end-to-end speed improvements, regardless of the hardware environment. Meanwhile, different components of the model exhibit varying sensitivities towards pruning, calling for non-uniform model compression. However, a pruning method should not only identify a capable substructure, but also account for post-compression training. To this end, we propose DarwinLM, a method for training-aware structured pruning. DarwinLM builds upon an evolutionary search process, generating multiple offspring models in each generation through mutation, and selecting the fittest for survival. To assess the effect of post-training, we incorporate a lightweight, multistep training process within the offspring population, progressively increasing the number of tokens and eliminating poorly performing models in each selection stage. We validate our method through extensive experiments on Llama-2-7B, Llama-3.1-8B and Qwen-2.5-14B-Instruct, achieving state-of-the-art performance for structured pruning. For instance, DarwinLM surpasses ShearedLlama while requiring 5x less training data during post-compression training. Code is at: https://github.com/IST-DASLab/DarwinLM

Oliver Sieberling, Denis Kuznedelev, Eldar Kurtic et al.

The high computational costs of large language models (LLMs) have led to a flurry of research on LLM compression, via methods such as quantization, sparsification, or structured pruning. A new frontier in this area is given by dynamic, non-uniform compression methods, which adjust the compression levels (e.g., sparsity) per-block or even per-layer in order to minimize accuracy loss, while guaranteeing a global compression threshold. Yet, current methods rely on estimating the importance of a given layer, implicitly assuming that layers contribute independently to the overall compression error. We begin from the motivating observation that this independence assumption does not generally hold for LLM compression: pruning a model further may even significantly recover performance. To address this, we propose EvoPress, a novel evolutionary framework for dynamic LLM compression. By formulating dynamic compression as a general optimization problem, EvoPress identifies optimal compression profiles in a highly efficient manner, and generalizes across diverse models and compression techniques. Via EvoPress, we achieve state-of-the-art performance for dynamic compression of Llama, Mistral, and Phi models, setting new benchmarks for structural pruning (block/layer dropping), unstructured sparsity, and quantization with dynamic bitwidths. Our code is available at https://github.com/IST-DASLab/EvoPress}.

Roberto L. Castro, Andrei Panferov, Soroush Tabesh et al.

Training large language models (LLMs) models directly in low-precision offers a way to address computational costs by improving both throughput and energy efficiency. For those purposes, NVIDIA's recent Blackwell architecture facilitates very low-precision operations using FP4 variants. Yet, current algorithms for training LLMs in FP4 precision face significant accuracy degradation and often rely on mixed-precision fallbacks. In this paper, we investigate hardware-supported FP4 training and introduce a new approach for accurate, end-to-end FP4 training with all the major computations (i.e., linear layers) in low precision. Through extensive evaluations on Llama-type models, we reveal a new low-precision scaling law that quantifies performance trade-offs across bit-widths and training setups. Guided by this investigation, we design an "optimal" technique in terms of accuracy-vs-computation, called Quartet. We implement Quartet using optimized CUDA kernels tailored for Blackwell, demonstrating that fully FP4-based training is a competitive alternative to FP16 half-precision and to FP8 training. Our code is available at https://github.com/IST-DASLab/Quartet.

Saleh Ashkboos, Mahdi Nikdan, Soroush Tabesh et al.

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 can have large weight and activation outlier values that make lower-precision optimization difficult. To address this, 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, which mitigate outliers, 2) high-performance kernel support, and 3) FSDP integration for low-precision communication. 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 1.41x end-to-end speedup for full fine-tuning on RTX 4090 GPUs. HALO efficiently supports both standard and parameterefficient fine-tuning (PEFT). Our results demonstrate the first practical approach to fully quantized LLM fine-tuning that maintains accuracy in 8-bit precision, while delivering performance benefits. Code is available at https://github.com/IST-DASLab/HALO.

Eugenia Iofinova, Andrej Jovanovic, Dan Alistarh

The recent surge in high-quality open-source Generative AI text models (colloquially: LLMs), as well as efficient finetuning techniques, have opened the possibility of creating high-quality personalized models that generate text attuned to a specific individual's needs and are capable of credibly imitating their writing style by refining an open-source model with that person's own data. The technology to create such models is accessible to private individuals, and training and running such models can be done cheaply on consumer-grade hardware. While these advancements are a huge gain for usability and privacy, this position paper argues that the practical feasibility of impersonating specific individuals also introduces novel safety risks. For instance, this technology enables the creation of phishing emails or fraudulent social media accounts, based on small amounts of publicly available text, or by the individuals themselves to escape AI text detection. We further argue that these risks are complementary to - and distinct from - the much-discussed risks of other impersonation attacks such as image, voice, or video deepfakes, and are not adequately addressed by the larger research community, or the current generation of open- and closed-source models.

Vage Egiazarian, Andrei Panferov, Denis Kuznedelev et al.

The emergence of accurate open large language models (LLMs) has led to a race towards performant quantization techniques which can enable their execution on end-user devices. In this paper, we revisit the problem of "extreme" LLM compression-defined as targeting extremely low bit counts, such as 2 to 3 bits per parameter-from the point of view of classic methods in Multi-Codebook Quantization (MCQ). Our algorithm, called AQLM, generalizes the classic Additive Quantization (AQ) approach for information retrieval to advance the state-of-the-art in LLM compression, via two innovations: 1) learned additive quantization of weight matrices in input-adaptive fashion, and 2) joint optimization of codebook parameters across each transformer blocks. Broadly, AQLM is the first scheme that is Pareto optimal in terms of accuracy-vs-model-size when compressing to less than 3 bits per parameter, and significantly improves upon all known schemes in the extreme compression (2bit) regime. In addition, AQLM is practical: we provide fast GPU and CPU implementations of AQLM for token generation, which enable us to match or outperform optimized FP16 implementations for speed, while executing in a much smaller memory footprint.

Soroush Tabesh, Mher Safaryan, Andrei Panferov et al.

Despite significant work on low-bit quantization-aware training (QAT), there is still an accuracy gap between such techniques and native training. To address this, we introduce CAGE (Curvature-Aware Gradient Estimation), a new QAT method that augments the straight-through estimator (STE) gradient with a curvature-aware correction designed to counteract the loss increase induced by quantization. CAGE is derived from a multi-objective view of QAT that balances loss minimization with the quantization constraints, yielding a principled correction term that depends on local curvature information. On the theoretical side, we introduce the notion of Pareto-optimal solutions for quantized optimization, and establish that CAGE yields strong convergence guarantees in the smooth non-convex setting. In terms of implementation, our approach is optimizer-agnostic, but we provide a highly-efficient implementation that leverages Adam statistics. CAGE significantly improves upon the prior state-of-the-art methods in terms of accuracy, for similar computational cost: for QAT fine-tuning, it halves the compression accuracy loss relative to the prior best method, while for QAT pre-training of Llama models, its accuracy for 3-bit weights-and-activations (W3A3) matches the accuracy achieved at 4-bits (W4A4) with the prior best method. The official implementation can be found over https://github.com/IST-DASLab/CAGE .

Ionut-Vlad Modoranu, Mher Safaryan, Erik Schultheis et al.

Low-rank optimization has emerged as a promising direction in training large language models (LLMs) to improve running time and 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) or QR-decomposition. Applying these techniques 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/QR-based gradient projections into lower-dimensional spaces by using a predefined orthogonal matrix of the Discrete Cosine Transform (DCT). We dynamically select columns from the DCT matrix based on their alignment with the gradient of each layer. The effective projection matrices are obtained via a simple matmul with the DCT matrix in $O(n^3)$ time, followed by a lightweight sorting step to identify the most relevant basis vectors. For large layers, DCT can be computed via Makhoul's $N$-point algorithm based on Fast Fourier Transform (FFT) in $O(n^2 \log(n))$ time. Due to the predefined nature of the orthogonal bases, they are computed once at the start of training. Our numerical experiments on both pre-training and fine-tuning tasks demonstrate the effectiveness of our dual strategy in approximating optimal low-rank projections, obtaining an approach with rank-independent running time that matches the performance of costly SVD/QR-based methods while achieving faster runtime and reduced memory usage by up to $25\%$ across different model sizes. Our code is available at \href{https://github.com/IST-DASLab/ISTA-DASLab-Optimizers}{\texttt{https://github.com/IST-DASLab/ISTA-DASLab-Optimizers}}.

Armand Nicolicioiu, Eugenia Iofinova, Andrej Jovanovic et al.

The availability of powerful open-source large language models (LLMs) opens exciting use-cases, such as using personal data to fine-tune these models to imitate a user's unique writing style. Two key requirements for such assistants are personalization - in the sense that the assistant should recognizably reflect the user's own writing style - and privacy - users may justifiably be wary of uploading extremely personal data, such as their email archive, to a third-party service. In this paper, we present a new design and evaluation for such an automated assistant, for the specific use case of email generation, which we call Panza. Panza's personalization features are based on a combination of fine-tuning using a variant of the Reverse Instructions technique together with Retrieval-Augmented Generation (RAG). We demonstrate that this combination allows us to fine-tune an LLM to reflect a user's writing style using limited data, while executing on extremely limited resources, e.g. on a free Google Colab instance. Our key methodological contribution is the first detailed study of evaluation metrics for this personalized writing task, and of how different choices of system components--the use of RAG and of different fine-tuning approaches-impact the system's performance. Additionally, we demonstrate that very little data - under 100 email samples - are sufficient to create models that convincingly imitate humans. This finding showcases a previously-unknown attack vector in language models - that access to a small number of writing samples can allow a bad actor to cheaply create generative models that imitate a target's writing style. We are releasing the full Panza code as well as three new email datasets licensed for research use at https://github.com/IST-DASLab/PanzaMail.

Eldar Kurtic, Amir Moeini, Dan Alistarh

We introduce Mathador-LM, a new benchmark for evaluating the mathematical reasoning on large language models (LLMs), combining ruleset interpretation, planning, and problem-solving. This benchmark is inspired by the Mathador game, where the objective is to reach a target number using basic arithmetic operations on a given set of base numbers, following a simple set of rules. We show that, across leading LLMs, we obtain stable average performance while generating benchmark instances \emph{dynamically}, following a target difficulty level. Thus, our benchmark alleviates concerns about test-set leakage into training data, an issue that often undermines popular benchmarks. Additionally, we conduct a comprehensive evaluation of both open and closed-source state-of-the-art LLMs on Mathador-LM. Our findings reveal that contemporary models struggle with Mathador-LM, scoring significantly lower than average 3rd graders. This stands in stark contrast to their strong performance on popular mathematical reasoning benchmarks. The implementation of Mathador-LM benchmark is available at \href{https://github.com/IST-DASLab/Mathador-LM}{github.com/IST-DASLab/Mathador-LM}.

Alexandra Peste, Eugenia Iofinova, Adrian Vladu et al.

The increasing computational requirements of deep neural networks (DNNs) have led to significant interest in obtaining DNN models that are sparse, yet accurate. Recent work has investigated the even harder case of sparse training, where the DNN weights are, for as much as possible, already sparse to reduce computational costs during training. Existing sparse training methods are often empirical and can have lower accuracy relative to the dense baseline. In this paper, we present a general approach called Alternating Compressed/DeCompressed (AC/DC) training of DNNs, demonstrate convergence for a variant of the algorithm, and show that AC/DC outperforms existing sparse training methods in accuracy at similar computational budgets; at high sparsity levels, AC/DC even outperforms existing methods that rely on accurate pre-trained dense models. An important property of AC/DC is that it allows co-training of dense and sparse models, yielding accurate sparse-dense model pairs at the end of the training process. This is useful in practice, where compressed variants may be desirable for deployment in resource-constrained settings without re-doing the entire training flow, and also provides us with insights into the accuracy gap between dense and compressed models. The code is available at: https://github.com/IST-DASLab/ACDC .

Sidak Pal Singh, Dan Alistarh

Second-order information, in the form of Hessian- or Inverse-Hessian-vector products, is a fundamental tool for solving optimization problems. Recently, there has been significant interest in utilizing this information in the context of deep neural networks; however, relatively little is known about the quality of existing approximations in this context. Our work examines this question, identifies issues with existing approaches, and proposes a method called WoodFisher to compute a faithful and efficient estimate of the inverse Hessian. Our main application is to neural network compression, where we build on the classic Optimal Brain Damage/Surgeon framework. We demonstrate that WoodFisher significantly outperforms popular state-of-the-art methods for one-shot pruning. Further, even when iterative, gradual pruning is considered, our method results in a gain in test accuracy over the state-of-the-art approaches, for pruning popular neural networks (like ResNet-50, MobileNetV1) trained on standard image classification datasets such as ImageNet ILSVRC. We examine how our method can be extended to take into account first-order information, as well as illustrate its ability to automatically set layer-wise pruning thresholds and perform compression in the limited-data regime. The code is available at the following link, https://github.com/IST-DASLab/WoodFisher.

David Dao, Dan Alistarh, Claudiu Musat et al.

It is safe to assume that, for the foreseeable future, machine learning, especially deep learning will remain both data- and computation-hungry. In this paper, we ask: Can we build a global exchange where everyone can contribute computation and data to train the next generation of machine learning applications? We present an early, but running prototype of DataBright, a system that turns the creation of training examples and the sharing of computation into an investment mechanism. Unlike most crowdsourcing platforms, where the contributor gets paid when they submit their data, DataBright pays dividends whenever a contributor's data or hardware is used by someone to train a machine learning model. The contributor becomes a shareholder in the dataset they created. To enable the measurement of usage, a computation platform that contributors can trust is also necessary. DataBright thus merges both a data market and a trusted computation market. We illustrate that trusted computation can enable the creation of an AI market, where each data point has an exact value that should be paid to its creator. DataBright allows data creators to retain ownership of their contribution and attaches to it a measurable value. The value of the data is given by its utility in subsequent distributed computation done on the DataBright computation market. The computation market allocates tasks and subsequent payments to pooled hardware. This leads to the creation of a decentralized AI cloud. Our experiments show that trusted hardware such as Intel SGX can be added to the usual ML pipeline with no additional costs. We use this setting to orchestrate distributed computation that enables the creation of a computation market. DataBright is available for download at https://github.com/ds3lab/databright.

Vladimir Malinovskii, Denis Mazur, Ivan Ilin et al.

There has been significant interest in "extreme" compression of large language models (LLMs), i.e., to 1-2 bits per parameter, which allows such models to be executed efficiently on resource-constrained devices. Existing work focused on improved one-shot quantization techniques and weight representations; yet, purely post-training approaches are reaching diminishing returns in terms of the accuracy-vs-bit-width trade-off. State-of-the-art quantization methods such as QuIP# and AQLM include fine-tuning (part of) the compressed parameters over a limited amount of calibration data; however, such fine-tuning techniques over compressed weights often make exclusive use of straight-through estimators (STE), whose performance is not well-understood in this setting. In this work, we question the use of STE for extreme LLM compression, showing that it can be sub-optimal, and perform a systematic study of quantization-aware fine-tuning strategies for LLMs. We propose PV-Tuning - a representation-agnostic framework that generalizes and improves upon existing fine-tuning strategies, and provides convergence guarantees in restricted cases. On the practical side, when used for 1-2 bit vector quantization, PV-Tuning outperforms prior techniques for highly-performant models such as Llama and Mistral. Using PV-Tuning, we achieve the first Pareto-optimal quantization for Llama 2 family models at 2 bits per parameter.

Gleb Rodionov, Roman Garipov, Alina Shutova et al.

Large Language Models (LLMs) have demonstrated the ability to tackle increasingly complex tasks through advanced reasoning, long-form content generation, and tool use. Solving these tasks often involves long inference-time computations. In human problem solving, a common strategy to expedite work is collaboration: by dividing the problem into sub-tasks, exploring different strategies concurrently, etc. Recent research has shown that LLMs can also operate in parallel by implementing explicit cooperation frameworks, such as voting mechanisms or the explicit creation of independent sub-tasks that can be executed in parallel. However, each of these frameworks may not be suitable for all types of tasks, which can hinder their applicability. In this work, we propose a different design approach: we run LLM "workers" in parallel , allowing them to synchronize via a concurrently-updated attention cache and prompt these workers to decide how best to collaborate. Our approach allows the LLM instances to come up with their own collaboration strategy for the problem at hand, all the while "seeing" each other's memory in the concurrent KV cache. We implement this approach via Hogwild! Inference: a parallel LLM inference engine where multiple instances of the same LLM run in parallel with the same attention cache, with "instant" access to each other's memory. Hogwild! Inference takes advantage of Rotary Position Embeddings (RoPE) to avoid recomputation while improving parallel hardware utilization. We find that modern reasoning-capable LLMs can perform inference with shared Key-Value cache out of the box, without additional fine-tuning.

Aniruddha Nrusimha, Mayank Mishra, Naigang Wang et al.

We consider the problem of accurate quantization for language models, where both the weights and activations are uniformly quantized to 4 bits per parameter, the lowest bitwidth format natively supported by GPU hardware. In this context, the key challenge is activation quantization: it is known that language models contain outlier channels whose values on average are orders of magnitude higher than than other channels, which prevents accurate low-bitwidth quantization with known techniques. We systematically study this phenomena and find that these outlier channels emerge early in training, and that they occur more frequently in layers with residual streams. We then propose a simple strategy which regularizes a layer's inputs via quantization-aware training (QAT) and its outputs via activation kurtosis regularization. We show that regularizing both the inputs and outputs is crucial for preventing a model's "migrating" the difficulty in input quantization to the weights, which makes post-training quantization (PTQ) of weights more difficult. When combined with weight PTQ, we show that our approach can obtain a W4A4 model that performs competitively to the standard-precision W16A16 baseline.

Vladimir Malinovskii, Andrei Panferov, Ivan Ilin et al.

Quantizing large language models has become a standard way to reduce their memory and computational costs. Typically, existing methods focus on breaking down the problem into individual layer-wise sub-problems, and minimizing per-layer error, measured via various metrics. Yet, this approach currently lacks theoretical justification and the metrics employed may be sub-optimal. In this paper, we present a "linearity theorem" establishing a direct relationship between the layer-wise $\ell_2$ reconstruction error and the model perplexity increase due to quantization. This insight enables two novel applications: (1) a simple data-free LLM quantization method using Hadamard rotations and MSE-optimal grids, dubbed HIGGS, which outperforms all prior data-free approaches such as the extremely popular NF4 quantized format, and (2) an optimal solution to the problem of finding non-uniform per-layer quantization levels which match a given compression constraint in the medium-bitwidth regime, obtained by reduction to dynamic programming. On the practical side, we demonstrate improved accuracy-compression trade-offs on Llama-3.1 and 3.2-family models, as well as on Qwen-family models. Further, we show that our method can be efficiently supported in terms of GPU kernels at various batch sizes, advancing both data-free and non-uniform quantization for LLMs.