Quantization has established itself as the primary approach for decreasing the computational and storage expenses associated with Large Language Models (LLMs) inference. The majority of current research emphasizes quantizing weights and activations to enable low-bit general-matrix-multiply (GEMM) operations, with the remaining non-linear operations executed at higher precision. In our study, we discovered that following the application of these techniques, the primary bottleneck in LLMs inference lies in the softmax layer. The softmax operation comprises three phases: exponent calculation, accumulation, and normalization, Our work focuses on optimizing the first two phases. We propose an analytical approach to determine the optimal clipping value for the input to the softmax function, enabling sub-4-bit quantization for LLMs inference. This method accelerates the calculations of both $e^x$ and $\sum(e^x)$ with minimal to no accuracy degradation. For example, in LLaMA1-30B, we achieve baseline performance with 2-bit quantization on the well-known "Physical Interaction: Question Answering" (PIQA) dataset evaluation. This ultra-low bit quantization allows, for the first time, an acceleration of approximately 4x in the accumulation phase. The combination of accelerating both $e^x$ and $\sum(e^x)$ results in a 36.9% acceleration in the softmax operation.

Quantization of the weights and activations is one of the main methods to reduce the computational footprint of Deep Neural Networks (DNNs) training. Current methods enable 4-bit quantization of the forward phase. However, this constitutes only a third of the training process. Reducing the computational footprint of the entire training process requires the quantization of the neural gradients, i.e., the loss gradients with respect to the outputs of intermediate neural layers. Previous works separately showed that accurate 4-bit quantization of the neural gradients needs to (1) be unbiased and (2) have a log scale. However, no previous work aimed to combine both ideas, as we do in this work. Specifically, we examine the importance of having unbiased quantization in quantized neural network training, where to maintain it, and how to combine it with logarithmic quantization. Based on this, we suggest a logarithmic unbiased quantization (LUQ) method to quantize both the forward and backward phases to 4-bit, achieving state-of-the-art results in 4-bit training without the overhead. For example, in ResNet50 on ImageNet, we achieved a degradation of 1.1%. We further improve this to a degradation of only 0.32% after three epochs of high precision fine-tuning, combined with a variance reduction method--where both these methods add overhead comparable to previously suggested methods. Deep neural networks (DNNs) training consists of three main general-matrix-multiply (GEMM) phases: the forward phase, backward phase, and update phase. Quantization has become one of the main methods to compress DNNs and reduce the GEMM computational resources. Previous works showed the weights and activations in the forward pass to 4 bits while preserving model accuracy (Banner et al., 2019; Nahshan et al., 2019; Bhalgat et al., 2020; Choi et al., 2018b).

In deep learning, fine-grained N:M sparsity reduces the data footprint and bandwidth of a General Matrix multiply (GEMM) up to x2, and doubles throughput by skipping computation of zero values. So far, it was mainly only used to prune weights to accelerate the forward and backward phases. We examine how this method can be used also for the neural gradients (i.e., loss gradients with respect to the intermediate neural layer outputs). To this end, we first establish a tensorlevel optimality criteria. Previous works aimed to minimize the mean-square-error (MSE) of each pruned block. We show that while minimization of the MSE works fine for pruning the weights and activations, it catastrophically fails for the neural gradients. Instead, we show that accurate pruning of the neural gradients requires an unbiased minimum-variance pruning mask. We design such specialized masks, and find that in most cases, 1:2 sparsity is sufficient for training, and 2:4 sparsity is usually enough when this is not the case. Further, we suggest combining several such methods together in order to potentially speed up training even more. Pruning Deep Neural Networks (DNNs) is one of the most effective and widely studied methods to improve DNN resource efficiency. Since DNNs are over-parametrized, most researchers focused on weights pruning.

Recently, researchers proposed pruning deep neural network weights (DNNs) using an $N:M$ fine-grained block sparsity mask. In this mask, for each block of $M$ weights, we have at least $N$ zeros. In contrast to unstructured sparsity, $N:M$ fine-grained block sparsity allows acceleration in actual modern hardware. So far, this was used for DNN acceleration at the inference phase. First, we suggest a method to convert a pretrained model with unstructured sparsity to a $N:M$ fine-grained block sparsity model, with little to no training. Then, to also allow such acceleration in the training phase, we suggest a novel transposable-fine-grained sparsity mask where the same mask can be used for both forward and backward passes. Our transposable mask ensures that both the weight matrix and its transpose follow the same sparsity pattern; thus the matrix multiplication required for passing the error backward can also be accelerated. We discuss the transposable constraint and devise a new measure for mask constraints, called mask-diversity (MD), which correlates with their expected accuracy. Then, we formulate the problem of finding the optimal transposable mask as a minimum-cost-flow problem and suggest a fast linear approximation that can be used when the masks dynamically change while training. Our experiments suggest 2x speed-up with no accuracy degradation over vision and language models. A reference implementation can be found at https://github.com/papers-submission/structured_transposable_masks.