Extreme low-bit quantization is critical for efficiently deploying Large Language Models (LLMs), yet it often leads to severe performance degradation at 2-bits and even 4-bits (e.g., MXFP4). We present SignRoundV2, a post-training quantization framework that is highly effective even without mixed-precision. SignRoundV2 introduces (1) a fast sensitivity metric that combines gradient information with quantization-induced deviations to guide layer-wise bit allocation, and (2) a lightweight pre-tuning search for quantization scales to improve extremely low-bit quantization. These components allow SignRoundV2 to close the gap with full-precision models. Extensive experiments indicate that our method sustains competitive accuracy for LLMs, achieving production-grade performance with about 1 percent variance at 4-5 bits and strong results even at 2 bits. The implementation is available at https://github.com/intel/auto-round.

Accurate speech emotion recognition is essential for developing human-facing systems. Recent advancements have included finetuning large, pretrained transformer models like Wav2Vec 2.0. However, the finetuning process requires substantial computational resources, including high-memory GPUs and significant processing time. As the demand for accurate emotion recognition continues to grow, efficient finetuning approaches are needed to reduce the computational burden. Our study focuses on dimensional emotion recognition, predicting attributes such as activation (calm to excited) and valence (negative to positive). We present various finetuning techniques, including full finetuning, partial finetuning of transformer layers, finetuning with mixed precision, partial finetuning with caching, and low-rank adaptation (LoRA) on the Wav2Vec 2.0 base model. We find that partial finetuning with mixed precision achieves performance comparable to full finetuning while increasing training speed by 67%. Caching intermediate representations further boosts efficiency, yielding an 88% speedup and a 71% reduction in learnable parameters. We recommend finetuning the final three transformer layers in mixed precision to balance performance and training efficiency, and adding intermediate representation caching for optimal speed with minimal performance trade-offs. These findings lower the barriers to finetuning speech emotion recognition systems, making accurate emotion recognition more accessible to a broader range of researchers and practitioners.

Large Language Models (LLMs) exhibit impressive performance across various tasks, but deploying them for inference poses challenges. Their high resource demands often necessitate complex, costly multi-GPU pipelines, or the use of smaller, less capable models. While quantization offers a promising solution utilizing lower precision for model storage, existing methods frequently experience significant performance drops at lower precision levels. Additionally, they typically provide only a limited set of solutions at specific bit levels, many of which are extensively manually tuned. To address these challenges, we propose a new method called SKIM: Scaled K-means clustering wIth Mixed precision. Our approach introduces two novel techniques: 1. A greedy algorithm to solve approximately optimal bit allocation across weight channels, and 2. A trainable scaling vector for non-differentiable K-means clustering. These techniques substantially improve performance and can be adapted to any given bit. Notably, in terms of model perplexity, our method narrows the gap between 3-bit quantized LLaMA models and their full precision counterparts by 16.3% on average.

This study was the 2nd part of my dissertation for my master degree and compared the power consumption using the Comma-Separated-Values (CSV) and parquet dataset format with the default floating point (32bit) and Nvidia mixed precision (16bit and 32bit) while training a regression ML model. The same custom PC as per the 1st part, which was dedicated to the classification testing and analysis, was built to perform the experiments, and different ML hyper-parameters, such as batch size, neurons, and epochs, were chosen to build Deep Neural Networks (DNN). A benchmarking test with default hyper-parameter values for the DNN was used as a reference, while the experiments used a combination of different settings. The results were recorded in Excel, and descriptive statistics were chosen to calculate the mean between the groups and compare them using graphs and tables. The outcome was positive when using mixed precision combined with specific hyper-parameters. Compared to the benchmarking, optimising the regression models reduced the power consumption between 7 and 11 Watts. The regression results show that while mixed precision can help improve power consumption, we must carefully consider the hyper-parameters. A high number of batch sizes and neurons will negatively affect power consumption. However, this research required inferential statistics, specifically ANOVA and T-test, to compare the relationship between the means. The results reported no statistical significance between the means in the regression tests and accepted H0. Therefore, choosing different ML techniques and the Parquet dataset format will not improve the computational power consumption and the overall ML carbon footprint. However, a more extensive implementation with a cluster of GPUs can increase the sample size significantly, as it is an essential factor and can change the outcome of the statistical analysis.

This study was part of my dissertation for my master degree and compares the power consumption using the default floating point (32bit) and Nvidia mixed precision (16bit and 32bit) while training a classification ML model. A custom PC with specific hardware was built to perform the experiments, and different ML hyper-parameters, such as batch size, neurons, and epochs, were chosen to build Deep Neural Networks (DNN). Additionally, various software was used during the experiments to collect the power consumption data in Watts from the Graphics Processing Unit (GPU), Central Processing Unit (CPU), Random Access Memory (RAM) and manually from a wattmeter connected to the wall. A benchmarking test with default hyper parameter values for the DNN was used as a reference, while the experiments used a combination of different settings. The results were recorded in Excel, and descriptive statistics were chosen to calculate the mean between the groups and compare them using graphs and tables. The outcome was positive when using mixed precision combined with specific hyper-parameters. Compared to the benchmarking, the optimisation for the classification reduced the power consumption between 7 and 11 Watts. Similarly, the carbon footprint is reduced because the calculation uses the same power consumption data. Still, a consideration is required when configuring hyper-parameters because it can negatively affect hardware performance. However, this research required inferential statistics, specifically ANOVA and T-test, to compare the relationship between the means. Furthermore, tests indicated no statistical significance of the relationship between the benchmarking and experiments. However, a more extensive implementation with a cluster of GPUs can increase the sample size significantly, as it is an essential factor and can change the outcome of the statistical analysis.

Scientific machine learning (SciML) has emerged as a versatile approach to address complex computational science and engineering problems. Within this field, physics-informed neural networks (PINNs) and deep operator networks (DeepONets) stand out as the leading techniques for solving partial differential equations by incorporating both physical equations and experimental data. However, training PINNs and DeepONets requires significant computational resources, including long computational times and large amounts of memory. In search of computational efficiency, training neural networks using half precision (float16) rather than the conventional single (float32) or double (float64) precision has gained substantial interest, given the inherent benefits of reduced computational time and memory consumed. However, we find that float16 cannot be applied to SciML methods, because of gradient divergence at the start of training, weight updates going to zero, and the inability to converge to a local minima. To overcome these limitations, we explore mixed precision, which is an approach that combines the float16 and float32 numerical formats to reduce memory usage and increase computational speed. Our experiments showcase that mixed precision training not only substantially decreases training times and memory demands but also maintains model accuracy. We also reinforce our empirical observations with a theoretical analysis. The research has broad implications for SciML in various computational applications.

Diffusion models have gained popularity for generating images from textual descriptions. Nonetheless, the substantial need for computational resources continues to present a noteworthy challenge, contributing to time-consuming processes. Quantization, a technique employed to compress deep learning models for enhanced efficiency, presents challenges when applied to diffusion models. These models are notably more sensitive to quantization compared to other model types, potentially resulting in a degradation of image quality. In this paper, we introduce a novel approach to quantize the diffusion models by leveraging both quantization-aware training and distillation. Our results show the quantized models can maintain the high image quality while demonstrating the inference efficiency on CPUs.

Reducing the number of bits needed to encode the weights and activations of neural networks is highly desirable as it speeds up their training and inference time while reducing memory consumption. It is unsurprising that considerable attention has been drawn to developing neural networks that employ lower-precision computation. This includes IEEE 16-bit, Google bfloat16, 8-bit, 4-bit floating-point or fixed-point, 2-bit, and various mixed-precision algorithms. Out of these low-precision formats, IEEE 16-bit stands out due to its universal compatibility with contemporary GPUs. This accessibility contrasts with bfloat16, which needs high-end GPUs, or other non-standard fewer-bit designs, which typically require software simulation. This study focuses on the widely accessible IEEE 16-bit format for comparative analysis. This analysis involves an in-depth theoretical investigation of the factors that lead to discrepancies between 16-bit and 32-bit models, including a formalization of the concepts of floating-point error and tolerance to understand the conditions under which a 16-bit model can approximate 32-bit results. Contrary to literature that credits the success of noise-tolerated neural networks to regularization effects, our study-supported by a series of rigorous experiments-provides a quantitative explanation of why standalone IEEE 16-bit floating-point neural networks can perform on par with 32-bit and mixed-precision networks in various image classification tasks. Because no prior research has studied IEEE 16-bit as a standalone floating-point precision in neural networks, we believe our findings will have significant impacts, encouraging the adoption of standalone IEEE 16-bit networks in future neural network applications.

Neural network quantization is frequently used to optimize model size, latency and power consumption for on-device deployment of neural networks. In many cases, a target bit-width is set for an entire network, meaning every layer get quantized to the same number of bits. However, for many networks some layers are significantly more robust to quantization noise than others, leaving an important axis of improvement unused. As many hardware solutions provide multiple different bit-width settings, mixed-precision quantization has emerged as a promising solution to find a better performance-efficiency trade-off than homogeneous quantization. However, most existing mixed precision algorithms are rather difficult to use for practitioners as they require access to the training data, have many hyper-parameters to tune or even depend on end-to-end retraining of the entire model. In this work, we present a simple post-training mixed precision algorithm that only requires a small unlabeled calibration dataset to automatically select suitable bit-widths for each layer for desirable on-device performance. Our algorithm requires no hyper-parameter tuning, is robust to data variation and takes into account practical hardware deployment constraints making it a great candidate for practical use. We experimentally validate our proposed method on several computer vision tasks, natural language processing tasks and many different networks, and show that we can find mixed precision networks that provide a better trade-off between accuracy and efficiency than their homogeneous bit-width equivalents.

One of the keys to optimizing the runtime performance of your deep neural network (DNN) training workloads is to maximize the utilization of your training instance's resources. This is particularly true of the resources of the GPU, or other training accelerator, typically the most expensive component of your training device. Our focus in this post will be on the memory utilization of the GPU (or alternative training accelerator). For additional tips on training performance optimization be sure to check out some of our other blog posts (e.g. For the sake of simplicity, whenever we refer to GPU memory, we are referring more generally to the memory of any training accelerator, including GPU, Google Cloud TPU, Habana Gaudi, etc. The most basic example of GPU memory optimization is increasing your batch size to increase the memory utilization up to as close to 100% as possible. Generally speaking (but not always), your overall training throughput will increase.