Yujia Zhai

Felix Abecassis, Anjulie Agrusa, Dong Ahn et al. · nvidia

Large Language Models (LLMs) today are powerful problem solvers across many domains, and they continue to get stronger as they scale in model size, training set size, and training set quality, as shown by extensive research and experimentation across the industry. Training a frontier model today requires on the order of tens to hundreds of yottaflops, which is a massive investment of time, compute, and energy. Improving pretraining efficiency is therefore essential to enable the next generation of even more capable LLMs. While 8-bit floating point (FP8) training is now widely adopted, transitioning to even narrower precision, such as 4-bit floating point (FP4), could unlock additional improvements in computational speed and resource utilization. However, quantization at this level poses challenges to training stability, convergence, and implementation, notably for large-scale models trained on long token horizons. In this study, we introduce a novel approach for stable and accurate training of large language models (LLMs) using the NVFP4 format. Our method integrates Random Hadamard transforms (RHT) to bound block-level outliers, employs a two-dimensional quantization scheme for consistent representations across both the forward and backward passes, utilizes stochastic rounding for unbiased gradient estimation, and incorporates selective high-precision layers. We validate our approach by training a 12-billion-parameter model on 10 trillion tokens -- the longest publicly documented training run in 4-bit precision to date. Our results show that the model trained with our NVFP4-based pretraining technique achieves training loss and downstream task accuracies comparable to an FP8 baseline. These findings highlight that NVFP4, when combined with our training approach, represents a major step forward in narrow-precision LLM training algorithms.

Yujia Zhai, Chengquan Jiang, Leyuan Wang et al.

Transformers have become keystone models in natural language processing over the past decade. They have achieved great popularity in deep learning applications, but the increasing sizes of the parameter spaces required by transformer models generate a commensurate need to accelerate performance. Natural language processing problems are also routinely faced with variable-length sequences, as word counts commonly vary among sentences. Existing deep learning frameworks pad variable-length sequences to a maximal length, which adds significant memory and computational overhead. In this paper, we present ByteTransformer, a high-performance transformer boosted for variable-length inputs. We propose a padding-free algorithm that liberates the entire transformer from redundant computations on zero padded tokens. In addition to algorithmic-level optimization, we provide architecture-aware optimizations for transformer functional modules, especially the performance-critical algorithm Multi-Head Attention (MHA). Experimental results on an NVIDIA A100 GPU with variable-length sequence inputs validate that our fused MHA outperforms PyTorch by 6.13x. The end-to-end performance of ByteTransformer for a forward BERT transformer surpasses state-of-the-art transformer frameworks, such as PyTorch JIT, TensorFlow XLA, Tencent TurboTransformer, Microsoft DeepSpeed-Inference and NVIDIA FasterTransformer, by 87\%, 131\%, 138\%, 74\% and 55\%, respectively. We also demonstrate the general applicability of our optimization methods to other BERT-like models, including ALBERT, DistilBERT, and DeBERTa.

Shixun Wu, Yitong Ding, Yujia Zhai et al.

K-means is a widely used algorithm in clustering, however, its efficiency is primarily constrained by the computational cost of distance computing. Existing implementations suffer from suboptimal utilization of computational units and lack resilience against soft errors. To address these challenges, we introduce FT K-means, a high-performance GPU-accelerated implementation of K-means with online fault tolerance. We first present a stepwise optimization strategy that achieves competitive performance compared to NVIDIA's cuML library. We further improve FT K-means with a template-based code generation framework that supports different data types and adapts to different input shapes. A novel warp-level tensor-core error correction scheme is proposed to address the failure of existing fault tolerance methods due to memory asynchronization during copy operations. Our experimental evaluations on NVIDIA T4 GPU and A100 GPU demonstrate that FT K-means without fault tolerance outperforms cuML's K-means implementation, showing a performance increase of 10\%-300\% in scenarios involving irregular data shapes. Moreover, the fault tolerance feature of FT K-means introduces only an overhead of 11\%, maintaining robust performance even with tens of errors injected per second.

Shuyin Xia, Cheng Wang, Guoyin Wang et al.

Pawlak rough set and neighborhood rough set are the two most common rough set theoretical models. Pawlak can use equivalence classes to represent knowledge, but it cannot process continuous data; neighborhood rough sets can process continuous data, but it loses the ability of using equivalence classes to represent knowledge. To this end, this paper presents a granular-ball rough set based on the granular-ball computing. The granular-ball rough set can simultaneously represent Pawlak rough sets, and the neighborhood rough set, so as to realize the unified representation of the two. This makes the granular-ball rough set not only can deal with continuous data, but also can use equivalence classes for knowledge representation. In addition, we propose an implementation algorithms of granular-ball rough sets. The experimental results on benchmark datasets demonstrate that, due to the combination of the robustness and adaptability of the granular-ball computing, the learning accuracy of the granular-ball rough set has been greatly improved compared with the Pawlak rough set and the traditional neighborhood rough set. The granular-ball rough set also outperforms nine popular or the state-of-the-art feature selection methods.

Yujia Zhai, Mohannad Ibrahim, Yiqin Qiu et al.

Homomorphic Encryption (HE) is an emerging encryption scheme that allows computations to be performed directly on encrypted messages. This property provides promising applications such as privacy-preserving deep learning and cloud computing. Prior works have been proposed to enable practical privacy-preserving applications with architectural-aware optimizations on CPUs, GPUs and FPGAs. However, there is no systematic optimization for the whole HE pipeline on Intel GPUs. In this paper, we present the first-ever SYCL-based GPU backend for Microsoft SEAL APIs. We perform optimizations from instruction level, algorithmic level and application level to accelerate our HE library based on the Cheon, Kim, Kimand Song (CKKS) scheme on Intel GPUs. The performance is validated on two latest Intel GPUs. Experimental results show that our staged optimizations together with optimizations including low-level optimizations and kernel fusion accelerate the Number Theoretic Transform (NTT), a key algorithm for HE, by up to 9.93X compared with the naïve GPU baseline. The roofline analysis confirms that our optimized NTT reaches 79.8% and85.7% of the peak performance on two GPU devices. Through the highly optimized NTT and the assembly-level optimization, we obtain 2.32X - 3.05X acceleration for HE evaluation routines. In addition, our all-together systematic optimizations improve the performance of encrypted element-wise polynomial matrix multiplication application by up to 3.10X.

Kai Zhao, Sheng Di, Sihuan Li et al.

Convolutional neural networks (CNNs) are becoming more and more important for solving challenging and critical problems in many fields. CNN inference applications have been deployed in safety-critical systems, which may suffer from soft errors caused by high-energy particles, high temperature, or abnormal voltage. Of critical importance is ensuring the stability of the CNN inference process against soft errors. Traditional fault tolerance methods are not suitable for CNN inference because error-correcting code is unable to protect computational components, instruction duplication techniques incur high overhead, and existing algorithm-based fault tolerance (ABFT) techniques cannot protect all convolution implementations. In this paper, we focus on how to protect the CNN inference process against soft errors as efficiently as possible, with the following three contributions. (1) We propose several systematic ABFT schemes based on checksum techniques and analyze their fault protection ability and runtime thoroughly.Unlike traditional ABFT based on matrix-matrix multiplication, our schemes support any convolution implementations. (2) We design a novel workflow integrating all the proposed schemes to obtain a high detection/correction ability with limited total runtime overhead. (3) We perform our evaluation using ImageNet with well-known CNN models including AlexNet, VGG-19, ResNet-18, and YOLOv2. Experimental results demonstrate that our implementation can handle soft errors with very limited runtime overhead (4%~8% in both error-free and error-injected situations).

Jinyang Liu, Yujia Zhai, Zizhong Chen

Neural Network based models have been state-of-the-art models for various Natural Language Processing tasks, however, the input and output dimension problem in the networks has still not been fully resolved, especially in text generation tasks (e.g. Machine Translation, Text Summarization), in which input and output both have huge sizes of vocabularies. Therefore, input-output embedding weight sharing has been introduced and adopted widely, which remains to be improved. Based on linear algebra and statistical theories, this paper locates the shortcoming of existed input-output embedding weight sharing method, then raises methods for improving input-output weight shared embedding, among which methods of normalization of embedding weight matrices show best performance. These methods are nearly computational cost-free, can get combined with other embedding techniques, and show good effectiveness when applied on state-of-the-art Neural Network models. For Transformer-big models, the normalization techniques can get at best 0.6 BLEU improvement compared to the original version of model on WMT'16 En-De dataset, and similar BLEU improvements on IWSLT 14' datasets. For DynamicConv models, 0.5 BLEU improvement can be attained on WMT'16 En-De dataset, and 0.41 BLEU improvement on IWSLT 14' De-En translation task is achieved.