Qijun Luo

Qijun Luo, Hengxu Yu, Xiao Li

This work presents BAdam, an optimization method that leverages the block coordinate descent (BCD) framework with Adam's update rule. BAdam offers a memory efficient approach to the full parameter finetuning of large language models. We conduct a theoretical convergence analysis for BAdam in the deterministic case. Experimentally, we apply BAdam to finetune the Llama 3-8B and Llama 3-70B models using a single RTX3090-24GB GPU and 4 A100-80GB GPUs, respectively. The results confirm BAdam's efficiency in terms of memory usage, running time, and optimization capability. Furthermore, the downstream performance evaluation based on MT-bench and math benchmarks shows that BAdam outperforms existing memory efficient baselines such as LoRA. It also demonstrates that BAdam can achieve comparable or even superior performance compared to Adam. Finally, the ablation study using SGD's update rule illustrates the suitability of BCD for finetuning LLMs. Our code can be easily integrated into any PyTorch-based codebase and is available at https://github.com/Ledzy/BAdam.

Qijun Luo, Xiao Li

Decentralized Actor-Critic (AC) algorithms have been widely utilized for multi-agent reinforcement learning (MARL) and have achieved remarkable success. Apart from its empirical success, the theoretical convergence property of decentralized AC algorithms is largely unexplored. Most of the existing finite-time convergence results are derived based on either double-loop update or two-timescale step sizes rule, and this is the case even for centralized AC algorithm under a single-agent setting. In practice, the \emph{single-timescale} update is widely utilized, where actor and critic are updated in an alternating manner with step sizes being of the same order. In this work, we study a decentralized \emph{single-timescale} AC algorithm.Theoretically, using linear approximation for value and reward estimation, we show that the algorithm has sample complexity of $\tilde{\mathcal{O}}(\varepsilon^{-2})$ under Markovian sampling, which matches the optimal complexity with a double-loop implementation (here, $\tilde{\mathcal{O}}$ hides a logarithmic term). When we reduce to the single-agent setting, our result yields new sample complexity for centralized AC using a single-timescale update scheme. The central to establishing our complexity results is \emph{the hidden smoothness of the optimal critic variable} we revealed. We also provide a local action privacy-preserving version of our algorithm and its analysis. Finally, we conduct experiments to show the superiority of our algorithm over the existing decentralized AC algorithms.

Qijun Luo, Mengqi Li, Lei Zhao et al.

Training language models on long sequence data is a demanding requirement for enhancing the model's capability on complex tasks, e.g., long-chain reasoning. However, as the sequence length scales up, the memory cost for storing activation values becomes huge during the Backpropagation (BP) process, even with the application of gradient checkpointing technique. To tackle this challenge, we propose a memory-efficient and exact BP method called StreamBP, which performs a linear decomposition of the chain rule along the sequence dimension in a layer-wise manner, significantly reducing the memory cost of activation values and logits. The proposed method is applicable to common objectives such as SFT, GRPO, and DPO. From an implementation perspective, StreamBP achieves less computational FLOPs and faster BP speed by leveraging the causal structure of the language model. Compared to gradient checkpointing, StreamBP scales up the maximum sequence length of BP by 2.8-5.5 times larger, while using comparable or even less BP time. Note that StreamBP's sequence length scaling ability can be directly transferred to batch size scaling for accelerating training. We further develop a communication-efficient distributed StreamBP to effectively support multi-GPU training and broaden its applicability. Our code can be easily integrated into the training pipeline of any transformer models and is available at https://github.com/Ledzy/StreamBP.

Wenbo Gong, Javier Zazo, Qijun Luo et al.

Matrix-based optimizers have attracted growing interest for improving LLM training efficiency, with significant progress centered on orthogonalization/whitening based methods. While yielding substantial performance gains, a fundamental question arises: can we develop new paradigms beyond orthogonalization, pushing the efficiency frontier further? We present \textbf{Adaptively Rotated Optimization (ARO}, a new matrix optimization framework that treats gradient rotation as a first class design principle. ARO accelerates LLM training by performing normed steepest descent in a rotated coordinate system, where the rotation is determined by a novel norm-informed policy. This perspective yields update rules that go beyond existing orthogonalization and whitening optimizers, improving sample efficiency in practice. To make comparisons reliable, we propose a rigorously controlled benchmarking protocol that reduces confounding and bias. Under this protocol, ARO consistently outperforms AdamW (by 1.3 $\sim$1.35$\times$) and orthogonalization methods (by 1.1$\sim$1.15$\times$) in LLM pretraining at up to 8B activated parameters, and up to $8\times$ overtrain budget, without evidence of diminishing returns. Finally, we discuss how ARO can be reformulated as a symmetry-aware optimizer grounded in rotational symmetries of residual streams, motivating advanced designs that enable computationally efficient exploitation of cross-layer/cross module couplings.

Qijun Luo, Zhili Liu, Lanqing Hong et al.

Semi-supervised domain adaptation (SSDA), which aims to learn models in a partially labeled target domain with the assistance of the fully labeled source domain, attracts increasing attention in recent years. To explicitly leverage the labeled data in both domains, we naturally introduce a conditional GAN framework to transfer images without changing the semantics in SSDA. However, we identify a label-domination problem in such an approach. In fact, the generator tends to overlook the input source image and only memorizes prototypes of each class, which results in unsatisfactory adaptation performance. To this end, we propose a simple yet effective Relaxed conditional GAN (Relaxed cGAN) framework. Specifically, we feed the image without its label to our generator. In this way, the generator has to infer the semantic information of input data. We formally prove that its equilibrium is desirable and empirically validate its practical convergence and effectiveness in image transfer. Additionally, we propose several techniques to make use of unlabeled data in the target domain, enhancing the model in SSDA settings. We validate our method on the well-adopted datasets: Digits, DomainNet, and Office-Home. We achieve state-of-the-art performance on DomainNet, Office-Home and most digit benchmarks in low-resource and high-resource settings.

Xiao-Yun Zhou, Jiacheng Sun, Nanyang Ye et al.

Deep Convolutional Neural Networks (DCNNs) are hard and time-consuming to train. Normalization is one of the effective solutions. Among previous normalization methods, Batch Normalization (BN) performs well at medium and large batch sizes and is with good generalizability to multiple vision tasks, while its performance degrades significantly at small batch sizes. In this paper, we find that BN saturates at extreme large batch sizes, i.e., 128 images per worker, i.e., GPU, as well and propose that the degradation/saturation of BN at small/extreme large batch sizes is caused by noisy/confused statistic calculation. Hence without adding new trainable parameters, using multiple-layer or multi-iteration information, or introducing extra computation, Batch Group Normalization (BGN) is proposed to solve the noisy/confused statistic calculation of BN at small/extreme large batch sizes with introducing the channel, height and width dimension to compensate. The group technique in Group Normalization (GN) is used and a hyper-parameter G is used to control the number of feature instances used for statistic calculation, hence to offer neither noisy nor confused statistic for different batch sizes. We empirically demonstrate that BGN consistently outperforms BN, Instance Normalization (IN), Layer Normalization (LN), GN, and Positional Normalization (PN), across a wide spectrum of vision tasks, including image classification, Neural Architecture Search (NAS), adversarial learning, Few Shot Learning (FSL) and Unsupervised Domain Adaptation (UDA), indicating its good performance, robust stability to batch size and wide generalizability. For example, for training ResNet-50 on ImageNet with a batch size of 2, BN achieves Top1 accuracy of 66.512% while BGN achieves 76.096% with notable improvement.