Siheng Li, Kejiao Li, Zenan Xu, Guanhua Huang, Kun Li, Haoyuan Wu, Wujiajia, Zihao Zheng, Chenchen Zhang, Kun Shi, Xue Gong, Qi Yi, Ruibin Xiong, Tingqiang Xu, Yuhao Jiang, Jianfeng Yan, Yuyuan Zeng, Guanghui Xu, Jinbao Xue, Zhijiang xu, Zheng Fang, Shuai LI, Qibin Liu, Xiaoxue Li, Zhuoyu Li, Yangyu Tao, Fei Gao, Cheng Jiang, Bochao Wang, Kai Liu, Jianchen Zhu, Wai Lam, Bo Zhou, Di Wang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Low-precision training is considered an effective strategy for reducing both training and downstream inference costs. Previous scaling laws for precision mainly focus on integer quantization, which pay less attention to the constituents in floating-point (FP) quantization, and thus cannot well fit the LLM losses in this scenario. In contrast, while FP quantization training is more commonly implemented in production, it's research has been relatively superficial. In this paper, we thoroughly explore the effects of FP quantization targets, exponent bits, mantissa bits, and the calculation granularity of the scaling factor in FP quantization training performance of LLM models. In addition to an accurate FP quantization unified scaling law, we also provide valuable suggestions for the community: (1) Exponent bits contribute slightly more to the model performance than mantissa bits. We provide the optimal exponent-mantissa bit ratio for different bit numbers, which is available for future reference by hardware manufacturers; (2) We discover the formation of the critical data size in low-precision LLM training. Too much training data exceeding the critical data size will inversely bring in degradation of LLM performance; (3) The optimal FP quantization precision is directly proportional to the computational power, but within a wide computational power range. We estimate that the best cost-performance precision should lie between 4-8 bits.

BERT-like models have been widely adopted in text mining and web search due to their high accuracy. However, large BERT-like models suffer from inefficient online inference on GPUs for two main reasons. First, their high accuracy relies on large model depth, which linearly increases sequential computation on GPUs. Second, stochastic and dynamic online workloads lead to extra costs due to batching and padding. To address the problem, we present Student Parallelism for efficient GPU inference of BERT-like models under real-world online workloads. At its core, Student Parallelism adopts stacking distillation and boosting ensemble, distilling the original deep model into a group of shallow but virtually stacked student models running in parallel. This enables Student Parallelism to achieve a low model depth (e.g., two layers), and thus low inference latency while maintaining accuracy. In addition, we design adaptive student pruning to adjust the number of students according to the dynamic online workloads. For example, during workload bursts, it can temporarily decrease the number of students with minimal accuracy loss to improve system throughput. Extensive experiments on real-world datasets and workloads show that Student Parallelism achieves up to 4.1× lower latency while maintaining accuracy and up to 22.27× higher throughput during workload bursts.

To optimize large Transformer model training, both efficient parallel computing and advanced data management are indispensable. However, current methods often assume a stable and uniform training workload, neglecting data-induced imbalances-arising from both sampling and packing processes-which can impede training performance. Specifically, data sampling imbalance arises from uneven sequence length distribution of the training data, while data packing imbalance stems from the discrepancy between the linear memory complexity and quadratic time complexity of the attention mechanism. To address these imbalance issues, we develop Hydraulis, which jointly optimizes the parallel strategies and data assignment. For one thing, we introduce large model training with dynamic heterogeneous parallel strategies in response to the sequence length variations within and across training iterations. For another, we devise a two-stage data assignment approach, which strikes a good balance in terms of the training workloads both within and across model replicas. Empirical results demonstrate that Hydraulis outperforms existing systems by 1.32-2.66×. Our source code is available: https://github.com/PKU-DAIR/Hetu.

Nowadays, Large Language Models (LLMs) have been trained using extended context lengths to foster more creative applications. However, long context training poses great challenges considering the constraint of GPU memory. It not only leads to substantial activation memory consumption during training, but also incurs considerable memory fragmentation. To facilitate long context training, existing frameworks have adopted strategies such as recomputation and various forms of parallelisms. Nevertheless, these techniques rely on redundant computation or extensive communication, resulting in low Model FLOPS Utilization (MFU). In this paper, we propose MEMO, a novel LLM training framework designed for fine-grained activation memory management. Given the quadratic scaling of computation and linear scaling of memory with sequence lengths when using FlashAttention, we offload memory-consuming activations to CPU memory after each layer's forward pass and fetch them during the backward pass. To maximize the swapping of activations without hindering computation, and to avoid exhausting limited CPU memory, we implement a token-wise activation recomputation and swapping mechanism. Furthermore, we tackle the memory fragmentation issue by employing a bi-level Mixed Integer Programming (MIP) approach, optimizing memory reuse across transformer layers. Empirical results demonstrate that MEMO achieves an average of 1.97x and 1.80x MFU compared to Megatron-LM and DeepSpeed, respectively. This improvement is attributed to MEMO's ability to minimize memory fragmentation, reduce recomputation and intensive communication, and circumvent the delays associated with the memory reorganization process due to fragmentation. By leveraging fine-grained activation memory management, MEMO facilitates efficient training of 7B LLM with 1 million sequence length on just 8 A800 GPUs, achieving an MFU of 52.30%.

In current deep learning tasks, Adam-style optimizers—such as Adam, Adagrad, RMSprop, Adafactor, and Lion—have been widely used as alternatives to SGD-style optimizers. These optimizers typically update model parameters using the sign of gradients, resulting in more stable convergence curves. The learning rate and the batch size are the most critical hyperparameters for optimizers, which require careful tuning to enable effective convergence. Previous research has shown that the optimal learning rate increases linearly (or follows similar rules) with batch size for SGD-style optimizers. However, this conclusion is not applicable to Adam-style optimizers. In this paper, we elucidate the connection between optimal learning rates and batch sizes for Adam-style optimizers through both theoretical analysis and extensive experiments. First, we raise the scaling law between batch sizes and optimal learning rates in the “sign of gradient” case, in which we prove that the optimal learning rate first rises and then falls as the batch size increases. Moreover, the peak value of the surge will gradually move toward the larger batch size as training progresses. Second, we conduct experiments on various CV and NLP tasks and verify the correctness of the scaling law.

This work studies the problem of out-of-distribution fluid dynamics modeling. Previous works usually design effective neural operators to learn from mesh-based data structures. However, in real-world applications, they would suffer from distribution shifts from the variance of system parameters and temporal evolution of the dynamical system. In this paper, we propose a novel approach named \underline{P}rompt Evol\underline{u}tion with G\underline{r}aph OD\underline{E} (\method{}) for out-of-distribution fluid dynamics modeling. The core of our \method{} is to learn time-evolving prompts using a graph ODE to adapt spatio-temporal forecasting models to different scenarios. In particular, our \method{} first learns from historical observations and system parameters in the frequency domain to explore multi-view context information, which could effectively initialize prompt embeddings. More importantly, we incorporate the interpolation of observation sequences into a graph ODE, which can capture the temporal evolution of prompt embeddings for model adaptation. These time-evolving prompt embeddings are then incorporated into basic forecasting models to overcome temporal distribution shifts. We also minimize the mutual information between prompt embeddings and observation embeddings to enhance the robustness of our model to different distributions. Extensive experiments on various benchmark datasets validate the superiority of the proposed \method{} in comparison to various baselines.

Recent years have witnessed the unprecedented achievements of large-scale pre-trained models, especially Transformer models. Many products and services in Tencent Inc., such as WeChat, QQ, and Tencent Advertisement, have been opted in to gain the power of pre-trained models. In this work, we present Angel-PTM, a productive deep learning system designed for pre-training and fine-tuning Transformer models. Angel-PTM can train extremely large-scale models with hierarchical memory efficiently. The key designs of Angel-PTM are a fine-grained memory management via the Page abstraction and a unified scheduling method that coordinates computations, data movements, and communications. Furthermore, Angel-PTM supports extreme model scaling with SSD storage and implements a lock-free updating mechanism to address the SSD I/O bottlenecks. Experimental results demonstrate that Angel-PTM outperforms existing systems by up to 114.8% in terms of maximum model scale as well as up to 88.9% in terms of training throughput. Additionally, experiments on GPT3-175B and T5-MoE-1.2T models utilizing hundreds of GPUs verify our strong scalability.