Guangming Tan

Jinwu Yang, Jiaan Wu, Zedong Liu et al.

The rapid scaling of Large Language Models presents significant challenges for their deployment and inference, particularly on resource-constrained specialized AI hardware accelerators such as Huawei's Ascend NPUs, where weight data transfer has become a critical performance bottleneck. While lossless compression can preserve model accuracy and reduce data volume, existing lossless compression algorithms exhibit extremely low throughput when ported to the Ascend NPU architecture. In this paper, we propose ENEC, a novel lossless compression method specifically customized for AI model weights and optimized for Ascend Neural Processing Units. ENEC adopts a block-based fixed-length encoding scheme and incorporates a series of NPU-specific optimizations: bit-width quantization with hierarchical halving bit-packing, vectorized branch-free integer transformation, and dependency-decoupled intra-segment scan for efficient prefix-sum computation. Experimental results demonstrate that ENEC outperforms existing state-of-the-art NPU compressors in both compression ratio and throughput. Compared to leading GPU solutions, ENEC achieves a 3.43X higher throughput than DietGPU and a 1.12X better compression ratio than nvCOMP. By reducing weight transmission overhead, ENEC significantly improves end-to-end inference performance, achieving up to a 6.3X speedup. On Ascend NPUs, ENEC is the first open-source lossless compression algorithm for model weights that achieves performance comparable to state-of-the-art GPU compressors, offering an effective solution for deploying large-scale AI models.

Yida Gu, Fakang Wang, Jianhao Fu et al.

As training scales grow, collective communication libraries (CCL) increasingly face anomalies arising from complex interactions among hardware, software, and environmental factors. These anomalies typically manifest as slow/hang communication, the most frequent and time-consuming category to diagnose. However, traditional diagnostic methods remain inaccurate and inefficient, frequently requiring hours or even days for root cause analysis. To address this, we propose CCL-D, a high-precision diagnostic system designed to detect and locate slow/hang anomalies in large-scale distributed training. CCL-D integrates a rank-level real-time probe with an intelligent decision analyzer. The probe measures cross-layer anomaly metrics using a lightweight distributed tracing framework to monitor communication traffic. The analyzer performs automated anomaly detection and root-cause location, precisely identifying the faulty GPU rank. Deployed on a 4,000-GPU cluster over one year, CCL-D achieved near-complete coverage of known slow/hang anomalies and pinpointed affected ranks within 6 minutes-substantially outperforming existing solutions.

Hongyu Wang, Weijian Liu, Hongtao Xu et al.

Discovering atom-level phenomena requires molecular dynamics (MD) simulations with ab initio accuracy. Machine learning interatomic potentials (MLIPs) enable stable, high-accuracy MD simulations, and their models exhibit scaling-law trends similar to large language models. However, the lack of scalable and efficient distributed training systems for conservative MLIPs makes them difficult to scale. This is because conservative MLIPs inherently follow a double-backward execution pattern, which involves computing gradients during the forward pass. This pattern creates a mismatch with existing distributed training systems, especially for pipeline parallelism. Therefore, we present JanusPipe, an efficient 3D-parallel (PP/DP/GP) training system tailored for conservative MLIPs. It integrates SymFold to enable memory-efficient pipeline parallelism for conservative MLIPs, and WaveK to reduce pipeline bubbles by balancing the four-phase compute time. Experimental results on 32 GPUs show that JanusPipe improves throughput by $1.51\times$ and $1.45\times$ on average over 1F1B and Hanayo, respectively.

Yuanchang Zhou, Hongyu Wang, Yiming Du et al.

Universal Machine Learning Interatomic Potentials (uMLIPs), pre-trained on massively diverse datasets encompassing inorganic materials and organic molecules across the entire periodic table, serve as foundational models for quantum-accurate physical simulations. However, uMLIP training requires second-order derivatives, which lack corresponding parallel training frameworks; moreover, scaling to the billion-parameter regime causes explosive growth in computation and communication overhead, making its training a tremendous challenge. We introduce MatRIS-MoE, a billion-parameter Mixture-of-Experts model built upon invariant architecture, and {Janus}, a pioneering high-dimensional distributed training framework for uMLIPs with hardware-aware optimizations. Deployed across two Exascale supercomputers, our code attains a peak performance of 1.2/1.0 EFLOPS (24\%/{35.5\%} of theoretical peak) in single precision at over 90\% parallel efficiency, compressing the training of billion-parameter uMLIPs from weeks to hours. This work establishes a new high-water mark for AI-for-Science (AI4S) foundation models at Exascale and provides essential infrastructure for rapid scientific discovery.

Daran Sun, Bowen Kan, Haoquan Long et al.

AI-driven methods have demonstrated considerable success in tackling the central challenge of accurately solving the Schrödinger equation for complex many-body systems. Among neural network quantum state (NNQS) approaches, the NNQS-SCI (Selected Configuration Interaction) method stands out as a state-of-the-art technique, recognized for its high accuracy and scalability. However, its application to larger systems is severely constrained by a hybrid CPU-GPU architecture. Specifically, centralized CPU-based global de-duplication creates a severe scalability barrier due to communication bottlenecks, while host-resident coupled-configuration generation induces prohibitive computational overheads. We introduce cuNNQS-SCI, a fully GPU-accelerated SCI framework designed to overcome these bottlenecks. cuNNQS-SCI first integrates a distributed, load-balanced global de-duplication algorithm to minimize redundancy and communication overhead at scale. To address compute limitations, it employs specialized, fine-grained CUDA kernels for exact coupled configuration generation. Finally, to break the single-GPU memory barrier exposed by this full acceleration, it incorporates a GPU memory-centric runtime featuring GPU-side pooling, streaming mini-batches, and overlapped offloading. This design enables much larger configuration spaces and shifts the bottleneck from host-side limitations back to on-device inference. Our evaluation demonstrates that cuNNQS-SCI fundamentally expands the scale of solvable problems. On an NVIDIA A100 cluster with 64 GPUs, cuNNQS-SCI achieves up to 2.32X end-to-end speedup over the highly-optimized NNQS-SCI baseline while preserving the same chemical accuracy. Furthermore, it demonstrates excellent distributed performance, maintaining over 90% parallel efficiency in strong scaling tests.

Zedong Liu, Xinyang Ma, Dejun Luo et al.

LLMs are widely adopted in production, pushing inference systems to their limits. Disaggregated LLM serving (e.g., PD separation and KV state disaggregation) improves scalability and cost efficiency, but it also turns KV into an explicit payload crossing network and storage boundaries, making KV a dominant end-to-end bottleneck. Existing KV compression are typically static runtime configurations, despite production service context varies over time in workload mix, bandwidth, and SLO/quality budgets. As a result, a fixed choice can be suboptimal or even increase latency. We present \emph{KVServe}, the first service-aware and adaptive KV communication compression framework for disaggregated LLM serving: KVServe (1) unifies KV compression into a modular strategy space with new components and cross-method recomposition; (2) introduces Bayesian Profiling Engine that efficiently searches this space and distills a 3D Pareto candidate set, reducing $50\times$ offline search overhead; and (3) deploys a Service-Aware Online Controller that combines an analytical latency model with a lightweight bandit to select profiles under constraints and correct offline-to-online mismatch. Integrated into vLLM and evaluated across datasets, models, GPUs and networks, KVServe achieves up to $9.13\times$ JCT speedup in PD-separated serving and up to $32.8\times$ TTFT reduction in KV-disaggregated serving.

Yuanchang Zhou, Siyu Hu, Xiangyu Zhang et al.

Foundation MLIPs demonstrate broad applicability across diverse material systems and have emerged as a powerful and transformative paradigm in chemical and computational materials science. Equivariant MLIPs achieve state-of-the-art accuracy in a wide range of benchmarks by incorporating equivariant inductive bias. However, the reliance on tensor products and high-degree representations makes them computationally costly. This raises a fundamental question: as quantum mechanical-based datasets continue to expand, can we develop a more compact model to thoroughly exploit high-dimensional atomic interactions? In this work, we present MatRIS (\textbf{Mat}erials \textbf{R}epresentation and \textbf{I}nteraction \textbf{S}imulation), an invariant MLIP that introduces attention-based modeling of three-body interactions. MatRIS leverages a novel separable attention mechanism with linear complexity $O(N)$, enabling both scalability and expressiveness. MatRIS delivers accuracy comparable to that of leading equivariant models on a wide range of popular benchmarks (Matbench-Discovery, MatPES, MDR phonon, Molecular dataset, etc). Taking Matbench-Discovery as an example, MatRIS achieves an F1 score of up to 0.847 and attains comparable accuracy at a lower training cost. The work indicates that our carefully designed invariant models can match or exceed the accuracy of equivariant models at a fraction of the cost, shedding light on the development of accurate and efficient MLIPs.

Chen Wang, Siyu Hu, Guangming Tan et al.

SO(3) equivariant graph neural networks have become the dominant paradigm for atomistic foundation models, achieving high accuracy and data efficiency by building rotational symmetry directly into the architecture. Yet the computational cost of their higher-order tensor operations creates a tough trade-off between model accuracy and inference efficiency. In this paper, we propose a structural pruning method for SO(3) equivariant atomistic foundation models to bridge this accuracy-efficiency gap. The pruning is applied along the channel and order dimensions, with each irreducible representation kept or removed as a complete block, thereby retaining SO(3) equivariance. Starting from a large checkpoint, the pruned model substantially reduces the inference cost while retaining higher accuracy than an independently trained small model. The pruned MACE-MP model outperforms the official from-scratch trained small model on 7 of 9 metrics on the Matbench Discovery leaderboard. In terms of efficiency, compressed MACE-MP and MACE-OFF models contain 1.5$\times$ to 4$\times$ fewer parameters and require 2.5$\times$ to 4$\times$ less pre-training compute than training a small model from scratch. For downstream applications, fine-tuning the pruned model reduces energy and force errors by 70.1% and 34.4% compared to training task-specific models from scratch across eight representative downstream datasets. We demonstrate that the method generalizes to other SO(3) equivariant architectures (SevenNet, eSCN) and can be combined with quantization and knowledge distillation for further gains.

Shiyun Zhang, Yibo Yao, Haoquan Long et al.

The classical material tetrahedron that represents the Structure-Property-Processing-Performance-Characterization relationship is the most important research paradigm in materials science so far. It has served as a protocol to guide experiments, modeling, and theory to uncover hidden relationships between various aspects of a certain material. This substantially facilitates knowledge accumulation and material discovery with desired functionalities to realize versatile applications. In recent years, with the advent of artificial intelligence (AI) techniques, the attention of AI towards scientific research is soaring. The trials of implementing AI in various disciplines are endless, with great potential to revolutionize the research diagram. Despite the success in natural language processing and computer vision, how to effectively integrate AI with natural science is still a grand challenge, bearing in mind their fundamental differences. Inspired by these observations and limitations, we delve into the current research paradigm dictated by the classical material tetrahedron and propose two new paradigms to stimulate data-driven and AI-augmented research. One tetrahedron focuses on AI for materials science by considering the Matter-Data-Model-Potential-Agent diagram. The other demonstrates AI research by discussing Data-Architecture-Encoding-Optimization-Inference relationships. The crucial ingredients of these frameworks and their connections are discussed, which will likely motivate both scientific thinking refinement and technology advancement. Despite the widespread enthusiasm for chasing AI for science, we must analyze issues rationally to come up with well-defined, resolvable scientific problems in order to better master the power of AI.

Tong Zhao, Jiacheng Li, Yuanchang Zhou et al.

Finding lower and better-generalizing minima is crucial for deep learning. However, most existing optimizers stop searching the parameter space once they reach a local minimum. Given the complex geometric properties of the loss landscape, it is difficult to guarantee that such a point is the lowest or provides the best generalization. To address this, we propose an adaptor "E" for gradient-based optimizers. The adapted optimizer tends to continue exploring along landscape valleys (areas with low and nearly identical losses) in order to search for potentially better local minima even after reaching a local minimum. This approach increases the likelihood of finding a lower and flatter local minimum, which is often associated with better generalization. We also provide a proof of convergence for the adapted optimizers in both convex and non-convex scenarios for completeness. Finally, we demonstrate their effectiveness in an important but notoriously difficult training scenario, large-batch training, where Lamb is the benchmark optimizer. Our testing results show that the adapted Lamb, ALTO, increases the test accuracy (generalization) of the current state-of-the-art optimizer by an average of 2.5% across a variety of large-batch training tasks. This work potentially opens a new research direction in the design of optimization algorithms.

Shaofeng Yang, Yunting Wang, Yingying Cheng et al.

The solution of sparse linear systems constitutes the dominant computational bottleneck in interior point methods (IPMs), frequently consuming over 70\% of the total solution time. As optimization problems scale to millions of variables, direct solvers encounter prohibitive fill-in, excessive memory consumption, and limited parallel scalability. We present SDSL-Solver, a scalable distributed sparse linear solver framework designed for IPMs. SDSL-Solver employs Krylov subspace methods, combined with numerics-based sparse filtering and diagonal correction techniques that produce high-quality preconditioners. To accommodate diverse problem characteristics, SDSL-Solver offers two complementary distributed parallel methods: Block Jacobi for well-conditioned, diagonally dominant systems, and Bordered Block Diagonal (BBD) for ill-conditioned problems requiring globally coupled preconditioning via Schur complement techniques. A preconditioner reuse strategy further amortizes construction costs across consecutive IPMs iterations. We evaluate SDSL-Solver on benchmark problems with matrix dimensions ranging from tens of thousands to over five million on multi-node clusters equipped with X86 processors. The experimental results show that under the Block Jacobi and BBD distributed methods, SDSL-Solver on a four-node configuration achieves average speedups of $6.23\times$ and $7.77\times$, respectively, compared to PETSc running on the same number of nodes. Relative to the single-node PARDISO, the average speedups reach $97.54\times$ and $5.85\times$, respectively.

Man Liu, Xingchen Liu, Xingjian Tian et al.

Handling communication overhead in large-scale tensor-parallel training remains a critical challenge due to the dense, near-zero distributions of intermediate tensors, which exacerbate errors under frequent communication and introduce significant computational overhead during compression. To this end, we propose TACO (Tensor-parallel Adaptive COmmunication compression), a robust FP8-based framework for compressing TP intermediate tensors. First, we employ a data-driven reshaping strategy combined with an Adaptive Scale-Hadamard Transform to enable high-fidelity FP8 quantization, while its Dual-Scale Quantization mechanism ensures numerical stability throughout training. Second, we design a highly fused compression operator to reduce memory traffic and kernel launch overhead, allowing efficient overlap with communication. Finally, we integrate TACO with existing state-of-the-art methods for Data and Pipeline Parallelism to develop a compression-enabled 3D-parallel training framework. Detailed experiments on GPT models and Qwen model demonstrate up to 1.87X end-to-end throughput improvement while maintaining near-lossless accuracy, validating the effectiveness and efficiency of TACO in large-scale training.

Yuanchang Zhou, Siyu Hu, Chen Wang et al.

Graph neural network universal interatomic potentials (GNN-UIPs) have demonstrated remarkable generalization and transfer capabilities in material discovery and property prediction. These models can accelerate molecular dynamics (MD) simulation by several orders of magnitude while maintaining \textit{ab initio} accuracy, making them a promising new paradigm in material simulations. One notable example is Crystal Hamiltonian Graph Neural Network (CHGNet), pretrained on the energies, forces, stresses, and magnetic moments from the MPtrj dataset, representing a state-of-the-art GNN-UIP model for charge-informed MD simulations. However, training the CHGNet model is time-consuming(8.3 days on one A100 GPU) for three reasons: (i) requiring multi-layer propagation to reach more distant atom information, (ii) requiring second-order derivatives calculation to finish weights updating and (iii) the implementation of reference CHGNet does not fully leverage the computational capabilities. This paper introduces FastCHGNet, an optimized CHGNet, with three contributions: Firstly, we design innovative Force/Stress Readout modules to decompose Force/Stress prediction. Secondly, we adopt massive optimizations such as kernel fusion, redundancy bypass, etc, to exploit GPU computation power sufficiently. Finally, we extend CHGNet to support multiple GPUs and propose a load-balancing technique to enhance GPU utilization. Numerical results show that FastCHGNet reduces memory footprint by a factor of 3.59. The final training time of FastCHGNet can be decreased to \textbf{1.53 hours} on 32 GPUs without sacrificing model accuracy.

Jiacheng Li, Jianchao Tan, Zhidong Yang et al.

Transformer architecture gradually dominates the LLM field. Recent advances in training optimization for Transformer-based large language models (LLMs) primarily focus on architectural modifications or optimizer adjustments. However, these approaches lack systematic optimization of weight patterns during training. Weight pattern refers to the distribution and relative magnitudes of weight parameters in a neural network. To address this issue, we propose a Weight Scaling method called WISCA to enhance training efficiency and model quality by strategically improving neural network weight patterns without changing network structures. By rescaling weights while preserving model outputs, WISCA indirectly optimizes the model's training trajectory. Experiments demonstrate that WISCA significantly improves convergence quality (measured by generalization capability and loss reduction), particularly in LLMs with Grouped Query Attention (GQA) architectures and LoRA fine-tuning tasks. Empirical results show 5.6% average improvement on zero-shot validation tasks and 2.12% average reduction in training perplexity across multiple architectures.

Zedong Liu, Shenggan Cheng, Guangming Tan et al.

Multimodal large language models (MLLMs) extend LLMs to handle images, videos, and audio by incorporating feature extractors and projection modules. However, these additional components -- combined with complex inference pipelines and heterogeneous workloads -- introduce significant inference overhead. Therefore, efficiently serving MLLMs remains a major challenge. Current tightly coupled serving architectures struggle to distinguish between mixed request types or adapt parallelism strategies to different inference stages, leading to increased time-to-first-token (TTFT) latency and poor resource utilization. To address this, we introduce Elastic Multimodal Parallelism (EMP), a new serving paradigm that elastically adapts to resource heterogeneity across request types and inference stages. Building upon EMP, we develop ElasticMM, an MLLM serving system that (1) separates requests into independent modality groups with dynamic resource allocation via a modality-aware load balancer; (2) decouples inference stages and enables parallelism adjustment and adaptive scaling via elastic partition scheduling; and (3) improves inference efficiency through unified multimodal prefix caching and non-blocking encoding. Experiments on diverse real-world datasets show that ElasticMM outperforms state-of-the-art (SOTA) serving systems, reducing TTFT by up to 4.2x and achieving 3.2-4.5x higher throughput while meeting service-level objectives (SLOs).

Xiaoyang Zhang, Junmin Xiao, Guangming Tan

Convolution is the most time-consuming part in the computation of convolutional neural networks (CNNs), which have achieved great successes in numerous applications. Due to the complex data dependency and the increase in the amount of model samples, the convolution suffers from high overhead on data movement (i.e., memory access). This work provides comprehensive analysis and methodologies to minimize the communication for the convolution in CNNs. With an in-depth analysis of the recent I/O complexity theory under the red-blue game model, we develop a general I/O lower bound theory for a composite algorithm which consists of several different sub-computations. Based on the proposed theory, we establish the data movement lower bound results of two representative convolution algorithms in CNNs, namely the direct convolution and Winograd algorithm. Next, derived from I/O lower bound results, we design the near I/O-optimal dataflow strategies for the two main convolution algorithms by fully exploiting the data reuse. Furthermore, in order to push the envelope of performance of the near I/O-optimal dataflow strategies further, an aggressive design of auto-tuning based on I/O lower bounds, is proposed to search an optimal parameter configuration for the direct convolution and Winograd algorithm on GPU, such as the number of threads and the size of shared memory used in each thread block. Finally, experiment evaluation results on the direct convolution and Winograd algorithm show that our dataflow strategies with the auto-tuning approach can achieve about 3.32x performance speedup on average over cuDNN. In addition, compared with TVM, which represents the state-of-the-art technique for auto-tuning, not only our auto-tuning method based on I/O lower bounds can find the optimal parameter configuration faster, but also our solution has higher performance than the optimal solution provided by TVM.