arXiv subjects

Mingzhen Li

Publications and source records attributed to Mingzhen Li.

At least 19 recordsLinked to original sources

arXiv2026

Scalable High-Fidelity Macromolecular Docking for GPU-Accelerated Supercomputers

Flexible macromolecular docking offers high-fidelity predictions of biomolecular interactions, but remains prohibitively expensive at scale. Among existing approaches, LightDock leverages Glowworm Swarm Optimization (GSO) for accuracy, yet suffers from limited parallelism, irregular computation, and severe load imbalance, preventing efficient execution on GPU supercomputers. We present SparkleDock, a scalable GSO-based docking framework enabling near-real-time flexible docking. We redesign GSO to expose massive fine-grained parallelism at the glowworm-agent level, and restructure the dominant energy scoring computation into a Tensor Core-compatible formulation, enabling efficient execution of irregular pairwise interactions through structured matrix operations. We further introduce a performance-model-driven scheduling for load balancing and out-of-core scaling across GPUs. SparkleDock achieves 9.7 $\times$ and 18.9 $\times$ speedups over LightDock on single A100 and H100 GPU, and delivers over two orders of magnitude acceleration at scale. On 512 GPUs, it reduces docking time from hours to seconds, enabling large-scale, high-fidelity virtual screening previously impractical with flexible docking.

cs.DC↗

arXiv2026

Libra: Taming Attention Workload Skew in Long-Context LLM Training with Bounded Sequence Pool

Long-context LLM training suffers from a load-balancing problem that sequence packing does not solve. Packing samples into fixed-token sequences balances memory and linear-cost operators, but the dominant attention cost scales with the sum of squared sequence lengths. Thus, equally sized packed sequences drawn from a long-tailed corpus can carry substantially different attention workloads, creating data-parallel stragglers and pipeline bubbles. Existing approaches either balance at the granularity of sequences or microbatches, where an outlier can dominate an assignment, or disaggregate attention over a global worker pool whose communication domain grows with the data-parallel (DP) degree. We present Libra, which operationalizes the law of large numbers (LLN) as a scaling principle for load balancing: the attention-balancing pool need not grow with the DP degree. Libra groups packed sequences and their CP groups into fixed-size sequence pools. As DP scales out, Libra adds pools rather than enlarging each one, bounding every attention exchange. Variance-Reduced Sequence Placement makes this effective for finite, long-tailed workloads by co-locating sequences with complementary attention workloads to reduce residual inter-pool skew. Within each pool, Tiled Attention Pooling dispatches sequence-head SH-Tiles across GPUs, while a pipelined runtime overlaps tile exchange with attention. Libra exposes a drop-in context-parallel attention operator and a pluggable data sampler, requiring no changes to model layers, optimizers, or pipeline schedules. On Qwen3-Turbo training with 256K- and 1M-token workloads, Libra improves end-to-end throughput by up to 2.54x over Ulysses, with up to 3.14x worst-step straggler-attention speedup in microbenchmarks. Libra has run for hundreds of thousands of GPU-hours in production on jobs spanning 32K to 1M tokens while preserving training semantics.

cs.DC↗

arXiv2026

Direct Model State Migration for Elastic Training of Large Language Models

Large language model (LLM) training shall adapt to dynamic resources in shared clusters to tackle the elasticity, including passive preemption and optimistic scaling. State migration across device sets is required when altering the hybrid-parallel configuration due to dynamic resources. Existing solutions rely on checkpoint-based mechanisms, which persist complete states to storage for resuming with re-assigned resources, forcing all GPUs to stall when transferring model states. Despite optimization efforts, checkpoint-based solutions incur prohibitive latency due to data movement across memory hierarchies. We propose ETC, a checkpoint-free state migration framework for elastic hybrid-parallel LLM training. We exploits the state locality to minimize inter-GPU data movement, replacing storage persistence with direct peer-to-peer communication. Besides, we eliminate node fragmentation through communication coalescing. Integrated with Megatron-LM, ETC reduces migration overhead by 2.33$\times$ to 6.37$\times$ compared to checkpoint-based solutions across diverse parallel configurations. By enabling efficient migration, ETC unlocks practical elastic training in production environments.

cs.DC↗

arXiv2026

JanusPipe: Efficient Pipeline Parallel Training for Machine Learning Interatomic Potentials

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.

cs.DC↗

arXiv2026

MetaSR: Content-Adaptive Metadata Orchestration for Generative Super-Resolution

We study generative super-resolution (SR) in real-world scenarios where content and degradations vary across domains, genres, and segments. For example, images and videos may alternate between text overlays, fast motion, smooth cartoons, and low-light faces, each benefiting from different forms of side information. Existing metadata-guided SR methods typically use a fixed conditioning design, which is suboptimal when useful cues are content dependent and transmission budgets are limited. We propose MetaSR, a Diffusion Transformer (DiT)-based framework that selects and injects task-relevant metadata to guide SR under resource constraints. Specifically, we use the DiT's own VAE and transformer backbone to fuse heterogeneous metadata, and adopt an efficient distillation strategy that enables one-step diffusion inference. Experiments across diverse content buckets and degradation regimes show that MetaSR outperforms reference solutions by up to 1.0~dB PSNR while achieving up to 50\% transmission bitrate saving at matched quality. We assess these gains under a rate--distortion optimization (RDO) framework that jointly accounts for sender-side bitrate and receiver/display quality metrics (e.g., PSNR and SSIM).

cs.CV↗

arXiv2026

Breaking the Training Barrier of Billion-Parameter Universal Machine Learning Interatomic Potentials

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.

cs.DC↗

arXiv2026

SparseBalance: Load-Balanced Long Context Training with Dynamic Sparse Attention

While sparse attention mitigates the computational bottleneck of long-context LLM training, its distributed training process exhibits extreme heterogeneity in both \textit{1)} sequence length and \textit{2)} sparsity sensitivity, leading to a severe imbalance problem and sub-optimal model accuracy. Existing algorithms and training frameworks typically focus on single issue, failing to systematically co-optimize these two problems. Therefore, we propose SparseBalance, a novel algorithm-system co-design framework, which exploits the sparsity and sequence heterogeneity to optimize model accuracy and system efficiency jointly. First, we propose workload-aware dynamic sparsity tuning, which employs a bidirectional sparsity adjustment to eliminate stragglers and exploit inherent bubbles for free accuracy. Second, we propose a sparsity-aware batching strategy to achieve coarse-grained balance, which complements dynamic sparsity tuning. Experimental results demonstrate that SparseBalance achieves up to a 1.33$\times$ end-to-end speedup while still improving the long-context capability by 0.46\% on the LongBench benchmark.

cs.LG↗

arXiv2025

Vision-Language Enhanced Foundation Model for Semi-Supervised Medical Image Segmentation

Semi-supervised learning (SSL) has emerged as an efficient paradigm for medical image segmentation, reducing the reliance on extensive expert annotations. Vision-language models (VLMs) have demonstrated strong generalization and few-shot capabilities across diverse visual domains. In this work, we integrate a VLM into a semi-supervised medical image segmentation model by adding a Vision-Language Enhanced Semi-supervised Segmentation Assistant (VESSA) that incorporates foundation-level visual-semantic understanding into SSL frameworks. Our approach consists of two stages. In Stage 1, the VLM-enhanced segmentation foundation model VESSA is trained as a reference-guided segmentation assistant using a template bank containing gold-standard exemplars, simulating learning from limited labeled data. Given an input-template pair, VESSA performs visual feature matching to extract representative semantic and spatial cues from exemplar segmentations, generating structured prompts for a Segment Anything Model (SAM)-inspired mask decoder to produce segmentation masks. In Stage 2, VESSA is integrated into an SSL framework as a plug-and-play teacher, providing template-guided pseudo-labels that complement the task-specific student model and strengthen supervision under scarce annotations. Extensive experiments across multiple segmentation datasets and domains show that VESSA-augmented SSL significantly enhances segmentation accuracy, outperforming state-of-the-art baselines under extremely limited annotation conditions.

cs.CV↗

arXiv2025

Large-scale Neural Network Quantum States for ab initio Quantum Chemistry Simulations on Fugaku

Solving quantum many-body problems is one of the fundamental challenges in quantum chemistry. While neural network quantum states (NQS) have emerged as a promising computational tool, its training process incurs exponentially growing computational demands, becoming prohibitively expensive for large-scale molecular systems and creating fundamental scalability barriers for real-world applications. To address above challenges, we present \ours, a high-performance NQS training framework for \textit{ab initio} electronic structure calculations. First, we propose a scalable sampling parallelism strategy with multi-layers workload division and hybrid sampling scheme, which break the scalability barriers for large-scale NQS training. Then, we introduce multi-level parallelism local energy parallelism, enabling more efficient local energy computation. Last, we employ cache-centric optimization for transformer-based \textit{ansatz} and incorporate it with sampling parallelism strategy, which further speedup up the NQS training and achieve stable memory footprint at scale. Experiments demonstrate that \ours accelerate NQS training with up to 8.41x speedup and attains a parallel efficiency up to 95.8\% when scaling to 1,536 nodes.

cs.DC↗

arXiv2025

Skrull: Towards Efficient Long Context Fine-tuning through Dynamic Data Scheduling

Long-context supervised fine-tuning (Long-SFT) plays a vital role in enhancing the performance of large language models (LLMs) on long-context tasks. To smoothly adapt LLMs to long-context scenarios, this process typically entails training on mixed datasets containing both long and short sequences. However, this heterogeneous sequence length distribution poses significant challenges for existing training systems, as they fail to simultaneously achieve high training efficiency for both long and short sequences, resulting in sub-optimal end-to-end system performance in Long-SFT. In this paper, we present a novel perspective on data scheduling to address the challenges posed by the heterogeneous data distributions in Long-SFT. We propose Skrull, a dynamic data scheduler specifically designed for efficient long-SFT. Through dynamic data scheduling, Skrull balances the computation requirements of long and short sequences, improving overall training efficiency. Furthermore, we formulate the scheduling process as a joint optimization problem and thoroughly analyze the trade-offs involved. Based on those analysis, Skrull employs a lightweight scheduling algorithm to achieve near-zero cost online scheduling in Long-SFT. Finally, we implement Skrull upon DeepSpeed, a state-of-the-art distributed training system for LLMs. Experimental results demonstrate that Skrull outperforms DeepSpeed by 3.76x on average (up to 7.54x) in real-world long-SFT scenarios.

cs.LG↗

arXiv2025

Scaling Neural-Network-Based Molecular Dynamics with Long-Range Electrostatic Interactions to 51 Nanoseconds per Day

Neural network-based molecular dynamics (NNMD) simulations incorporating long-range electrostatic interactions have significantly extended the applicability to heterogeneous and ionic systems, enabling effective modeling critical physical phenomena such as protein folding and dipolar surface and maintaining ab initio accuracy. However, neural network inference and long-range force computation remain the major bottlenecks, severely limiting simulation speed. In this paper, we target DPLR, a state-of-the-art NNMD package that supports long-range electrostatics, and propose a set of comprehensive optimizations to enhance computational efficiency. We introduce (1) a hardware-offloaded FFT method to reduce the communication overhead; (2) an overlapping strategy that hides long-range force computations using a single core per node, and (3) a ring-based load balancing method that enables atom-level task evenly redistribution with minimal communication overhead. Experimental results on the Fugaku supercomputer show that our work achieves a 37x performance improvement, reaching a maximum simulation speed of 51 ns/day.

cs.DC↗

arXiv2025

Efficient Long Context Fine-tuning with Chunk Flow

Long context fine-tuning of large language models(LLMs) involves training on datasets that are predominantly composed of short sequences and a small proportion of longer sequences. However, existing approaches overlook this long-tail distribution and employ training strategies designed specifically for long sequences. Moreover, these approaches also fail to address the challenges posed by variable sequence lengths during distributed training, such as load imbalance in data parallelism and severe pipeline bubbles in pipeline parallelism. These issues lead to suboptimal training performance and poor GPU resource utilization. To tackle these problems, we propose a chunk-centric training method named ChunkFlow. ChunkFlow reorganizes input sequences into uniformly sized chunks by consolidating short sequences and splitting longer ones. This approach achieves optimal computational efficiency and balance among training inputs. Additionally, ChunkFlow incorporates a state-aware chunk scheduling mechanism to ensure that the peak memory usage during training is primarily determined by the chunk size rather than the maximum sequence length in the dataset. Integrating this scheduling mechanism with existing pipeline scheduling algorithms further enhances the performance of distributed training. Experimental results demonstrate that, compared with Megatron-LM, ChunkFlow can be up to 4.53x faster in the long context fine-tuning of LLMs. Furthermore, we believe that ChunkFlow serves as an effective solution for a broader range of scenarios, such as long context continual pre-training, where datasets contain variable-length sequences.

cs.DC↗

arXiv2024

Scaling Molecular Dynamics with ab initio Accuracy to 149 Nanoseconds per Day

Physical phenomena such as chemical reactions, bond breaking, and phase transition require molecular dynamics (MD) simulation with ab initio accuracy ranging from milliseconds to microseconds. However, previous state-of-the-art neural network based MD packages such as DeePMD-kit can only reach 4.7 nanoseconds per day on the Fugaku supercomputer. In this paper, we present a novel node-based parallelization scheme to reduce communication by 81%, then optimize the computationally intensive kernels with sve-gemm and mixed precision. Finally, we implement intra-node load balance to further improve the scalability. Numerical results on the Fugaku supercomputer show that our work has significantly improved the time-to-solution of the DeePMD-kit by a factor of 31.7x, reaching 149 nanoseconds per day on 12,000 computing nodes. This work has opened the door for millisecond simulation with ab initio accuracy within one week for the first time.

cs.DC↗

arXiv2024

Real-Time Human Action Recognition on Embedded Platforms

With advancements in computer vision and deep learning, video-based human action recognition (HAR) has become practical. However, due to the complexity of the computation pipeline, running HAR on live video streams incurs excessive delays on embedded platforms. This work tackles the real-time performance challenges of HAR with four contributions: 1) an experimental study identifying a standard Optical Flow (OF) extraction technique as the latency bottleneck in a state-of-the-art HAR pipeline, 2) an exploration of the latency-accuracy tradeoff between the standard and deep learning approaches to OF extraction, which highlights the need for a novel, efficient motion feature extractor, 3) the design of Integrated Motion Feature Extractor (IMFE), a novel single-shot neural network architecture for motion feature extraction with drastic improvement in latency, 4) the development of RT-HARE, a real-time HAR system tailored for embedded platforms. Experimental results on an Nvidia Jetson Xavier NX platform demonstrated that RT-HARE realizes real-time HAR at a video frame rate of 30 frames per second while delivering high levels of recognition accuracy.

cs.CV↗

arXiv2023

Vicinal Feature Statistics Augmentation for Federated 3D Medical Volume Segmentation

Federated learning (FL) enables multiple client medical institutes collaboratively train a deep learning (DL) model with privacy protection. However, the performance of FL can be constrained by the limited availability of labeled data in small institutes and the heterogeneous (i.e., non-i.i.d.) data distribution across institutes. Though data augmentation has been a proven technique to boost the generalization capabilities of conventional centralized DL as a "free lunch", its application in FL is largely underexplored. Notably, constrained by costly labeling, 3D medical segmentation generally relies on data augmentation. In this work, we aim to develop a vicinal feature-level data augmentation (VFDA) scheme to efficiently alleviate the local feature shift and facilitate collaborative training for privacy-aware FL segmentation. We take both the inner- and inter-institute divergence into consideration, without the need for cross-institute transfer of raw data or their mixup. Specifically, we exploit the batch-wise feature statistics (e.g., mean and standard deviation) in each institute to abstractly represent the discrepancy of data, and model each feature statistic probabilistically via a Gaussian prototype, with the mean corresponding to the original statistic and the variance quantifying the augmentation scope. From the vicinal risk minimization perspective, novel feature statistics can be drawn from the Gaussian distribution to fulfill augmentation. The variance is explicitly derived by the data bias in each individual institute and the underlying feature statistics characterized by all participating institutes. The added-on VFDA consistently yielded marked improvements over six advanced FL methods on both 3D brain tumor and cardiac segmentation.

eess.IV↗

arXiv2022

Mimose: An Input-Aware Checkpointing Planner for Efficient Training on GPU

Larger deep learning models usually lead to higher model quality with an ever-increasing GPU memory footprint. Although tensor checkpointing techniques have been proposed to enable training under a restricted GPU memory budget, the input tensor dynamics have been unexploited for optimizing performance while reducing GPU memory footprint. Specifically, due to the diverse datasets and subsequent data argumentation, the input tensor size per mini-batch is dynamic during the training process, leading to a changing GPU memory footprint. However, to leverage such input tensor dynamics in checkpointing, there are two challenges to be solved. First, the checkpointing plan needs to be determined during runtime due to the dynamics of input tensors. Second, the checkpointing plan needs to be applied on the fly without significantly deteriorating the performance. In this paper, we propose Mimose, an input-aware tensor checkpointing planner respecting the memory budget while enabling efficient model training on GPU. Mimose builds a lightweight but accurate prediction model of GPU memory usage online, without pre-analyzing the model. It generates a tensor checkpointing plan based on per-layer memory prediction and applies it to training progress on the fly. It also adopts a caching strategy to avoid having to regenerate the plan for repeated input size. Our experiments show that Mimose achieves superior training throughput compared to state-of-the-art memory planners under the same GPU memory budgets.

cs.DC↗

arXiv2022

EasyScale: Accuracy-consistent Elastic Training for Deep Learning

Distributed synchronized GPU training is commonly used for deep learning. The resource constraint of using a fixed number of GPUs makes large-scale training jobs suffer from long queuing time for resource allocation, and lowers the cluster utilization. Adapting to resource elasticity can alleviate this but often introduces inconsistent model accuracy, due to lacking of capability to decouple model training procedure from resource allocation. We propose EasyScale, an elastic training system that achieves consistent model accuracy under resource elasticity for both homogeneous and heterogeneous GPUs. EasyScale preserves the data-parallel training behaviors strictly, traces the consistency-relevant factors carefully, utilizes the deep learning characteristics for EasyScaleThread abstraction and fast context-switching. To utilize heterogeneous cluster, EasyScale dynamically assigns workers based on the intra-/inter-job schedulers, minimizing load imbalance and maximizing aggregated job throughput. Deployed in an online serving cluster, EasyScale powers the training jobs to utilize idle GPUs opportunistically, improving overall cluster utilization by 62.1%.

cs.DC↗

arXiv2022

swTVM: Towards Optimized Tensor Code Generation for Deep Learning on Sunway Many-Core Processor

The flourish of deep learning frameworks and hardware platforms has been demanding an efficient compiler that can shield the diversity in both software and hardware in order to provide application portability. Among the existing deep learning compilers, TVM is well known for its efficiency in code generation and optimization across diverse hardware devices. In the meanwhile, the Sunway many-core processor renders itself as a competitive candidate for its attractive computational power in both scientific computing and deep learning workloads. This paper combines the trends in these two directions. Specifically, we propose swTVM that extends the original TVM to support ahead-of-time compilation for architecture requiring cross-compilation such as Sunway. In addition, we leverage the architecture features during the compilation such as core group for massive parallelism, DMA for high bandwidth memory transfer and local device memory for data locality, in order to generate efficient codes for deep learning workloads on Sunway. The experiment results show that the codes generated by swTVM achieves 1.79x on average compared to the state-of-the-art deep learning framework on Sunway, across six representative benchmarks. This work is the first attempt from the compiler perspective to bridge the gap of deep learning and Sunway processor particularly with productivity and efficiency in mind. We believe this work will encourage more people to embrace the power of deep learning and Sunway many-core processor.

cs.LG↗