arXiv subjects

Jinhua Zhou

Publications and source records attributed to Jinhua Zhou.

9 recordsLinked to original sources

arXiv2026

psRL: Efficient Training for Agentic AI via Training-Time Prefix Sharing

In modern agentic AI training, the system bottleneck is shifting from rollout to update. Emerging sampling strategies such as tree-structured and step-wise RL greatly increase training sample volume while incurring relatively low marginal rollout cost, causing the update phase to dominate the end-to-end execution time. Crucially, this shift exposes a new optimization opportunity, as production traces reveal substantial prefix redundancy across training samples. In this paper, we propose psRL (prefix sharing for RL), a new training system for agentic AI designed to exploit prefix redundancy among training samples. Leveraging the global visibility and data immutability inherent to the update phase, psRL achieves efficient workload scheduling and memory management for distributed training. Specifically, psRL introduces two novel prefix-sharing mechanisms that enable flexible, fine-grained workload distribution across GPU workers, simultaneously optimizing prefix reuse and achieving load balancing. Moreover, psRL implements a new underlying KV cache manager that facilitates adaptable block-size allocation and dynamic KV caching, maximizing memory utilization while maintaining a high prefix hit rate. Evaluations using production traces demonstrate that psRL outperforms existing systems by up to 5.2x in throughput. The source code will be publicly available soon.

cs.DC↗

arXiv2026

SLAI T-Rex: Full-Parameter Post-training of the DeepSeek-V4 Family on Ascend SuperPOD

Full-parameter post-training of trillion-parameter-scale MoE models introduces substantial system-level challenges for large-scale distributed training, including severe memory pressure, non-overlapped communication overhead, and inefficient kernel execution. While most large-scale LLM training systems are built around GPU-based clusters, this report presents an end-to-end optimization practice on the Ascend NPU SuperPOD. Using the DeepSeek-V4 model family as the target workload, we develop a hierarchical optimization framework spanning model-level parallelism, computation-communication orchestration, and low-level kernel execution. The resulting system achieves 34.22% Model FLOPs Utilization (MFU) with a 2.93x improvement over the open-source baseline recipe while maintaining training stability. Building on this optimized infrastructure, we further establish a CPT and SFT workflow for complex Operations Research (OR) tasks. We refer to the integrated framework as SLAI T-Rex. Using DeepSeek-V4-Flash, we develop OR-oriented CPT and SFT data pipelines that combine collected domain resources with solver-verified synthetic optimization documents. The resulting dataset contains 10K high-quality SFT samples spanning four task categories and three problem representations. The specialized model achieves the highest average zero-shot Pass@1 score among the evaluated models, reaching 71.81% and outperforming GPT-5.4-Mini and the base DeepSeek-V4-Flash model by 3.98 and 11.27 percentage points, respectively. Overall, this work demonstrates a full-stack pathway from efficient trillion-parameter model post-training on Ascend infra to domain-specialized Flash models for solver-grounded mathematical modeling, advancing frontier-model systems for complex reasoning.

cs.CL↗

arXiv2025

High-Throughput LLM inference on Heterogeneous Clusters

Nowadays, many companies possess various types of AI accelerators, forming heterogeneous clusters. Efficiently leveraging these clusters for high-throughput large language model (LLM) inference services can significantly reduce costs and expedite task processing. However, LLM inference on heterogeneous clusters presents two main challenges. Firstly, different deployment configurations can result in vastly different performance. The number of possible configurations is large, and evaluating the effectiveness of a specific setup is complex. Thus, finding an optimal configuration is not an easy task. Secondly, LLM inference instances within a heterogeneous cluster possess varying processing capacities, leading to different processing speeds for handling inference requests. Evaluating these capacities and designing a request scheduling algorithm that fully maximizes the potential of each instance is challenging. In this paper, we propose a high-throughput inference service system on heterogeneous clusters. First, the deployment configuration is optimized by modeling the resource amount and expected throughput and using the exhaustive search method. Second, a novel mechanism is proposed to schedule requests among instances, which fully considers the different processing capabilities of various instances. Extensive experiments show that the proposed scheduler improves throughput by 122.5% and 33.6% on two heterogeneous clusters, respectively.

cs.DC↗

arXiv2024

WindGP: Efficient Graph Partitioning on Heterogenous Machines

Graph Partitioning is widely used in many real-world applications such as fraud detection and social network analysis, in order to enable the distributed graph computing on large graphs. However, existing works fail to balance the computation cost and communication cost on machines with different power (including computing capability, network bandwidth and memory size), as they only consider replication factor and neglect the difference of machines in realistic data centers. In this paper, we propose a general graph partitioning algorithm WindGP, which can support fast and high-quality edge partitioning on heterogeneous machines. WindGP designs novel preprocessing techniques to simplify the metric and balance the computation cost according to the characteristics of graphs and machines. Also, best-first search is proposed instead of BFS and DFS, in order to generate clusters with high cohesion. Furthermore, WindGP adaptively tunes the partition results by sophisticated local search methods. Extensive experiments show that WindGP outperforms all state-of-the-art partition methods by 1.35 - 27 times on both dense and sparse distributed graph algorithms, and has good scalability with graph size and machine number.

cs.DC↗

arXiv2024

FPM-WSI: Fourier ptychographic whole slide imaging via feature-domain backdiffraction

Fourier ptychographic microscopy (FPM), characterized by high-throughput computational imaging, theoretically provides a cunning solution to the trade-off between spatial resolution and field of view (FOV), which has a promising prospect in the application of digital pathology. However, block reconstruction and then stitching has currently become an unavoidable procedure due to vignetting effects. The stitched image tends to present color inconsistency in different image segments, or even stitching artifacts. In response, we reported a computational framework based on feature-domain backdiffraction to realize full-FOV, stitching-free FPM reconstruction. Different from conventional algorithms that establish the loss function in the image domain, our method formulates it in the feature domain, where effective information of images is extracted by a feature extractor to bypass the vignetting effect. The feature-domain error between predicted images based on estimation of model parameters and practically captured images is then digitally diffracted back through the optical system for complex amplitude reconstruction and aberration compensation. Through massive simulations and experiments, the method presents effective elimination of vignetting artifacts, and reduces the requirement of precise knowledge of illumination positions. We also found its great potential to recover the data with a lower overlapping rate of spectrum and to realize automatic blind-digital refocusing without a prior defocus distance.

physics.optics↗

arXiv2023

Pupil-driven quantitative differential phase contrast imaging

In this research, we reveal the inborn but hitherto ignored properties of quantitative differential phase contrast (qDPC) imaging: the phase transfer function being an edge detection filter. Inspired by this, we highlighted the duality of qDPC between optics and pattern recognition, and propose a simple and effective qDPC reconstruction algorithm, termed Pupil-Driven qDPC (pd-qDPC), to facilitate the phase reconstruction quality for the family of qDPC-based phase reconstruction algorithms. We formed a new cost function in which modified L0-norm was used to represent the pupil-driven edge sparsity, and the qDPC convolution operator is duplicated in the data fidelity term to achieve automatic background removal. Further, we developed the iterative reweighted soft-threshold algorithms based on split Bregman method to solve this modified L0-norm problem. We tested pd-qDPC on both simulated and experimental data and compare against state-of-the-art (SOTA) methods including L2-norm, total variation regularization (TV-qDPC), isotropic-qDPC, and Retinex qDPC algorithms. Results show that our proposed model is superior in terms of phase reconstruction quality and implementation efficiency, in which it significantly increases the experimental robustness while maintaining the data fidelity. In general, the pd-qDPC enables the high-quality qDPC reconstruction without any modification of the optical system. It simplifies the system complexity and benefits the qDPC community and beyond including but not limited to cell segmentation and PTF learning based on the edge filtering property.

math.OC↗

arXiv2022

ELFPIE: an error-laxity Fourier ptychographic iterative engine

We present a simple but efficient and robust reconstruction algorithm for Fourier ptychographic microscopy, termed error-laxity Fourier ptychographic iterative engine (Elfpie), that is simultaneously robust to (1) noise signal (including Gaussian, Poisson, and salt & pepper noises), (2) problematic background illumination problem, (3) vignetting effects, and (4) misaligning of LED positions, without the need of calibrating or recovering of these system errors. In Elfpie, we embedded the inverse problem of FPM under the framework of feature extraction/recovering and proposed a new data fidelity cost function regularized by the global second-order total-variation regularization (Hessian regularization). The closed-form complex gradient for the cost function is derived and is back-propagated using the AdaBelief optimizer with an adaptive learning rate to update the entire Fourier spectrum of the sample and system pupil function. The Elfpie is tested on both simulation data and experimental data and is compared against the state-of-the-art (SOTA) algorithm. Results show the superiority of the Elfpie among other SOTA methods, in both reconstruction quality under different degeneration issues, and implementation efficiency. In general, compared against SOTA methods, the Elfpie is robust to Gaussian noise with 100 times larger noise strength, salt & pepper noise with 1000 times larger noise strength, and Poisson noise with 10 times noise strength. The Elfpie is able to reconstruct high-fidelity sample field under large LED position misalignments up to 2 mm. It can also bypass the vignetting effect in which all SOTA methods fail to reconstruct the sample pattern. With parallel computation, the Elfpie is able to be K times faster than traditional FPM, where K is the number of used LEDs.

math.CV↗

arXiv2022

Retinex-qDPC: automatic background rectified quantitative differential phase contrast imaging

The quality of quantitative differential phase contrast reconstruction (qDPC) can be severely degenerated by the mismatch of the background of two oblique illuminated images, yielding problematic phase recovery results. These background mismatches may result from illumination patterns, inhomogeneous media distribution, or other defocusing layers. In previous reports, the background is manually calibrated which is time-consuming, and unstable, since new calibrations are needed if any modification to the optical system was made. It is also impossible to calibrate the background from the defocusing layers, or for high dynamic observation as the background changes over time. To tackle the mismatch of background and increases the experimental robustness, we propose the Retinex-qDPC in which we use the images edge features as data fidelity term yielding L2-Retinex-qDPC and L1-Retinex-qDPC for high background-robustness qDPC reconstruction. The split Bregman method is used to solve the L1-Retinex DPC. We compare both Retinex-qDPC models against state-of-the-art DPC reconstruction algorithms including total-variation regularized qDPC, and isotropic-qDPC using both simulated and experimental data. Results show that the Retinex qDPC can significantly improve the phase recovery quality by suppressing the impact of mismatch background. Within, the L1-Retinex-qDPC is better than L2-Retinex and other state-of-the-art DPC algorithms. In general, the Retinex-qDPC increases the experimental robustness against background illumination without any modification of the optical system, which will benefit all qDPC applications.

eess.IV↗

arXiv2022

High-fidelity quantitative differential phase contrast deconvolution using dark-field sparse prior

Differential phase contrast (DPC) imaging plays an important role in the family of quantitative phase measurement. However, the reconstruction algorithm for quantitative DPC (qDPC) imaging is not yet optimized, as it does not incorporate the inborn properties of qDPC imaging. In this research, we propose a simple but effective image prior, the dark-field sparse prior (DSP), to facilitate the phase reconstruction quality for all DPC-based phase reconstruction algorithms. The DSP is based on the key observation that most pixel values for an idea differential phase contrast image are zeros since the subtraction of two images under anti-symmetric illumination cancels all background components. With this DSP prior, we formed a new cost function in which L0-norm was used to represent the DSP. Further, we developed two different algorithms based on (1) the Half Quadratic Splitting, and (2) the Richardson-Lucy deconvolution to solve this NP-hard L0-norm problem. We tested our new model on both simulated and experimental data and compare against state-of-the-art methods including L2-norm and total variation regularizations. Results show that our proposed model is superior in terms of phase reconstruction quality and implementation efficiency, in which it significantly increases the experimental robustness, while maintaining the data fidelity.

math.OC↗