arXiv subjects

Mingyu Sun

Publications and source records attributed to Mingyu Sun.

5 recordsLinked to original sources

arXiv2025

Collaborative Lossless LLM Inference Serving with Offloading-based Pipeline Parallelism on Edge Devices

Providing lossless inference services of LLMs on edge devices remains challenging, especially given the extremely tight memory budgets. The existing offloading techniques inevitably introduce numerous loading bubbles, which further inflate the end-to-end latency of the entire inference pipeline. Meanwhile, dynamically fluctuating network bandwidth and diverse user request patterns pose additional obstacles to efficient lossless inference on edge devices. To address this, we propose LOIP, a collaborative lossless LLM inference system that employs an offloading-based interleaved pipeline parallelism to better overlap model offloading with computing and communicating. Specifically, LOIP first constructs an offloading-aware cost model to characterize inference latency and memory overhead under heterogeneous device capabilities and limited bandwidth. Based on this cost model, LOIP develops a fine-grained allocation scheduler that determines latency-efficient layer partitions across devices while explicitly accounting for offloading overhead, along with a unified memory architecture (UMA)-aware loading optimization using customized CUDA operators to reduce runtime loading overhead. LOIP further designs an online memory adaptation strategy to handle the increasing KV cache pressure and dynamic bandwidth fluctuations during inference. We implement LOIP with 2500+ lines of Python and 500+ lines of C++/CUDA code, and deploy it on five heterogeneous NVIDIA Jetson edge devices for lossless collaborative inference of LLaMA3.3-70B-Instruct. Extensive experiments demonstrate that LOIP achieves 8.8$\times$$\sim$20.3$\times$ speedups over the SOTA baselines under different bandwidth conditions and request patterns without compromising model accuracy.

cs.DC↗

arXiv2025

Efficient Fidelity Estimation with Few Local Pauli Measurements

As quantum devices scale, quantifying how close an experimental state aligns with a target becomes both vital and challenging. Fidelity is the standard metric, but existing estimators either require full tomography or apply only to restricted state/measurement families. Huang, Preskill, and Soleimanifar (Nature Physics, 2025) introduced an efficient certification protocol for Haar-random states using only a polynomial number of non-adaptive, single-copy, local Pauli measurements. Here, we adopt the same data collection routine but recast it as a fidelity estimation protocol with rigorous performance guarantees and broaden its applicability. We analyze the bias in this estimator, linking its performance to the mixing time $\tau$ of a Markov chain induced by the target state, and resolve the three open questions posed by Huang, Preskill, and Soleimanifar (Nature Physics, 2025). Our analysis extends beyond Haar-random states to state $t$-designs, states prepared by low-depth random circuits, physically relevant states and families of mixed states. We introduce a $k$-generalized local escape property that identifies when the fidelity estimation protocol is both efficient and accurate, and design a practical empirical test to verify its applicability for arbitrary states. This work enables scalable benchmarking, error characterization, and tomography assistance, supports adaptive quantum algorithms in high dimensions, and clarifies fundamental limits of learning from local measurements.

quant-ph↗

arXiv2024

Optimizing MCMC-Driven Bayesian Neural Networks for High-Precision Medical Image Classification in Small Sample Sizes

This paper discusses the application of a Bayesian neural network based on the Markov Chain Monte Carlo method in medical image classification with small samples. Experimental results on two medical image datasets, including lung X-ray images and breast tissue slice images, show that this MCMC-based BNN model works very well on small-sample data and greatly improves the robustness and accuracy of classification. Model accuracy reached 85% for the lung X-ray dataset and 88% for the breast tissue slice dataset. To this end, we combine data augmentation techniques such as rotation, flipping, and scaling with regularization methods like dropout and weight decay to improve effectively the diversity of the training data and the generalization ability of the model. The performance of the model was evaluated by many indicators of the results, including accuracy, precision, recall, and the F1 score. All of these have proven the advantages of BNN in small-sample medical image classification. This study not only enriches the application of BNN in the field of medical image classification, but also provides specific implementation paths and optimization methods, providing new solutions for future medical image analysis.

stat.CO↗

arXiv2020

Efficient correction of multiqubit measurement errors

State preparation and measurement (SPAM) errors limit the performance of near-term quantum computers and their potential for practical application. SPAM errors are partly correctable after a calibration step that requires, for a complete implementation on a register of $n$ qubits, $2^n$ additional measurements. Here we introduce an approximate but efficient method for multiqubit SPAM error characterization and mitigation requiring the classical processing of $2^n \! \times 2^n$ matrices, but only $O(4^k n^2)$ measurements, where $k=O(1)$ is the number of qubits in a correlation volume. We demonstrate and validate the technique using an IBM Q processor on registers of 4 and 8 superconducting qubits.

quant-ph↗

arXiv2018

Efficient characterization of correlated SPAM errors

State preparation and measurement (SPAM) errors limit the performance of many gate-based quantum computing architecures, but are partly correctable after a calibration step that requires, for an exact implementation on a register of $n$ qubits, $2^n$ additional characterization experiments, as well as classical post-processing. Here we introduce an approximate but efficient method for SPAM error characterization requiring the {\it classical} processing of $2^n \! \times 2^n$ real matrices, but only $O(n^2)$ measurements. The technique assumes that multi-qubit measurement errors are dominated by pair correlations, which are estimated with $n(n-1)k/2$ two-qubit experiments, where $k$ is a parameter related to the accuracy. We demonstrate the technique on the IBM and Rigetti online superconducting quantum computers, allowing comparison of their SPAM errors in both magnitude and degree of correlation. We also study the correlations as a function of the register's geometric layout. We find that the pair-correlation model is fairly accurate on linear arrays of superconducting qubits. However qubits arranged in more closely spaced two-dimensional geometries exhibit significant higher-order (such as 3-qubit) SPAM error correlations.

quant-ph↗