arXiv subjects

Yichuan Bai

Publications and source records attributed to Yichuan Bai.

6 recordsLinked to original sources

arXiv2026

CD-PIM: A High-Bandwidth and Compute-Efficient LPDDR5-Based PIM for Low-Batch LLM Acceleration on Edge-Device

Edge deployment of low-batch large language models (LLMs) faces critical memory bandwidth bottlenecks when executing memory-intensive general matrix-vector multiplications (GEMV) operations. While digital processing-in-memory (PIM) architectures promise to accelerate GEMV operations, existing PIM-equipped edge devices still suffer from three key limitations: limited bandwidth improvement, component under-utilization in mixed workloads, and low compute capacity of computing units (CUs). In this paper, we propose CD-PIM to address these challenges through three key innovations. First, we introduce a high-bandwidth compute-efficient mode (HBCEM) that enhances bandwidth by dividing each bank into four pseudo-banks through segmented global bitlines. Second, we propose a low-batch interleaving mode (LBIM) to improve component utilization by overlapping GEMV operations with GEMM operations. Third, we design a compute-efficient CU that performs enhanced GEMV operations in a pipelined manner by serially feeding weight data into the computing core. Forth, we adopt a column-wise mapping for the key-cache matrix and row-wise mapping for the value-cache matrix, which fully utilizes CU resources. Our evaluation shows that compared to a GPU-only baseline and state-of-the-art PIM designs, our CD-PIM achieves 11.42x and 4.25x speedup on average within a single batch in HBCEM mode, respectively. Moreover, for low-batch sizes, the CD-PIM achieves an average speedup of 1.12x in LBIM compared to HBCEM.

cs.AR↗

arXiv2026

A Scheduling Framework for Efficient MoE Inference on Edge GPU-NDP Systems

Mixture-of-Experts (MoE) models facilitate edge deployment by decoupling model capacity from active computation, yet their large memory footprint drives the need for GPU systems with near-data processing (NDP) capabilities that offload experts to dedicated processing units. However, deploying MoE models on such edge-based GPU-NDP systems faces three critical challenges: 1) severe load imbalance across NDP units due to non-uniform expert selection and expert parallelism, 2) insufficient GPU utilization during expert computation within NDP units, and 3) extensive data pre-profiling necessitated by unpredictable expert activation patterns for pre-fetching. To address these challenges, this paper proposes an efficient inference framework featuring three key optimizations. First, the underexplored tensor parallelism in MoE inference is exploited to partition and compute large expert parameters across multiple NDP units simultaneously towards edge low-batch scenarios. Second, a load-balancing-aware scheduling algorithm distributes expert computations across NDP units and GPU to maximize resource utilization. Third, a dataset-free pre-fetching strategy proactively loads frequently accessed experts to minimize activation delays. Experimental results show that our framework enables GPU-NDP systems to achieve 2.41x on average and up to 2.56x speedup in end-to-end latency compared to state-of-the-art approaches, significantly enhancing MoE inference efficiency in resource-constrained environments.

cs.DC↗

arXiv2025

A Robust Framework for Graph-based Two-Sample Tests Using Weights

Graph-based tests are a class of non-parametric two-sample tests useful for analyzing high-dimensional data. The test statistics are constructed from similarity graphs (such as K-minimum spanning tree), and consequently, their performance is sensitive to the structure of the graph. When the graph has problematic structures (for example, hubs), as is common for high-dimensional data, this can result in low power and unstable performance among existing graph-based tests. We address this challenge by proposing new test statistics that are robust to problematic structures of the graph and can provide reliable inferences. We employ an edge-weighting strategy using intrinsic characteristics of the graph that are computationally simple and efficient to obtain. The limiting null distribution of the robust test statistics is derived and shown to work well for finite sample sizes. Simulation studies and data analysis of Chicago taxi-trip travel patterns demonstrate the new tests' improved performance across a range of settings.

stat.ME↗

arXiv2025

A Graph-based Approach to Estimating the Number of Clusters in High-dimensional Settings

We consider the problem of estimating the number of clusters (k) in a dataset. We propose a non-parametric approach to the problem that utilizes similarity graphs to construct a robust statistic that effectively captures similarity information among observations. This graph-based statistic is applicable to datasets of any dimension, is computationally efficient to obtain, and can be paired with any kind of clustering technique. Asymptotic theory is developed to establish the selection consistency of the proposed approach. Simulation studies demonstrate that the graph-based statistic outperforms existing methods for estimating k, especially in the high-dimensional setting. We illustrate its utility on an imaging dataset and an RNA-seq dataset.

stat.ME↗

arXiv2025

Graph-theoretic Inference for Random Effects in High-dimensional Studies

We study the problem of testing for the presence of random effects in mixed models with high-dimensional fixed effects. To this end, we propose a rank-based graph-theoretic approach to test whether a collection of random effects is zero. Our approach is non-parametric and model-free in the sense that we not require correct specification of the mixed model nor estimation of unknown parameters. Instead, the test statistic evaluates whether incorporating group-level correlation meaningfully improves the ability of a potentially high-dimensional covariate vector $X$ to predict a response variable $Y$. We establish the consistency of the proposed test and derive its asymptotic null distribution. Through simulation studies and a real data application, we demonstrate the practical effectiveness of the proposed test.

stat.ME↗

arXiv2025

FBQuant: FeedBack Quantization for Large Language Models

Deploying Large Language Models (LLMs) on edge devices is increasingly important, as it eliminates reliance on network connections, reduces expensive API calls, and enhances user privacy. However, on-device deployment is challenging due to the limited computational resources of edge devices. In particular, the key bottleneck stems from memory bandwidth constraints related to weight loading. Weight-only quantization effectively reduces memory access, yet often induces significant accuracy degradation. Recent efforts to incorporate sub-branches have shown promise for mitigating quantization errors, but these methods either lack robust optimization strategies or rely on suboptimal objectives. To address these gaps, we propose FeedBack Quantization (FBQuant), a novel approach inspired by negative feedback mechanisms in automatic control. FBQuant inherently ensures that the reconstructed weights remain bounded by the quantization process, thereby reducing the risk of overfitting. To further offset the additional latency introduced by sub-branches, we develop an efficient CUDA kernel that decreases 60% of extra inference time. Comprehensive experiments demonstrate the efficiency and effectiveness of FBQuant across various LLMs. Notably, for 3-bit Llama2-7B, FBQuant improves zero-shot accuracy by 1.2%.

cs.LG↗