arXiv subjects

Zhisheng Ye

Publications and source records attributed to Zhisheng Ye.

13 recordsLinked to original sources

arXiv2026

OctoPipe: Reducing Pipeline Bubbles for Heterogeneous Models via Co-Optimizing Partitioning, Placement, and Scheduling

Pipeline parallelism is widely used to train large language models (LLMs). However, increasing heterogeneity in model architectures exacerbates pipeline bubbles, thereby reducing training efficiency. Prior approaches typically optimize a single phase of the pipeline schedule (i.e., partitioning, placement, or scheduling), leaving substantial pipeline bubbles. While promising, co-optimization poses three key challenges: (1) complex performance modeling, (2) a combinatorial search space, and (3) irregular execution orders. To address these challenges, we propose OctoPipe, a pipeline parallelism system to jointly optimize partitioning, placement, and scheduling. First, we build a graph-based pipeline simulator to model heterogeneous pipeline execution for co-optimization. Second, on top of the simulator, we develop an iterative bubble-aware tuner to efficiently explore the combinatorial search space. Third, we implement a unified pipeline executor that dynamically orchestrates computation and communication to support irregular execution orders without deadlocks while maximizing communication-computation overlap. Experiments show that OctoPipe achieves 1.09--1.49$\times$ throughput improvement over the state-of-the-art pipeline parallelism approaches across various heterogeneous model configurations and GPU cluster scales.

cs.DC↗

arXiv2026

SpecGen: Accelerating Agentic Kernel Optimization with Speculative Generation

Agentic kernel optimization automates manual GPU kernel tuning via iterative generation, validation, and profiling with reasoning LLMs, casting the optimization task as feedback-guided search. However, our workload characterization reveals three system-level inefficiencies that limit search efficiency: (1) long generation latency due to LLM reasoning, (2) insufficient profiling feedback, and (3) underutilized validation/profiling resources. Our key insight is that the ongoing reasoning generation exposes a window for producing additional candidate kernels before it completes, allowing the system to terminate reasoning early once a satisfactory kernel appears. We present SpecGen, an agentic kernel optimization system with \emph{speculative generation}. First, SpecGen forks non-reasoning generations at well-chosen trigger points in the reasoning trace to yield kernels, increasing the candidate kernel count per iteration. These kernels are validated and profiled in parallel with the ongoing reasoning, increasing profiling feedback, and keeping resources busy during generation. When a kernel meets the termination criterion, SpecGen terminates the reasoning generation early to reduce the generation latency. Second, SpecGen dynamically reallocates validation and profiling GPU pools based on the arrival rate and prioritizes requests to reduce profiling feedback latency under bursty speculative generation load. Furthermore, SpecGen utilizes spare memory of the validation/profiling GPUs as remote KV cache storage to eliminate prefix recomputation of speculative generations under limited memory budget. Experiments with two reasoning LLMs on H200 show that SpecGen reduces end-to-end time over three baseline systems, while producing more profiling feedback, increasing resource utilization, and improving kernel speedup under a fixed time and token budget.

cs.DC↗

arXiv2026

ResiHP: Taming LLM Training Failures with Dynamic Hybrid Parallelism

Hybrid parallelism underpins large-scale LLM training across tens of thousands of GPUs. At such scale, hardware failures on individual devices lead to performance skew across devices, diminishing overall training efficiency. Existing resilient systems overlook sequence length variability in datasets and device performance skew under hybrid parallelism. As a result, (1) iteration time fluctuations induced by sequence length variability can trigger spurious fail-slow detections, and (2) failures are mitigated through individual adaptations in hybrid parallelism, leading to unnecessary detection overhead and inefficient resilient training. To respond, this paper presents ResiHP, a resilient system that enables robust failure detection and fine-grained adaptation for hybrid parallel training. First, we develop a Detector to accurately identify failures. In particular, it employs a workload-aware execution time predictor that disentangles failures from iteration time fluctuations while remaining lightweight for online detection. Second, we design a Scheduler that dynamically adapts parallelism group sizes, model partitioning, and workload scheduling policies to improve training efficiency under failures. Experiments show that ResiHP improves training throughput by 1.04-4.39$\times$ compared with state-of-the-art resilient training systems under diverse failure scenarios in a 256-GPU cluster.

cs.DC↗

arXiv2026

CONCUR: High-Throughput Agentic Batch Inference of LLM via Congestion-Based Concurrency Control

Batch inference for agentic workloads stresses the GPU key-value (KV) cache in a sustained and cumulative manner, often causing severe throughput degradation well before memory capacity is exhausted. We identify this phenomenon as middle-phase thrashing, a previously under-characterized pathology in which cache efficiency collapses as long-lived agents accumulate state over time. We argue that mitigating this pathology requires moving beyond reactive, request-level cache management to proactive, agent-level admission control. Drawing inspiration from congestion control in distributed systems, we view the KV cache as a shared resource whose efficient utilization depends on feedback-driven regulation. Based on this insight, we present CONCUR, a lightweight control layer that regulates agent admission to bound aggregate cache pressure while preserving execution continuity. CONCUR adapts a cache-aware control algorithm to dynamically adjust the number of active agents using runtime cache signals. Across large models and real-world agent workloads, CONCUR prevents middle-phase thrashing and improves batch inference throughput by up to 4.09x on Qwen3-32B and 1.9x on DeepSeek-V3, while remaining compatible with existing LLM serving systems.

cs.DC↗

arXiv2025

LEMUR: Large scale End-to-end MUltimodal Recommendation

Traditional ID-based recommender systems often struggle with cold-start and generalization challenges. Multimodal recommendation systems, which leverage textual and visual data, offer a promising solution to mitigate these issues. However, existing industrial approaches typically adopt a two-stage training paradigm: first pretraining a multimodal model, then applying its frozen representations to train the recommendation model. This decoupled framework suffers from misalignment between multimodal learning and recommendation objectives, as well as an inability to adapt dynamically to new data. To address these limitations, we propose LEMUR, the first large-scale multimodal recommender system trained end-to-end from raw data. By jointly optimizing both the multimodal and recommendation components, LEMUR ensures tighter alignment with downstream objectives while enabling real-time parameter updates. Constructing multimodal sequential representations from user history often entails prohibitively high computational costs. To alleviate this bottleneck, we propose a novel memory bank mechanism that incrementally accumulates historical multimodal representations throughout the training process. After one month of deployment in Douyin Search, LEMUR has led to a 0.843% reduction in query change rate decay and a 0.81% improvement in QAUC. Additionally, LEMUR has shown significant gains across key offline metrics for Douyin Advertisement. Our results validate the superiority of end-to-end multimodal recommendation in real-world industrial scenarios.

cs.IR↗

arXiv2025

Memory Offloading for Large Language Model Inference with Latency SLO Guarantees

Offloading large language models (LLMs) state to host memory during inference promises to reduce operational costs by supporting larger models, longer inputs, and larger batch sizes. However, the design of existing memory offloading mechanisms does not take latency service-level objectives (SLOs) into consideration. As a result, they either lead to frequent SLO violations or underutilize host memory, thereby incurring economic loss and thus defeating the purpose of memory offloading. This paper presents Select-N, a latency-SLO-aware memory offloading system for LLM serving. A key challenge in designing Select-N is to reconcile the tension between meeting SLOs and maximizing host memory usage. Select-N overcomes it by exploiting a unique characteristic of modern LLMs: during serving, the computation time of each decoder layer is deterministic. Leveraging this, Select-N introduces offloading interval, an internal tunable knob that captures the tradeoff between SLOs and host memory usage, thereby reducing the aforementioned challenge to pick an optimal offloading interval. With that, Select-N proposes a two-stage approach to automatically pick the offloading interval. The first stage is offline that generates the range of optimal offloading interval, while the second stage adjusts offloading interval at the granularity of inference iteration based on runtime hardware status. Our evaluation shows that Select-N consistently meets SLOs and improves the serving throughput over existing mechanisms by 1.85X due to maximizing the use of host memory.

cs.DC↗

arXiv2024

Characterization of Large Language Model Development in the Datacenter

Large Language Models (LLMs) have presented impressive performance across several transformative tasks. However, it is non-trivial to efficiently utilize large-scale cluster resources to develop LLMs, often riddled with numerous challenges such as frequent hardware failures, intricate parallelization strategies, and imbalanced resource utilization. In this paper, we present an in-depth characterization study of a six-month LLM development workload trace collected from our GPU datacenter Acme. Specifically, we investigate discrepancies between LLMs and prior task-specific Deep Learning (DL) workloads, explore resource utilization patterns, and identify the impact of various job failures. Our analysis summarizes hurdles we encountered and uncovers potential opportunities to optimize systems tailored for LLMs. Furthermore, we introduce our system efforts: (1) fault-tolerant pretraining, which enhances fault tolerance through LLM-involved failure diagnosis and automatic recovery. (2) decoupled scheduling for evaluation, which achieves timely performance feedback via trial decomposition and scheduling optimization.

cs.DC↗

arXiv2022

Deep Learning Workload Scheduling in GPU Datacenters: Taxonomy, Challenges and Vision

Deep learning (DL) shows its prosperity in a wide variety of fields. The development of a DL model is a time-consuming and resource-intensive procedure. Hence, dedicated GPU accelerators have been collectively constructed into a GPU datacenter. An efficient scheduler design for such GPU datacenter is crucially important to reduce the operational cost and improve resource utilization. However, traditional approaches designed for big data or high performance computing workloads can not support DL workloads to fully utilize the GPU resources. Recently, substantial schedulers are proposed to tailor for DL workloads in GPU datacenters. This paper surveys existing research efforts for both training and inference workloads. We primarily present how existing schedulers facilitate the respective workloads from the scheduling objectives and resource consumption features. Finally, we prospect several promising future research directions. More detailed summary with the surveyed paper and code links can be found at our project website: https://github.com/S-Lab-System-Group/Awesome-DL-Scheduling-Papers

cs.DC↗

arXiv2021

The Internet of Federated Things (IoFT): A Vision for the Future and In-depth Survey of Data-driven Approaches for Federated Learning

The Internet of Things (IoT) is on the verge of a major paradigm shift. In the IoT system of the future, IoFT, the cloud will be substituted by the crowd where model training is brought to the edge, allowing IoT devices to collaboratively extract knowledge and build smart analytics/models while keeping their personal data stored locally. This paradigm shift was set into motion by the tremendous increase in computational power on IoT devices and the recent advances in decentralized and privacy-preserving model training, coined as federated learning (FL). This article provides a vision for IoFT and a systematic overview of current efforts towards realizing this vision. Specifically, we first introduce the defining characteristics of IoFT and discuss FL data-driven approaches, opportunities, and challenges that allow decentralized inference within three dimensions: (i) a global model that maximizes utility across all IoT devices, (ii) a personalized model that borrows strengths across all devices yet retains its own model, (iii) a meta-learning model that quickly adapts to new devices or learning tasks. We end by describing the vision and challenges of IoFT in reshaping different industries through the lens of domain experts. Those industries include manufacturing, transportation, energy, healthcare, quality & reliability, business, and computing.

cs.LG↗

arXiv2021

Asymptotic Analysis for Data-Driven Inventory Policies

We study periodic review stochastic inventory control in the data-driven setting where the retailer makes ordering decisions based only on historical demand observations without any knowledge of the probability distribution of the demand. Since an (s, S)-policy is optimal when the demand distribution is known, we investigate the statistical properties of the data-driven (s, S)-policy obtained by recursively computing the empirical cost-to-go functions. This policy is inherently challenging to analyze because the recursion induces propagation of the estimation error backwards in time. In this work, we establish the asymptotic properties of this data-driven policy by fully accounting for the error propagation. First, we rigorously show the consistency of the estimated parameters by filling in some gaps (due to unaccounted error propagation) in the existing studies. In this setting, empirical process theory (EPT) cannot be directly applied to show asymptotic normality. To explain, the empirical cost-to-go functions for the estimated parameters are not i.i.d. sums due to the error propagation. Our main methodological innovation comes from an asymptotic representation for multi-sample U-processes in terms of i.i.d. sums. This representation enables us to apply EPT to derive the influence functions of the estimated parameters and to establish joint asymptotic normality. Based on these results, we also propose an entirely data-driven estimator of the optimal expected cost and we derive its asymptotic distribution. We demonstrate some useful applications of our asymptotic results, including sample size determination and interval estimation. The results from our numerical simulations conform to our theoretical analysis.lations conform to our theoretical analysis.

math.ST↗

arXiv2021

A Unifying Framework for Variance Reduction Algorithms for Finding Zeroes of Monotone Operators

It is common to encounter large-scale monotone inclusion problems where the objective has a finite sum structure. We develop a general framework for variance-reduced forward-backward splitting algorithms for this problem. This framework includes a number of existing deterministic and variance-reduced algorithms for function minimization as special cases, and it is also applicable to more general problems such as saddle-point problems and variational inequalities. With a carefully constructed Lyapunov function, we show that the algorithms covered by our framework enjoy a linear convergence rate in expectation under mild assumptions. We further consider Catalyst acceleration and asynchronous implementation to reduce the algorithmic complexity and computation time. We apply our proposed framework to a policy evaluation problem and a strongly monotone two-player game, both of which fall outside of function minimization.

stat.ML↗

arXiv2018

Index-Based Policy for Risk-Averse Multi-Armed Bandit

The multi-armed bandit (MAB) is a classical online optimization model for the trade-off between exploration and exploitation. The traditional MAB is concerned with finding the arm that minimizes the mean cost. However, minimizing the mean does not take the risk of the problem into account. We now want to accommodate risk-averse decision makers. In this work, we introduce a coherent risk measure as the criterion to form a risk-averse MAB. In particular, we derive an index-based online sampling framework for the risk-averse MAB. We develop this framework in detail for three specific risk measures, i.e. the conditional value-at-risk, the mean-deviation and the shortfall risk measures. Under each risk measure, the convergence rate for the upper bound on the pseudo regret, defined as the difference between the expectation of the empirical risk based on the observation sequence and the true risk of the optimal arm, is established.

math.OC↗

arXiv2014

On analysis of incomplete field failure data

Many commercial products are sold with warranties and indirectly through dealers. The manufacturer-retailer distribution mechanism results in serious missing data problems in field return data, as the sales date for an unreturned unit is generally unknown to the manufacturer. This study considers a general setting for field failure data with unknown sales dates and a warranty limit. A stochastic expectation-maximization (SEM) algorithm is developed to estimate the distributions of the sales lag (time between shipment to a retailer and sale to a customer) and the lifetime of the product under study. Extensive simulations are used to evaluate the performance of the SEM algorithm and to compare with the imputation method proposed by Ghosh [Ann. Appl. Stat. 4 (2010) 1976-1999]. Three real examples illustrate the methodology proposed in this paper.

stat.AP↗