arXiv subjects

Yiheng Tao

Publications and source records attributed to Yiheng Tao.

5 recordsLinked to original sources

arXiv2026

Ranking Before Serving: Low-Latency LLM Serving via Pairwise Learning-to-Rank

Efficient scheduling of large language model (LLM) inference tasks is critical for achieving low latency and high throughput, a challenge that is becoming increasingly acute with the rise of reasoning-capable LLMs whose generation lengths are highly variable. Traditional strategies like First Come, First-Serve (FCFS) often suffer from Head-of-Line (HOL) blocking, where long-running tasks delay shorter ones queued behind them. In this paper, we introduce PARS, a prompt-aware LLM task scheduler that mitigates HOL blocking by approximating shortest-job-first (SJF) scheduling through pairwise ranking with a margin ranking loss. PARS effectively predicts response-length-based task ordering directly from prompts, thereby optimizing scheduling decisions with minimal overhead. In addition, it integrates seamlessly with vLLM, a state-of-the-art LLM serving system, for the research community. Extensive experiments across multiple LLM models and real-world inference use cases, including chat, math, and code generation, demonstrate that PARS significantly reduces latency by up to 15.7x compared to the vLLM default scheduler. Cross-model evaluations demonstrate that our design generalizes effectively, allowing effective scheduling across diverse LLMs without requiring model-specific retraining.

cs.LG↗

arXiv2026

The Devil is in the Spectrum: Mitigating Representation Collapse in LLMs via Topologically Regularized Side-Path

Large Language Models (LLMs) are fundamentally limited by representation collapse, a bottleneck that severely degrades long-context performance. We identify that existing approaches risk drifting into one of two pathological extremes: homogenization collapse (e.g., attention sinks causing rank deficiency) and isolation collapse (e.g., local attention causing context disconnection). Through spectral analysis of attention dynamics, we derive an intrinsic trade-off between mixing efficiency (spectral gap) and information capacity (effective rank) that standard mechanisms struggle to balance. To resolve this dilemma, we propose the Topologically Regularized Side-Path (TRSP), a non-invasive architectural intervention that achieves spectral balance. TRSP employs a parameter-free Triangular Box mechanism, scaled by a lightweight, length-aware gate, to regularize the token interaction topology. By integrating proximal coupling to preserve effective rank and distal propagation to support non-degenerate mixing, TRSP promotes a geometrically healthier transition operator without altering core attention. Experiments show significant improvements across general capabilities and long-context benchmarks. Notably, on NoLiMa at $8\times$ the training length, TRSP retains $83\%$ accuracy and surpasses the Differential Transformer and Gated Attention by approximately 30 and 50 percentage points, respectively. Code available at: https://github.com/Eziotao-tyd/TRSP.

cs.AI↗

arXiv2026

Leveraging LLMs to Automate Energy-Aware Refactoring of Parallel Scientific Codes

Large language models (LLMs) are increasingly used for generating parallel scientific codes, with a primary focus on generating functionally correct code. Recent work has focused on generating performant code, with an emphasis on its execution time. However, energy efficiency is now recognized as a critical objective, given the significant power demands of large-scale compute systems. This paper addresses the research question of whether LLMs can generate energy-efficient parallel scientific codes when guided by empirical execution feedback. To answer this question, we propose LASSI-EE, an automated LLM-based refactoring framework that generates energy-efficient parallel codes through a multi-stage, iterative approach integrating runtime power profiling, energy-aware prompting, self-correcting feedback loops, and an LLM-as-a-Judge agent for screening generated code. We evaluate LASSI-EE using twenty-two representative scientific benchmarks and applications on NVIDIA A100 and AMD MI100 GPUs. The results indicate an average energy reduction of 36% for MI100 and 34% for A100, across trials that produced passing energy-reducing refactorings.

cs.AI↗

arXiv2025

A Real-Time Digital Twin for Adaptive Scheduling

High-performance computing (HPC) workloads are becoming increasingly diverse, exhibiting wide variability in job characteristics, yet cluster scheduling has long relied on static, heuristic-based policies. In this work we present SchedTwin, a real-time digital twin designed to adaptively guide scheduling decisions using predictive simulation. SchedTwin periodically ingests runtime events from the physical scheduler, performs rapid what-if evaluations of multiple policies using a high-fidelity discrete-event simulator, and dynamically selects the one satisfying the administrator configured optimization goal. We implement SchedTwin as an open-source software and integrate it with the production PBS scheduler. Preliminary results show that SchedTwin consistently outperforms widely used static scheduling policies, while maintaining low overhead (a few seconds per scheduling cycle). These results demonstrate that real-time digital twins offer a practical and effective path toward adaptive HPC scheduling.

cs.DC↗

arXiv2025

LASSI: An LLM-based Automated Self-Correcting Pipeline for Translating Parallel Scientific Codes

This paper addresses the problem of providing a novel approach to sourcing significant training data for LLMs focused on science and engineering. In particular, a crucial challenge is sourcing parallel scientific codes in the ranges of millions to billions of codes. To tackle this problem, we propose an automated pipeline framework called LASSI, designed to translate between parallel programming languages by bootstrapping existing closed- or open-source LLMs. LASSI incorporates autonomous enhancement through self-correcting loops where errors encountered during the compilation and execution of generated code are fed back to the LLM through guided prompting for debugging and refactoring. We highlight the bi-directional translation of existing GPU benchmarks between OpenMP target offload and CUDA to validate LASSI. The results of evaluating LASSI with different application codes across four LLMs demonstrate the effectiveness of LASSI for generating executable parallel codes, with 80% of OpenMP to CUDA translations and 85% of CUDA to OpenMP translations producing the expected output. We also observe approximately 78% of OpenMP to CUDA translations and 62% of CUDA to OpenMP translations execute within 10% of or at a faster runtime than the original benchmark code in the same language.

cs.SE↗