paperbot · PL 论文追踪

全部论文 待读 收藏

P1 TELLER: Non-intrusive Cross-Layer Root-Cause Analysis for LLM Inference

TELLER：面向 LLM 推理的非侵入式跨层根因分析
ASE Proceedings of the 41st IEEE/ACM International Conference on Automated Software Engineering2026匹配度 5/10 · 该论文面向 LLM 推理服务栈的跨层根因分析，属于机器学习系统 + 程序语言（测试/调试/故障定位）的交叉，且用多模态模型作为分析工具，与“AI→PL 用模型改进程序分析”与“ML 系统分析”有部分重合；但它基于运行时 trace/log 的动态诊断，而非用户核心关注的静态分析、浮点程序分析或贝叶斯程序分析，也不涉及 Agentic 程序，因此只落在研究版图的边缘而非核心。引用 1 inbox
Ruilin Xu, Junyi Li, Pengfei Chen, Zongxuan Xie
测试、调试与修复机器学习与程序语言
crossref: 2026-10-09

AI 速览

一句话：提出非侵入式跨层根因分析框架 TELLER，用结构化 trace 表示与多模态模型定位 LLM 推理异常步骤并生成解释。

LLM 推理已成为持续运行的在线服务，单次请求横跨推理引擎、Python/C++ 后端、CUDA API、GPU kernel 与分布式通信，现有 profiler 只暴露原始时间线、日志诊断缺乏跨层语义与请求级结构，根因分析困难。TELLER 不修改模型二进制即可采集 NVTX/CUPTI trace 与服务日志，重建每请求调用链树并对齐日志；提出依赖感知因果上下文切片与 Trace Pair Encoding 分词器，将切片压缩为带父节点、深度、时长属性的结构化 token 序列，再以数值候选定位结合多模态根因模型联合预测异常步骤、定位可疑算子并生成自然语言解释。多节点 GPU 负载上，中等词表在压缩每步 trace 80% 以上的同时取得横纵向视图最佳性能，过度压缩则显著降低诊断质量。

问题
LLM 推理服务中单次请求的执行横跨推理引擎、Python/C++ 后端、主机 CUDA API、GPU kernel 与分布式通信，现有 profiler 仅提供原始时间线、日志诊断又缺乏跨层执行语义与请求级结构，导致根因分析困难。
方法
在不修改模型二进制的前提下采集 NVTX/CUPTI trace 与服务日志，重建请求级调用链树并将日志行对齐到执行步骤；提出依赖感知的因果上下文切片和 Trace Pair Encoding 分词器把切片编码为紧凑结构 token 序列，再用数值候选定位结合多模态根因模型联合预测异常步骤、定位可疑算子并生成自然语言解释。
评估
在多节点 GPU 推理负载上与基线对比，并进行低保真先验、模态消融、解释质量与 tracing 开销分析；结果显示中等 TPE 词表可把每步 trace 长度压缩 80% 以上并取得横向（跨节点通信）与纵向（节点内执行栈）视图的最佳整体诊断性能，而过度压缩会显著降低诊断质量。
原文摘要（Abstract）

Large language model (LLM) inference has evolved from an offline workload into a continuously operated software service, yet root-cause analysis remains difficult because a single request spans the inference engine, Python/C++ backend, host CUDA APIs, GPU kernels, and distributed communication. Existing profilers expose raw timelines, while log-based diagnosis often misses cross-layer execution semantics and request-level structure. We present TELLER, a non-intrusive trace- and log-aware LLM inference root-cause analysis framework. TELLER first collects NVTX/CUPTI traces and service logs without modifying model binaries, then reconstructs per-request call-chain trees and aligns log lines with the corresponding execution steps. We introduce a dependency-aware causal-context slice that preserves parent-child structure, temporal order, and communication relations, and a Trace Pair Encoding (TPE) tokenizer that compresses such slices into compact structural token sequences with parent, depth, and duration attributes. On top of these representations, TELLER combines numeric candidate localization with a multimodal root-cause model that jointly predicts abnormal steps, localizes suspicious operators, and generates natural-language explanations. Experiments on multi-node GPU inference workloads show a clear compression-accuracy trade-off: a moderate TPE vocabulary reduces per-step trace length by more than 80% while achieving the best overall performance on both horizontal (cross-node communication) and vertical (within-node execution stack) views, whereas more aggressive compression substantially degrades diagnosis quality. Further analyses under low-fault priors, strengthened baselines, modality ablations, explanation-quality checks, and tracing overhead show that TELLER provides a practical triage and evidence-localization substrate for LLM inference RCA.

链接与引用

DOI 原文 · arXiv ·

BibTeX 
@inproceedings{paperbot12267,
  title = {TELLER: Non-intrusive Cross-Layer Root-Cause Analysis for LLM Inference},
  author = {Ruilin Xu and Junyi Li and Pengfei Chen and Zongxuan Xie},
  booktitle = {ASE},
  year = {2026},
  doi = {10.1145/3832783.3837440}
}