Understanding and simplifying bug-inducing inputs for deep-learning compilers

Liu, Jiawei

Permalink

https://hdl.handle.net/2142/124525 Copy

Description

Title

Understanding and simplifying bug-inducing inputs for deep-learning compilers

Author(s)

Liu, Jiawei

Issue Date

2024-04-18

Director of Research (if dissertation) or Advisor (if thesis)

Zhang, Lingming

Department of Study

Computer Science

Discipline

Computer Science

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• Compiler Testing
• Test Reduction
• Deep Learning Compilers

Language

eng

Abstract

With the recent evolution of parallel computing and compilation techniques, deep learning (DL) models have increasingly been deployed using DL compilers to achieve real-time performance in resource-constrained environments. The compiler stack is complex and initially drafted, making the debugging of DL compilers challenging, especially given the limited time and number of compiler experts available. Complicating matters further, the massive and often duplicated bug reports (e.g., fuzzing bugs) may contain obscure or even no error messages (e.g., semantic bugs) for models consisting of thousands of layers. This situation makes it extremely challenging to understand and localize compiler failures. To address this emerging challenge, we propose new techniques for automatically reducing and deduplicating DL compiler bug reports: We define the test-case reduction problem in DL compilers for general graph-level inputs and propose sub-graph minimality as the reduction target. Based on the Delta Debugging framework, we perform validity-preserving graph partitioning guided by heuristics from error-prone patterns. Using graph matching, we deduplicate test cases with graph isomorphism and expedite reductions by matching minimized graphs against incoming ones. We have integrated our techniques into a new tool, NNReduce, which is the first graph-level reducer with a sub-graph minimality guarantee, implemented for ONNX model formats. We applied NNReduce to reduce and deduplicate both fuzzer-generated and real-world bug reports, creating a new dataset for extensive evaluation. Our results show that NNReduce can reduce test cases by up to 3.14× (for fuzzing bugs) and 90.9x (for real-world bugs) smaller than the existing half-automated reduction tool, Polygraphy by NVIDIA. NNReduce also improves error-message-based deduplication by achieving up to a 2.36x smaller distinct set of bug reports.

Graduation Semester

2024-05

Type of Resource

Text

Handle URL

https://hdl.handle.net/2142/124525

Copyright and License Information

Copyright 2024 Jiawei Liu

Owning Collections