TensorBound study maps tensor binding failures in AI stacks
Original title: TensorBound: Characterizing Tensor Binding Failures in AI Compiler and Runtime Stacks
Of 149 reconstructed repaired causes in IREE, XLA, and PyTorch, 69.1% were missing checks, bad mappings, or lost relations.
TensorBound is an empirical study of failures that break bindings between tensor values and their operations, storage, or execution state in IREE and the XLA and PyTorch ecosystems. The authors analyzed 264 reports and reconstructed 149 distinct repaired causes by tracing sources and comparing pre-fix and repaired revisions. Missing eligibility checks, incorrect mappings or translations, and lost or omitted relations accounted for 69.1% of those causes. They identify 101 first invalid decisions in compiler code as candidate checking points before execution.
Unconfirmed: Cause shares come from the authors’ reconstructions of archived reports, not a live incident.
- The study analyzed 264 reports and reconstructed 149 repaired causes.
- Four archived IREE diagnostic attempts separated effect localization from cause explanation.
- Authors argue for checks that preserve binding facts across representations.
From the original article
1 min read at arxiv.orgAI compiler and runtime stacks must preserve bindings between tensor values and their operations, storage, and execution state. These bindings can fail while the underlying objects remain present, with symptoms appearing in another layer. We present TensorBound, an empirical study of tensor binding failures in IREE and the XLA and PyTorch ecosystems.
By Wen Li · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.