New method boosts bug detection in tensor kernels for AI computations

Researchers have developed a new method for calibrating mixed-precision tolerance in tensor kernels, which are crucial for machine learning computations. This operator-aware approach analyzes element-wise error distributions from cloud GPU runs to establish tighter, empirically justified absolute tolerances. When applied to LLM-style buggy variants, this method significantly improves bug-detection recall by over 9 percentage points, identifying more than 200 additional bugs. AI

IMPACT Improves the reliability of AI model training and inference by enhancing the accuracy of tensor kernel correctness tests.

RANK_REASON The item is an academic paper detailing a new method for calibrating tensor kernels. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• alphaXiv
• arXiv
• attention_triton
• CatalyzeX Code Finder for Papers
• Connected Papers
• CORE Recommender
• DagsHub
• Gotit.pub
• gpuemu
• half-precision floating-point format
• Hugging Face
• IArxiv Recommender
• Influence Flower
• Litmaps
• ScienceCast
• scite Smart Citations

• paper
• infra

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →