📺 AI Model Debugging: Find the First Divergence Across AI Accelerators
The presentation explains how to investigate cases where a model produces different results after being moved from a validated reference machine to a new AI accelerator. It focuses on separating expected floating-point variation from actual correctness problems and on narrowing the search before diagnosing the root cause.
■ Debugging across the software stack
- Model operations, shapes, compiler lowering and fusion, runtime scheduling and buffering, accelerator precision and layout
- Expected precision-related differences versus observable correctness issues
■ Localize before diagnosing
- Run the same controlled inputs through a validated reference and the target accelerator
- Compare intermediate tensors at selected boundaries such as layers or operators
- Identify the first observed divergence and focus on the region between the last matching checkpoint and the first failing one
■ Diagnose and prevent recurrence
- Check operation inputs, isolate the operation, and build a minimal reproduction
- Inspect precision, layout, lowering, reduction behavior, and broadcasting
- Rerun the full model and retain a regression test for CI/CD
It is intended for engineers and technical teams working with model portability, compiler and runtime behavior, or accelerator debugging. Viewers can gain a structured approach for turning a model-wide numerical mismatch into a smaller, more manageable investigation.
この動画を紹介した IBM Technology の最新動画も、紹介付きで読めます。
いつも見ているチャンネルも、同じように紹介できます。
いつも見ているチャンネルを追加する📄 このページの紹介文は AI が独自に生成したものであり、著作権をはじめとする他者の権利(商標権・名誉権・プライバシー等)を侵害しないよう配慮しています。動画の著作権は各作成者に帰属します。