EmbeddingGemma 2: an open, lightweight multimodal embedding model
deepmind.google·6d ago
AI Engineering
Lobsters·Aug 26·5 min read
Catching NaN at the MLIR Pass Boundary
TL;DR
A method to catch not-a-number (NaN) values at the MLIR pass boundary was developed, preventing runtime accuracy regressions. This approach ensures that invalid values are identified early in the compilation process.
✦ Why It Matters
Implement pass-boundary verification in your MLIR workflows to catch NaN values early and improve model accuracy.
Key Takeaways
- Implement pass-boundary verification to catch NaN values early.
- Use targeted flags to identify problematic passes in MLIR.
- Employ local operation verifiers to enforce numeric contracts.
- Utilize mlir-reduce to minimize valid inputs triggering NaN.
How It Works
The method involves tracing the first occurrence of NaN values through the MLIR pass pipeline. By implementing local operation verifiers, the compiler can check the validity of operation attributes immediately after each pass.
This ensures that any NaN values are caught before they propagate to runtime, thus preventing accuracy regressions in machine learning models.