AI BriefingContract-Grade Verifier for LLM-Generated GPU Kernels and Native Blackwell Backward for Gated-Linear-Recurrence
Key point
This proposal introduces a tool for precisely verifying the reliability of LLM-generated GPU kernels and new kernels for Blackwell.
Details
Currently, accuracy verification of GPU kernels generated by LLMs relies on single tests and loose tolerances, posing a high risk of misclassifying incorrect kernels as normal.
We propose the Contract-Grade Verifier, which rigorously validates kernel accuracy through 12 adversarial gates. Beyond simple numerical comparison, this tool precisely checks for NaN/Infinity occurrences, behavior under shape changes, and cumulative errors across precisions (fp16 vs fp32).
An audit of 2,638 kernels previously deemed normal by existing systems revealed:
- 39.5% of kernels were completely broken, exceeding tolerance limits.
- 62.1% of kernels contained at least one violation.
- This verifier flagged 1,487 kernels as errors that had passed existing standard tests.
Additionally, leveraging this verifier, we developed the first native Blackwell backward kernel for the Gated-Linear-Recurrence (GLR) family and demonstrated its accuracy using a double-precision oracle.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.