The publication claims cuDNN, FA4 and FlexAttention have precision problems in backward passes and presents KohakuFA as an alternative kernel, with a public GitHub repository for inspection.
Radar, October 3, 2026. The publication claims that the backward passes of cuDNN, FA4 and FlexAttention have precision problems, and presents KohakuFA as an alternative kernel. The text links the project's GitHub repository. The source text gives no benchmark numbers, detailed methodology or comparison results, so the precision claims should be treated as the author's assertions until they are checked.
To consult and verify, open the original repository through the link in the publication, read the kernel code, tests and reproduction instructions, and compare backward-pass numerical results against a known reference in your own configuration. Engineers assessing training stability can use this check to decide whether the topic deserves attention. Privacy is secondary for this item, so the main care concerns organizational data: if you use AI to study code or training logs, remove names, tax IDs, e-mails and other identifiers before pasting the content.