Thomas Neff

icon indicating a website link thomasneff.github.io icon indicating an email [email protected]

icon company Luma AI icon location Graz icon location Leading the Accel team @lumalabs. PhD from Graz University of Technology (Graphics / ML / VR). Intern at @facebookresearch

Results 1 issues of Thomas Neff

CUDNN SDPA backward results in NaN outputs

1 comment

**Describe the bug** We noticed `NaN` being generated from the CUDNN SDPA backward in our training runs. After digging a bit, we narrowed it down to a minimal repro case...