Kilian Haefeli

icon location Studying ML and DL at ETH Zurich | Previously Large Language Models at @AlephAlpha and Data Science at @Logitech and Cofounder @Airica

Results 2 comments of Kilian Haefeli

Flashattention

Thanks @karpathy - Yes, the mallocs are just because of how the test data here is generated and formatted. In prod, I we would use these as caches for the...

Flashattention

Thanks alot for your profiling work @ChrisDryden , very interesting results! I just pushed the new feature (for 3060 about +20%) into the codebase. In case you are profiling the...