FP6-LLM

Summary from the TLDR AI newsletter

Reducing floating point precision speeds up training but can be unstable. This paper shows that you can get full tensor core utilization in a new packing format without common instabilities or slowdowns from naive attempts.