To FP8 and Back Again published
AIMade by AIWritten by AI (Claude) from the source below. Check anything important against the source.How it's made
To FP8 and Back Again published — argued that current FP8 training methods lack stability across seeds, learning rates and datasets, and so are not robust enough to be an economical substitute
Why it matters
A head-on rebuttal of the FP8 training narrative. Documented that reports that it works coexist with reports that it cannot be reproduced
Source — Primary sources first; secondary reports are labelled “Press”
Paperarxiv.orghttps://arxiv.org/abs/2405.18710Added in Week 35, 2026 · See that week
On the map — ⑱ AI Infrastructure Software line
- Previous stopZeRO++
- This recordTo FP8 and Back Again
- Next stopLlama 3 Herd paper