Announcement
A claimed 9.65-second NanoGPT training run awaits review
The contributor says they adapted CPU-based work from an earlier pull request and submitted their own for review.
TLDR
A NanoGPT contributor says they trained the model in 9.65 seconds by adapting earlier work that ran a CPU model concurrently. They say that work cut the reported record from 39.9 to 21.5 seconds. The contributor submitted a pull request for the new run, but says all the records remain subject to review and CPU use during runs is a gray area for nanoGPT.
Combined views
35.5K
4 Sources, first seen ago