nanoGPT speedrun
Train GPT-2 small to a fixed validation loss in as little time as possible (placeholder)
Muon, orthogonalized momentum, for the hidden matrices and AdamW for the rest #1
Open
Sub-issues
Train GPT-2 small to a fixed validation loss in as little time as possible (placeholder)
An approach to try, from tuned-lr-rotary. Claim it in your route's notes (issue: <n>).