nanoGPT speedrun
Train GPT-2 small to a fixed validation loss in as little time as possible (placeholder)
FlexAttention with a growing sliding window #4
Open
Sub-issues
Train GPT-2 small to a fixed validation loss in as little time as possible (placeholder)
An approach to try, from unet-value-embeds. Claim it in your route's notes (issue: <n>).