NotesDraft
Close readings of single papers, mine and other people's. What the paper claims, what it actually shows, and the part that only makes sense once you have tried to reimplement it.
A plain-language walkthrough of our NCAA 2026 paper. Why we expected FSDP to win everywhere, why it did not, and the one plot that changed how we set up every experiment afterwards.
Aug 4, 2026
The paper is usually summarised as a fused kernel. That undersells it. The real contribution is choosing to recompute rather than store, and the arithmetic that makes it come out ahead.
Jul 21, 2026
Low rank adaptation is well explained. Where to put the adapters, and why rank matters less than placement, is left almost entirely to folklore.
Jun 30, 2026