The GPU worked—and then I discovered the CPU regression
The first GPU implementation worked. Then a CPU benchmark exposed a regression: a novel memory-management system had been replaced by seven nested loops.
Lynn Cole / 标签
A thread through the work.
The first GPU implementation worked. Then a CPU benchmark exposed a regression: a novel memory-management system had been replaced by seven nested loops.
A Go-native transformer almost works. Tokenizers, vocabulary, norms, cache initialization and attention all take their turn in a night of increasingly interesting nonsense.