1 comment

[ 3.9 ms ] story [ 12.6 ms ] thread
"DiffusionBlocks, a principled framework that partitions transformers into independently trainable blocks, reducing memory requirements proportionally while maintaining competitive performance across diverse architectures and tasks."