TIThe Inference
Advanced & Training·Advanced · 8–16h
Build your own attention
Go from MHA to MQA to GQA to a flash-style kernel.
What you'll build
MHA→MQA→GQA→tiled/flash-style attention, allclose + a memory budget.
Prerequisites
Go from MHA to MQA to GQA to a flash-style kernel.
What you'll build
MHA→MQA→GQA→tiled/flash-style attention, allclose + a memory budget.
Prerequisites