Jia-Bin Huang(@jbhuang0604):One thing I hoped to convey in the video is how coherent/similar these methods are. • MQA is just MHA, but forces sharing KV matrices. • GQA is just MQA, but changes 1 -> n_g. • MLA is just GQA, but learns the up-proj matrices. • DSA is just MLA, but adds top-k selection.

Jia-Bin Huang

@jbhuang0604

I am the one who wears a jacket.

가입 May 2015

267 팔로잉 중 80.9K 팬

Jia-Bin Huang@jbhuang0604

2025.12.02 03:42

One thing I hoped to convey in the video is how coherent/similar these methods are. • MQA is just MHA, but forces sharing KV matrices. • GQA is just MQA, but changes 1 -> n_g. • MLA is just GQA, but learns the up-proj matrices. • DSA is just MLA, but adds top-k selection.