Add sparse MLA backward (DSA top-k), sliding-window attention + attention sink, and block-sparse MLA forward/backward for SM100 - #198

Open

kmswin1 wants to merge 3 commits into

deepseek-ai:mainfrom

kmswin1:feat/mla-training-extensions

Commits on Jul 16, 2026

Add sparse MLA backward kernel for SM100 (per-token top-k, DSA)
kmswin1
committed
8c37910
Add sliding-window attention and attention sink to dense MLA
kmswin1
committed
b86c615
Add block-sparse MLA forward and KV-outer backward (192/128 training path)
kmswin1
committed
091c15e