feat: Add DeepSeek-V4 CSA attention kernels - #13

Open

aws-zifan-he wants to merge 7 commits into

mainfrom

zifan/deepseek_v4_csa

Commits on Sep 1, 2026

create experimental kernel for deepseek v4 compressed sparse attention across 8 cores in a single trainium 3 device
aws-zifan-he
committed
9ca48fd

Commits on Sep 9, 2026

style: apply ruff format and trim verbose comments in deepseek v4 csa
aws-zifan-he
committed
d5e9467
test: add integration test for the deepseek v4 csa attention block
aws-zifan-he
committed
0fb4176

Commits on Sep 11, 2026

Commits on Sep 12, 2026

feat: sparse prefill dispatch, hadamard indexer compressor, and 2-core rms+rope grid
aws-zifan-he
committed
8e649d4

Commits on Sep 14, 2026

feat: fuse rope+hadamard into the indexer q-projection and add a packed q_b prefill kernel
aws-zifan-he
committed
2cbd299