[QSA] Tune sm120 sparse prefill geometry - #37275

Draft

Dev-Jahn wants to merge 12 commits into

sgl-project:mainfrom

Dev-Jahn:task/qsa-prefill

Commits on Sep 17, 2026

[QSA] Add direct-paged Triton sparse decode on sm120
Dev-Jahn
committed
7a98e66
[sm120] Fix serving fallbacks and graph capture
Dev-Jahn
committed
270edf3
[MoE] Add tuned fused MoE configs for RTX PRO 6000 Server Edition
Dev-Jahn
committed
aa9f9d8
[QSA] Add split-K sparse decode for small batches
Dev-Jahn
committed
2739594
[QSA] Add Triton MQA scoring on sm120
Dev-Jahn
committed
2cb4fb4
[QSA] Fuse split-K merge into sparse decode
Dev-Jahn
committed
5704af5
[PLE] Interleave the pinned table across NUMA nodes
Dev-Jahn
committed
61aef49
[HC] Tune sm120 Triton geometry by row bucket
Dev-Jahn
committed
e67f044
[MoE] Add fused small-row NVFP4 kernel for sm120
Dev-Jahn
committed
87a15dc
[QSA] Bound sparse indices and use 64-bit KV addresses
Dev-Jahn
committed
defe99e
[PLE] Select the prefetch buffer from capture state
Dev-Jahn
committed
689225c
[QSA] Tune sm120 sparse prefill geometry
Dev-Jahn
committed
1835488