perf(prims-ts): Optimize&refine PrimsTS block sparse attention - #5002

Merged

qsang-nv merged 12 commits into

flashinfer-ai:mainfrom

heyuhhh:yuhangh/primts-proxy-perf2

Sep 9, 2026

Commits on Sep 7, 2026

perf(prims-ts): stream the KV256 P fragments from one rolled loop
heyuhhh
committed
5d8d79a
perf(prims-ts): let proxy routes use the shared scheduler selection
heyuhhh
committed
f5b4a6f
perf(prims-ts): trim the KV256 per-tile fixed costs
heyuhhh
committed
b1c737f
perf(prims-ts): rebalance KV256 registers and free the K/V ring from the tail
heyuhhh
committed
02f83c0
refactor(prims-ts): consolidate the KV256 fragment and tail paths and the shared dense/sparse cadence
heyuhhh
committed
58107d1
perf(prims-ts): stream Q128/KV128 P as K32 fragments
heyuhhh
committed
c7b42e4
refactor(prims-ts): one streamed Keeps max pass for dense and block-sparse tiles
heyuhhh
committed
37f351c
perf(prims-ts): prepare structural score words for dense Keeps plans
heyuhhh
committed
c2557b6
perf(prims-ts): prefetch block-sparse route records one iteration ahead
heyuhhh
committed
aea0b1e
perf(prims-ts): keep dense Q128/KV128 on the complete-row P path
heyuhhh
committed
ee119fe
fix(prims-ts): size KV256 split-KV partial columns by the partial width
heyuhhh
committed
679bf00
test(prims-ts): keep the proxy_bk128 tail block in range
heyuhhh
committed
f2f7cf5