Add Blelloch parallel prefix scan for LASP#2

Open

petrpan26 wants to merge 6 commits into

OpenNLPLab:mainfrom

petrpan26:feature/blelloch-clean

Commits on Nov 4, 2025

Add Blelloch parallel prefix scan for LASP
Hoang Phan
committed
Fix Blelloch exclusive scan: avoid numerical instability
Hoang Phan
committed
Fix suffix scan rank shift: reverse communication direction
Hoang Phan
committed
Fix rank shift deadlock: use non-blocking communication
Hoang Phan
committed
Add more gpu tuning for other gpus
Hoang Phan
committed
Optimize memory usage
Hoang Phan
committed