Vamana build optimization set and fp16 support#2264

Open

bkarsin wants to merge 28 commits into

NVIDIA:mainfrom

bkarsin:vamana-build-opt

Commits on Jun 15, 2026

Initial multi-warp greedy search optimization
bkarsin
committed
Add fp16 support for Vamana build/serialize
bkarsin
committed
Fix recall issue with multi-warp optimization
bkarsin
committed
Speed up fp16 vamana build with promoted query vectors and optimized L2 comparators.
bkarsin
committed
perf(vamana): P1_cache_candidate_vector — Cache the accepted candidate vector in shared memory in the RobustPrune occlusion loop
bkarsin
committed
perf(vamana): O1_parallel_scan_buffer_pool — Replace single-thread prefix sum with cub scan and hoist per-batch reverse-edge allocations
bkarsin
committed
perf(vamana): N3_multiwarp_robust_prune — Parallelize RobustPrune occlusion across multiple warps per query (raise occupancy)
bkarsin
committed
perf(vamana): N7_reuse_search_distances_in_prune — Reuse GreedySearch query->existing-edge distances in the RobustPrune merge (avoid recompute)
bkarsin
committed

Commits on Jun 16, 2026

Commits on Jul 15, 2026

Merge branch 'main' into vamana-build-opt
cjnolet
authored

Commits on Jul 17, 2026

Commits on Jul 24, 2026

Change distribution of fp16 vamana unit tests
bkarsin
committed