What's been cooking — August 2026

1 merged PR across 1 repo

A quiet month by PR count, but the one thing that shipped is the kind of surgical perf fix that makes you wince at the previous code and then feel better about the world. All the action was in vector search land.

NVIDIA/cuvs

The lone merge, cuvs#2321, took a scalpel to brute_force_search_filtered, which had been picking between an SDDMM-based sparse path and a dense post-filter path based on a single hardcoded threshold of sparsity < 0.9f. In other words: run SDDMM whenever 10% or fewer rows survive the filter, and hope for the best. The trouble is that SDDMM's cost scales with the number of passing rows, so that one magic number was wrong across most of the interesting regimes. The PR retunes the dispatch thresholds and adds a scatter-gather path so the sparse branch stops doing dumb things when the filter is only mildly selective. Net result: filtered brute-force search stops falling off a cliff in the middle of its own decision tree.

That's the month. Small diff, big "why was it ever like that" energy — see you in September.