llama.cpp adds a Vulkan top-k radix sort shader for k >= 1024, targeted at Qwen 3.8 Flash Next

GitHub·medium signalXBlueskyLinkedInCopy link

Build b10712, released 2026-08-31, adds a Vulkan radix-sort top-k path for k >= 1024 plus top-k QSA fusion, with tests written specifically against Qwen 3.8 Flash Next (#28032). Large-k sampling was the bottleneck for that model's sparse-attention design on Vulkan backends, so this is the runtime catching up to a model architecture rather than a generic optimization. Build b10714 the same morning tunes RDNA3 mat-vec to a static 4 rows above four columns on Strix Halo, which benched faster than the defaults across types and batch sizes (#27909).