CYBER WATCH / EQUINOXE INTELLIGENCE
← All articles
b12250: vulkan: add tiled-transpose concat shader for Qwen models (#30591)
llama.cpp Releases · 2026-11-10 20:01:17
The commit adds a Vulkan shader that performs tiled transpose for dim‑0 concatenation when the source is transposed. By staging a 32×32 tile through shared memory, both load and store remain coalesced, eliminating bank conflicts and improving cache locality. The new kernel is gated to 4‑byte unquantized elements, contiguous src0/dst rows, and a truly transposed src1. On a gfx1151/RADV GPU, the concat operation drops from 11,877 µs to 957 µs, and text generation throughput improves. The shader is enabled by default, with smaller sizes and cleaned‑up comments.Read original source ↗