llama.cpp fuses DeepSeek-V4 ops eating 32% of Vulkan decode time

llama.cpp's Vulkan backend closed its last major gap in DeepSeek-V4 decoding speed. On DeepSeek-V4-Flash, the unfused hyper-connection math was burning about 32% of decode op time on gfx1151, the GPU inside AMD's Strix Halo APUs, spread across roughly 16,000 dispatches per token.

llama.cpp PR #26578 adds three fused kernels, DSV4_HC_COMB, DSV4_HC_PRE and DSV4_HC_POST, for the Sinkhorn comb chain DeepSeek-V4 runs during decode. Vulkan was the last major llama.cpp backend still running that chain unfused, one dispatch at a time.

The fused shader runs the full 20-iteration Sinkhorn algorithm inside GPU registers, using subgroupShuffleXor to move data within 16-lane blocks, skipping the round trip to memory for intermediate results. What used to take about 137 strictly ordered node executions per site now runs as a single dispatch. The author checked the output against a float64 reference and against DeepSeek's own modeling_deepseek_v4.py code, matching to about 6e-7.

Strix Halo owners running DeepSeek-V4 through llama.cpp's Vulkan backend should update. It doesn't change what the model outputs, only how long the wait is for it.