llama.cpp

806eee98 - vulkan: fix stale prealloc_y reuse across flash attention and soft_max (#29591)