llama.cpp

635ef78e - vulkan: work around Intel fp16 bug in mmq (#18814)