llama.cpp

8960efd0 - Vulkan: Add f32 accumulator support to quantized mul mat to fix GLM4 32B incoherence (#13607)