llama.cpp Release b11126 Adds Vulkan IQ4_XS Quantization Kernels
Release b11126 of llama.cpp introduces optimized MMQ and MMV matrix multiplication kernels for IQ4_XS quantization within the Vulkan backend. Additionally, it addresses a 27.3% text generation performance regression on Intel GPUs by selectively disabling MMVQ for IQ4_XS on Intel hardware. This update enhances cross-platform GPU inference speed for users executing low-bit quantized models via Vulkan. It also maintains consistent performance stability across non-Nvidia GPUs, particularly Intel Arc graphics cards. The release removes redundant shared-memory branch logic for IQ4_XS in Vulkan GLSL shaders and refines memory vector handling. The optimizations were developed with code assistance from OpenAI Codex and Claude Opus 5.
## BACKGROUND
llama.cpp is a popular open-source library that enables efficient, local inference of Large Language Models across various CPU and GPU architectures. IQ4_XS is a 4-bit non-linear quantization format that compresses LLM weights to approximately 4.25 bits per weight to reduce memory footprint. Vulkan provides a cross-platform API for GPU compute, allowing hardware acceleration on GPUs from AMD, Intel, and mobile vendors.