llama.cpp Release b11460 Adds IQ3_S Multi-Column MMVQ to SYCL Backend
llama.cpp release b11460 adds support for IQ3_S multi-column Matrix-Vector Multiplication Quantized (MMVQ) kernels to the SYCL backend. The release also updates standard prebuilt binaries across supported platforms including Linux, Windows, macOS, Android, and Snapdragon. This update improves inference performance for 3-bit IQ3_S quantized models on hardware architectures accelerated via SYCL, such as Intel GPUs. It continues the ongoing effort to optimize low-bit quantization performance across non-CUDA hardware backends. The release integrates PR #29500, which extends the SYCL backend's MMVQ execution path to handle multi-column matrix-vector operations for IQ3_S quantized tensors. Prebuilt binaries continue to support various acceleration runtimes, including CUDA 12/13, ROCm, Vulkan, OpenVINO, and SYCL.
## BACKGROUND
llama.cpp is a widely used C/C++ library designed for fast, local LLM inference across diverse hardware architectures. SYCL is an open standard spearheaded by the Khronos Group for heterogeneous programming, commonly used to accelerate compute workloads on Intel hardware via oneAPI. IQ3_S is an importance-matrix quantization scheme (i-quant) in llama.cpp that compresses model weights to roughly 3 bits per parameter while maintaining higher output precision.