llama.cpp Release b10887 Adds q4_0 Repack Support for IBM s390x Architectures
llama.cpp build b10887 has been released, introducing memory repack support for q4_0 quantized weights on the IBM s390x CPU architecture. Contributed by IBM developer Aaron Teo via PR #28667, the update also cleans up internal code comments in the GGML CPU backend. This release expands large language model (LLM) inference capabilities to enterprise IBM mainframes running s390x Linux. It enables enterprise infrastructure running mainframe hardware to execute quantized AI models more efficiently directly on CPU resources. The change specifically targets the `ggml-cpu` backend for s390x, implementing memory repacking for the 4-bit q4_0 quantization format. Official pre-built binaries for Ubuntu s390x CPU execution are provided alongside standard builds for x64, ARM64, CUDA, Vulkan, ROCm, and SYCL environments.
## BACKGROUND
llama.cpp is a high-performance open-source LLM inference framework built on top of the GGML tensor library, designed to run models across diverse hardware platforms. IBM s390x is a 64-bit mainframe architecture primarily used in enterprise IBM zSystems for mission-critical operations. Quantization techniques like q4_0 reduce LLM memory footprints by storing weights in 4-bit representations, where repacking reorganizes memory layouts to optimize vector CPU instruction performance.