The llama.cpp project released version b9885, which includes a critical fix for the AIX platform by enabling tiled matrix multiplication. This change reduces the local stack buffer footprint to prevent segmentation faults caused by large buffers in the matmul_tiled path.

  • Performance improvements of approximately 2x in PP_Speed were observed for FP32, Q4_0, and Q8_0 models on AIX.
  • The release provides binaries for macOS (Apple Silicon and Intel), Linux (Ubuntu x64, arm64, s390x), Windows (CPU, CUDA, Vulkan, ROCm, etc.), Android, and openEuler.
  • KleidiAI support for macOS Apple Silicon is disabled in this build.

This update ensures stability on AIX systems while delivering significant inference speedups for supported quantization formats.