The llama.cpp project has released version b10025, which includes a change to relax tensor contiguity requirements for quantized concatenation operations.

  • The core update modifies the CUDA backend and the underlying ggml library to allow non-contiguous tensors during quantized concat.
  • Test cases have been added to verify the functionality with non-contiguous quantized concat inputs.
  • Binaries are provided for macOS (Apple Silicon and Intel), Linux (CPU, Vulkan, ROCm, OpenVINO, SYCL), Android, Windows (CPU, OpenCL, CUDA 12/13, Vulkan, OpenVINO, SYCL, HIP), and openEuler.

This release updates the build environment and fixes specific tensor handling constraints for quantized operations.