The ggml-org project released version b9878, which includes a fix for stale tensor-split parameters affecting draft models. This update addresses metadata issues specifically related to GQA attention within the ggml-backend-meta component.

  • Fixed tensor split metadata for GQA attention.
  • Reverted a previous commit that introduced an asset hack in ggml-backend-meta.
  • Provided binaries for macOS (Apple Silicon and Intel), iOS, Linux (CPU, Vulkan, ROCm, OpenVINO, SYCL), Android, Windows (CPU, OpenCL, CUDA 12/13, Vulkan, OpenVINO, SYCL, HIP), and openEuler platforms.

This release ensures correct parameter handling for draft models while distributing updated binaries across supported hardware architectures.