The llama.cpp project released version b9890, which includes significant backend refactoring and bug fixes. The update removes the -sm row for CUDA and refactors cuBLAS integration to improve stability.

  • Fixes CDNA + BF16 logic errors
  • Corrects bad return values in operations
  • Resolves issues with src0 strides and contiguous memory requirements
  • Fixes GGML_CUDA_FORCE_CUBLAS configuration
  • Corrects casts to BF16 format

This release provides updated binaries for macOS, Linux, Windows, Android, and openEuler across CPU, GPU, and various accelerator backends.