llama.cpp b10077 fixes OpenVINO backend with missing DL_IMPL macro
A critical patch for OpenVINO users in llama.cpp's latest release b10077
The llama.cpp project, led by ggml-org, has released version b10077 on 21 July, addressing a critical bug in the OpenVINO backend. The release adds the missing GGML_BACKEND_DL_IMPL() macro invocation, which was absent compared to other backends like Vulkan or CUDA. Without this macro, the OpenVINO backend could not properly register itself with the dynamic loading system, causing failures when users attempted to run models on Intel CPUs or integrated GPUs via OpenVINO.
This fix is essential for professionals deploying large language models (LLMs) on Intel hardware, especially in enterprise environments where OpenVINO is used for optimized inference across CPUs and accelerators. The release also includes a verified GPG signature for security and provides pre-built binaries for macOS (Apple Silicon and Intel), Linux (multiple architectures and compute backends), Windows (CPU, CUDA, Vulkan, OpenVINO, and more), and Android.
Benchmarks or performance numbers were not included in the release notes, but the fix restores full functionality for the OpenVINO path. Users who were previously unable to use llama.cpp with OpenVINO due to issue #25586 should now be able to run models like Llama, Mistral, or CodeLlama on Intel hardware without workarounds.
- Version b10077 adds GGML_BACKEND_DL_IMPL() macro for OpenVINO backend, fixing issue #25586
- Release includes verified GPG signature and pre-built binaries for macOS, Linux, Windows, and Android
- Supports multiple compute backends: CPU, Vulkan, ROCm 7.2, SYCL, CUDA 12/13, and OpenVINO itself
Why It Matters
Essential stability fix for llama.cpp users deploying LLMs on Intel hardware via OpenVINO backend