b9859
Breaking Changes📦 llama-cppView on GitHub →
⚠ 1 breaking✨ 2 features🐛 1 fixes🔧 4 symbols
Summary
This release introduces the capability to load precompiled binary kernels for the OpenCL backend, resolving a dependency cycle in the process. It also provides numerous precompiled binaries across different operating systems and hardware configurations.
⚠️ Breaking Changes
- The OpenCL backend now loads precompiled binary kernels from a library. This required adding `libdl.h` to break a cyclic dependency between `ggml-opencl` and `ggml-backend-dl` (which depends on backend libs). If you rely on specific kernel loading behavior, ensure the new library loading mechanism works correctly in your setup.
Migration Steps
- If you are using the OpenCL backend, be aware that kernel loading behavior has changed to use precompiled binary libraries.
✨ New Features
- OpenCL backend now allows loading precompiled binary kernels from a library.
- Added support for loading `gemm_moe_mxfp4_f32_ns`, `q8_0 gemm`, `q4_0 moe gemm`, `q4_1 moe gemm`, and `q4_k moe gemm` from the kernel library if available on OpenCL.
🐛 Bug Fixes
- Fixed a cyclic dependency issue in the OpenCL backend by adding `libdl.h`.