Llama.cpp drops new grammar fix for b10286
Llama.cpp v1.0-b10286 fixes max repetition rule with grammar engine
The Llama.cpp team (maintained under the ggml-org umbrella) has pushed v1.0-b10286, a point release focused on grammar-engine stability. The headline change is the removal of the hard 2,000-token ceiling on max repetition in grammar-constrained decoding, effectively making it unbounded—useful for long-form generation tasks that need strict adherence to grammars without artificial limits. The binary ships with a signed GitHub Actions workflow (commit cd0fa60, signed with B5690EEEBB952194), giving users cryptographic provenance on macOS, Linux, Windows, Android, and openEuler across CPU, Vulkan, ROCm 7.2, CUDA 12/13, SYCL FP32/FP16, OpenVINO, and HIP backends.
On the tooling side, UI assets have been refreshed (v27) and platform coverage expanded to include openEuler on x86 (310p, 910b) and aarch64 (310p, 910b with ACL Graph). The release disables certain configurations (e.g., macOS Intel x64, openEuler x86 310p) where support remains experimental or pending further testing.
- Llama.cpp v1.0-b10286 removes the 2,000-token max repetition limit in grammar-constrained decoding
- Supports 25+ platform/backend combinations: macOS, Linux (CPU/Vulkan/ROCm/…), Windows (CPU/Vulkan/CUDA/OpenVINO), Android arm64, and openEuler variants
- Binaries are signed via GitHub Actions; UI assets updated to v27; some configurations remain disabled
Why It Matters
Enables unbounded grammar-constrained generation across diverse hardware, reducing constraints for long-form and structured outputs.