llama.cpp releases b10267 with speculative decoding refactor

ggml-org's llama.cpp, the wildly popular open-source LLM inference library with 123k stars, shipped release b10267 on August 4. This update is focused on code hygiene rather than new capability: it refactors common_speculative_init, the helper that sets up speculative decoding configurations, to reduce duplication. Speculative decoding is a technique that uses a small, fast draft model to propose tokens, which are then verified in parallel by the larger target model, speeding up inference by 2-3x without changing output quality.

The refactor consolidates the initialization logic when speculative decoding is enabled, making the codebase easier to maintain and extend. Notably, the maintainers explicitly stated that no tests were added because the existing server tests (unit/test_speculative.py) still pass after the change, confirming the behavior is unchanged. The release also bundles prebuilt artifacts for a wide range of platforms, including macOS Apple Silicon (with optional KleidiAI acceleration), Linux variants with Vulkan, ROCm, OpenVINO, and SYCL, plus Windows CUDA 12/13 and ARM builds. For developers, this means a cleaner foundation for future speculative decoding features, while end users get a stable, drop-in update.