llama.cpp b9587 fixes speculative decoding logging for ngram-map-k4v
A minor logging fix in the popular LLM inference engine goes live.
The open-source LLM inference engine llama.cpp has released version b9587, tagged on June 10. This release introduces a single non-functional change: a fix for the speculative decoding logging. When using the speculative type `ngram-map-k4v`, the log messages at startup and runtime were incorrectly displaying `ngram-map-k`. The commit adds logic in the constructor of `common_speculative_impl_ngram_map_k` to correctly pass `COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V` when `config.key_only` is false. This ensures that logging accurately reflects the selected speculative mode.
Alongside the fix, the release provides pre-built binaries and configurations for a wide range of platforms: macOS (Apple Silicon arm64 with and without KleidiAI, Intel x64, iOS XCFramework), Linux (Ubuntu x64, arm64, s390x, plus GPU variants with Vulkan, ROCm 7.2, OpenVINO, and SYCL FP32), Android (arm64), and Windows (x64 CPU, arm64 CPU, CUDA 12, CUDA 13, Vulkan, SYCL, HIP). Some configurations are disabled (e.g., macOS KleidiAI, Linux SYCL FP32, openEuler builds). The release was automatically published by GitHub Actions and is signed with a verified GPG key.
- Fixes logging for speculative decoding mode ngram-map-k4v (was incorrectly logged as ngram-map-k)
- Non-functional change – no impact on performance or model output
- Available for macOS, Linux, Android, Windows across CPU and multiple GPU backends (CUDA, Vulkan, ROCm, etc.)
Why It Matters
Maintains accuracy in critical logs for developers debugging speculative decoding workflows in llama.cpp.