llama.cpp b10282 adds vLLM-aligned spec-decode metrics to /metrics

llama.cpp, the wildly popular C/C++ inference engine for LLMs, just dropped release b10282. With 123k GitHub stars and 21.4k forks, it's the go-to tool for running local models, and this update tightens its production observability story. The headline change: the /metrics endpoint now exposes spec-decode counters, and the parameters are named exactly as vLLM names them. That means teams running mixed inference stacks can finally use the same Grafana dashboards and Prometheus alerts for both engines without custom mapping.

Speculative decoding is a key speed-up technique where a small 'draft' model proposes tokens and a larger model verifies them in parallel. Until now, llama.cpp server users had no built-in way to see draft acceptance rates, verification latency, or how many tokens were accepted per batch. This release fixes that gap by exposing those counters directly through the existing metrics endpoint. The commit also includes review fixes that align parameter names precisely with vLLM, reducing confusion in mixed environments. As usual, b10282 comes with a flood of prebuilt binaries: Windows x64 with CUDA 12.4/12.13, Linux variants for Vulkan, ROCm 7.2, OpenVINO, SYCL, and more, plus brand-new openEuler ARM builds for Ascend 310p and 910b. Whether you're running local inference on a MacBook or serving at scale on GPU clusters, this release makes llama.cpp easier to monitor and tune.