llama.cpp b10326 fixes TTS timing with vocoder pass measurement

llama.cpp, the wildly popular open-source inference engine for LLaMA models, just dropped release b10326. The headline change is a TTS (text-to-speech) timing fix: the team adjusted the timings line to account for the vocoder pass. Previously, the pipeline deferred waveform generation work to a separate get_output run, leading to inflated or inconsistent time reports. Now the reported total time and audio-to-processing ratio honestly reflect the full vocoder workload, helping developers benchmark TTS performance more reliably.

Beyond the fix, this release is notable for its sheer distribution breadth. The b10326 tag includes prebuilt artifacts for macOS (Apple Silicon, Intel, and a KleidiAI-optimized arm64 build), iOS, Linux across x64/arm64/s390x with CPU, Vulkan, ROCm 7.2, OpenVINO, and SYCL variants, plus Android, Windows (CPU, CUDA 12 and 13, Vulkan, OpenVINO, SYCL, HIP), and OpenEuler builds. With 123k stars and 21.4k forks, llama.cpp remains the go-to choice for developers running LLMs on edge devices and data centers alike. This update is purely incremental—no new model support—but the TTS timing correction improves trust in telemetry, making it easier to profile audio generation pipelines and compare hardware.