llama.cpp b9585 fixes Granite speech model inference with embedding scale

The open-source large language model inference engine llama.cpp, maintained by ggml-org, has released b9585 with a critical fix for the Granite speech model. Previously, when the deepstack parameter was disabled, the model's embedding scale was not applied, leading to incorrect inference results. This release applies the embedding scale in non-deepstack scenarios, restoring accurate output for speech tasks. The fix was contributed by gabe-l-hart and co-authored by Xuan Son Nguyen from Hugging Face, ensuring the model works reliably across all supported backends.

With over 116k stars and 19.4k forks, llama.cpp is a cornerstone of the open-source AI community, enabling users to run large language models locally on commodity hardware. This fix expands the Granite speech model's accessibility, allowing it to run on CPU-only, Vulkan, and other platforms that lack deepstack support. The update is available for macOS (Apple Silicon and Intel), Linux (x64, arm64, s390x), Windows (x64 and arm64, with CUDA and Vulkan backends), iOS, and Android. For professionals self-hosting AI assistants or building on-device speech applications, this patch removes a significant roadblock to production use without proprietary accelerators.