llama.cpp b10274 fixes duplicate empty audio chunks for short inputs

On August 4th, ggml-org shipped llama.cpp b10274, a maintenance release that resolves a subtle but annoying audio encoding bug. The issue, tracked as PR #26536, caused duplicate empty audio chunks to be generated for short inputs in mtmd (the multi-modal transformer decoder). This affected scenarios where brief audio snippets needed to be processed by local multimodal models, potentially producing garbled outputs or wasted compute. The fix restores correct handling of empty chunks, ensuring that only meaningful audio data is passed to the decoder. For developers running speech-to-text or audio-to-action pipelines with llama.cpp, this means cleaner outputs and more predictable performance on short utterances.

The release also demonstrates the project's incredible breadth: prebuilt binaries are available for macOS (Apple Silicon and Intel), Linux (x64, arm64, s390x, plus Vulkan, ROCm, OpenVINO, SYCL variants), Windows (CPU, CUDA 12/13, Vulkan, OpenCL Adreno, HIP), and Android (arm64 CPU). This wide support, combined with 123k stars and 21.3k forks on GitHub, makes llama.cpp one of the most important open-source projects for AI. It enables developers to run LLMs like LLaMA, Mistral, and community models entirely offline, on ordinary hardware, without cloud dependencies. The b10274 patch may be small, but its focus on audio inputs is a signal that the maintainers are hardening multi-modal support, a key trend for edge AI. For tech professionals, this is a low-risk update worth adopting to ensure their local AI stack remains stable and current.