llama.cpp b10216 brings Vulkan POOL_1D support for edge inference

llama.cpp, the popular open-source C++ library for running LLMs locally, released build b10216. The headline feature is full support for GGML_OP_POOL_1D in the Vulkan backend. This introduces a new pool1d compute shader (pool1d.comp), pipeline fields, and push constants, mirroring the existing 2D pooling implementation. It also fixed a shader logic bug where tensor dimension indices and average pooling scale were incorrect, plus a crash when the pooling boundary was negative (when k0 < p0).

The release is significant because Vulkan is a low-overhead, cross-platform GPU API that works across Windows, Linux, Android, and macOS (via MoltenVK). Adding 1D pooling expands the types of models llama.cpp can run efficiently on Vulkan-capable hardware—important for edge deployments and devices without CUDA or ROCm. The project's release assets now include a wide range of prebuilt binaries: Ubuntu x64/arm64 with Vulkan, Windows x64 with CUDA 12/13, Android arm64, and even s390x. This b10216 update continues llama.cpp's strategy of broadening GPU backend coverage, making local AI inference more accessible across modest and mobile GPUs.