llama.cpp b9935 adds Vision RoPE for multimodal models
New release supports rotary position embeddings for vision tasks
Deep Dive
llama.cpp released tag b9935 with hexagon: add VISION RoPE support, including support for strided half-dim views and non-contiguous dst across all modes. Builds available for macOS (Apple Silicon & Intel), Linux (CPU, Vulkan, ROCm, OpenVINO, SYCL), Windows (CPU, CUDA 12 & 13, Vulkan, OpenVINO, SYCL, HIP, OpenCL Adreno), Android (arm64 CPU), iOS, and more.
Key Points
- llama.cpp b9935 adds VISION RoPE support for multimodal transformer models
- Supports strided half-dim views and non-contiguous destinations for RoPE
- Available on macOS, Linux, Windows, Android, and iOS with multiple acceleration backends
Why It Matters
Enables efficient vision-language model inference on edge devices, expanding local AI capabilities.