Llama.cpp b9773 adds Vulkan GET_ROWS_BACK for faster inference

The open-source llama.cpp project, known for enabling local execution of large language models (LLMs) like LLaMA and Mistral, has released version b9773. This update introduces Vulkan support for the GET_ROWS_BACK operation, a key optimization for efficient data retrieval during inference. The feature enhances performance on Vulkan-compatible GPUs, broadening the hardware range for running LLMs beyond CUDA or Apple Metal. The release is accompanied by pre-built binaries for multiple platforms, including Ubuntu (x64, ARM64, s390x), Windows (x64, ARM64), macOS (Apple Silicon and Intel), and Android ARM64. Specialized builds are also available for Vulkan, CUDA 12 and 13, ROCm 7.2, OpenVINO, SYCL, and HIP.

This release underscores the active development of llama.cpp, which remains a vital tool for developers and researchers who need to run LLMs locally without cloud dependencies. The addition of GET_ROWS_BACK in Vulkan specifically improves throughput for attention mechanisms and batch operations, making it more practical for real-time applications and fine-tuning workflows. The commit is signed with a verified GPG key (B5690EEEBB952194), ensuring code integrity. As the project continues to optimize across diverse hardware backends, b9773 represents incremental but meaningful progress toward democratizing AI inference.