Llama.cpp b9781 improves Vulkan to prevent GPU timeouts

The ggml-org llama.cpp project has released version b9781, focused on improving Vulkan GPU backend reliability. The key change allows reducing the number of graph submission batches, which prevents GPU timeouts during long-running or complex inference tasks. This is particularly useful for users running large language models on Vulkan-compatible hardware, as timeouts could previously crash or hang the application. The patch was merged with a verified GPG signature and is part of an ongoing effort to make local LLM inference more robust across diverse hardware.

This release also reaffirms llama.cpp's commitment to broad platform support. Build artifacts are available for macOS Apple Silicon (both standard and KleidiAI-enabled), macOS Intel, iOS, Linux (x64 and ARM64 with CPU or Vulkan), Ubuntu with ROCm 7.2, OpenVINO, SYCL FP32/FP16, Android ARM64, and Windows (x64/ARM64 CPU, CUDA 12/13, Vulkan, OpenVINO, SYCL, and HIP). The project now has over 118k stars on GitHub and is widely used for running LLMs locally. This update ensures that Vulkan users experience fewer interruptions and more reliable performance.