Llama.cpp b9105 fixes CUDA stability with direct iterator inclusion

Llama.cpp, the highly popular open-source C++ implementation for running large language models locally, has released version b9105. With over 110,000 stars and 18,100 forks on GitHub, this project is a cornerstone for developers seeking efficient local inference across diverse hardware. The release addresses a subtle but critical CUDA issue: previously, the build relied on a transient import from cub/cub.cuh to access cuda/iterator. This practice was fragile because cub does not consistently expose that header, leading to compilation failures or runtime instability on certain CUDA configurations. By directly including cuda/iterator, the fix ensures reliable GPU-accelerated inference, particularly for those using custom build pipelines or newer CUDA toolkits.

The b9105 release maintains llama.cpp's reputation for broad platform support. Prebuilt binaries are available for macOS Apple Silicon (both standard and KleidiAI-enabled), Linux (x64, arm64, s390x with Vulkan, ROCm 7.2, OpenVINO, and SYCL FP32/FP16), Android arm64, Windows (x64 and arm64 CPU builds, plus CUDA 12.4/13.1 DLLs, Vulkan, and HIP), and openEuler (x86 and aarch64 with ACL Graph). This comprehensive coverage means developers running LLMs on everything from gaming PCs to cloud VMs to edge devices can immediately benefit from the fix. The release also signals ongoing maintenance of a project that has become essential for the local AI community, where stability and performance are paramount.