llama.cpp b9848 fixes CUDA bug for large model tables

The popular open‑source project llama.cpp, which enables running large language models locally on consumer hardware, has released version b9848. This patch addresses a critical CUDA kernel bug in the `get_rows_back` operation that caused errors when handling tables with more than 65,535 rows. By clamping the grid‑y dimension and adjusting stride, the fix ensures correct execution for large batch sizes or long contexts. The release is signed with GitHub’s verified signature and includes pre‑compiled binaries for a wide range of platforms — macOS (Apple Silicon with optional KleidiAI acceleration, Intel), Ubuntu (CPU, Vulkan, ROCm 7.2, OpenVINO, SYCL FP32/FP16), Windows (CPU, CUDA 12/13, Vulkan, OpenVINO, SYCL, HIP), Android (arm64), and even openEuler. The extensive build matrix reflects llama.cpp’s commitment to broad compatibility, from high‑end data center GPUs to mobile devices.

For developers and enthusiasts running self‑hosted LLMs on NVIDIA GPUs, this update is essential. The bug could silently produce incorrect token probabilities or cause crashes during attention‑heavy tasks. With this fix, users can reliably handle larger models and longer sequences without hitting GPU grid limits. Llama.cpp remains the go‑to tool for private, offline AI inference, and this release further solidifies its stability for production workloads. While no new features are introduced, the reliability improvement will be welcomed by the community.