b8245

The open-source project Llama.cpp, maintained by the ggml-org team, has rolled out a new release tagged b8245. This update primarily addresses a specific server-side bug (issue #20232) related to the creation of checkpoints immediately after processing mtmd (mixture of experts) chunks. The fix prevents potential instability or data corruption during long-running inference sessions, making the server component more robust for developers deploying local language models.

Beyond the core fix, the release significantly expands the availability of pre-compiled binaries, lowering the barrier to entry. New builds for Windows now include versions for CUDA 12.4 and CUDA 13.1 DLLs, as well as a Vulkan backend, joining the existing CPU, CUDA, SYCL, and HIP options. This broadens hardware support for users with NVIDIA GPUs and those leveraging alternative graphics APIs. The release also maintains its comprehensive suite of builds for macOS (Apple Silicon and Intel), Linux (CPU, Vulkan, ROCm), iOS, and openEuler, cementing Llama.cpp's position as a cross-platform powerhouse for efficient LLM inference.