Llama.cpp b9180: MTP support landed (github.com)

The latest update to Llama.cpp, release b9180, has introduced significant support for Model Training Protocol (MTP), enhancing the processing capabilities of the Llama framework. This update includes a host of technical improvements such as batch size adjustments, the introduction of a partial rollback feature for Gated Delta Networks (GDN), and optimizations for various platforms including macOS, Linux, Windows, and Android. Notably, the rollout allows models to store intermediate states during training, which can prevent the waste of computational resources during speculative decoding, significantly improving efficiency. This development is vital for the AI/ML community as it simplifies processes related to model checkpointing and enhances compatibility with existing architectures. By allowing models to revert to earlier states during training, developers can streamline the debugging process and save valuable time and computational power. Furthermore, improvements in documentations and compatibility checks with n-gram models ensure that Llama maintains a competitive edge in the rapidly evolving landscape of machine learning frameworks. Overall, this update reinforces Llama's commitment to facilitating cutting-edge AI research and development.