Small patch dedicated to optimizing gemma4, fixing inference with usecache=False due to k/v states sharing between layers, as well as conversion mappings for some models that woul...