Fabulous stuff!

Oh my request … the vision head on the Gemma models is super slow on CPU inferencing (and via Vulcan), even via llama.cpp. Any chance your team can figure out a solve? Other ViTs don’t have the same problem.