Architectures (like Gemma4) have things like sliding window attention that make it pretty impractical to do on a GPU, you would have to do additional things to make it work, and the talent to do that, even with LLM assistance is very small.