https://www.percepta.ai/blog/can-llms-be-computers

They implemented wasm in transformer with cpu decoding for the wasm part.