Yes I agree you can tune autoregressive LLMs

You can also tune diffusion LLMs

After doing so, the diffusion LLM will be able to generate more tokens/sec during inference