How fast can grammar-structured LLM generation be?

null

Read more here: External Link