Speculative decoding: when and why it actually speeds up inference

· Dev.to

Read full story at source