Comment by bawana

How is speculative decoding helpful if you still have to run the full model against which you check the results?