Article 3 Speculative Decoding in Practice: How EAGLE3 Makes LLMs Faster Without Changing Their Outputs
