#optimization

Speculative Decoding: How Draft-Model Parallelism Cuts LLM Inference Latency