#speculative-decoding

DFlash vs MTP: Two Paths to Breaking the Autoregressive Decoding Bottleneck

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