One Config Line Made My 27B Model 2.7× Faster
Deploying Qwen3.8-27B on a DGX Spark shows why dense-model decode is bandwidth-bound, and how MTP speculative decoding lifts throughput from 11 to 29 tok/s.
term
Speculative Decoding