User benchmarks speculative decoding in llama.cpp on Qwen 3.

On an RTX 6000 PRO, DFlash achieves 3.7x speedup and n-gram stack achieves 6x speedup on coding tasks. MTP achieves 2.7x speedup, demonstrating significant inference gains via speculative decoding techniques.