A user reports running Qwen 3.6 27B MTP model on two Radeon R9700 GPUs via llama.cpp with ROCm 7.2.1. Tests show stable decode speeds (40–67 t/s) and prefill throughput (up to 1,500 t/s for prompts under 10k tokens), with MTP draft acceptance rates between 0.33 and 0.61.
2× Radeon R9700 with Qwen 3.6 27B Q8 MTP on llama.cpp
from English