Tests show vLLM achieves significantly higher generation speeds on Qwen3.6 models, with 35B-A3B reaching 156 t/s using ROCm and AITER. ROCm outperforms Vulkan in both 35B and 27B models, with speeds of ~106 t/s and ~44 t/s respectively, while Vulkan achieves ~87 t/s and ~41 t/s.
ROCm vs Vulkan vs vLLM Performance on Dual R9700s
from English