Qwen 27B runs slowly on RX 7900 XTX in local test
A firsthand local test of Qwen3.6 27B on an RX 7900 XTX produced worse speed than expected. The setup used Ubuntu 24.04.4, 6.8, llama.cpp, ROCm 7.2.4, AMD driver 6.16.13, and Vulkan. In raw backend tests, ROCm processed the prompt at 235.73 and generated at 31.14 .
Vulkan processed the prompt faster at 634.80 , but generated more slowly at 13.32 . In a real API test with 32,201 prompt tokens and 128 generated tokens, the normal 27B model averaged 238.42 prompt , 26.84 generated , and 139.8 seconds total. The MTP n=3 setup averaged 226.09 prompt , 17.14 generated , and 149.9 seconds total, even with a 78.76% draft .
vLLM was also not a practical path on this hardware. The server settings included a 65,565 , 99 GPU layers, , continuous batching, and MTP draft inference.
Key points
- Qwen3.6 27B ran slower than expected on an RX 7900 XTX local setup.
- ROCm generated faster than Vulkan, while Vulkan processed the prompt faster than ROCm.
- The normal 27B model took 139.8 seconds on the real API test.
- MTP n=3 took longer at 149.9 seconds despite a 78.76% draft .
- For , actual tokens-per-second tests matter more than the name of an optimization feature.