Feedback: can't reproduce the headline numbers

[dero]

I have a dual 7900 XTX system.

Compiled the project, ran server with Qwen3.6-35B-A3B-UD-Q3_K_M.gguf and got 110 tps decode. Current llama.cpp gives me ~135 tps with Qwen3.6-35B-A3B-UD-IQ4_XS.gguf (Vulkan, single card).

Ubuntu 26.04