Qwen 3.8 with 8GB Ram

#GPU poor people: 30 tok/s with an AMD Radeon RX 7600 8GB on Qwen3.8-35B-A3B. Sauce:

llama-server -m Qwen3.8-35B-A3B-IQ4_XS.gguf -ngl 99 —n-cpu-moe 36 -c 65535 -b 1024 -ub 256 —flash-attn on -lm none —fit off —spec-type draft-mtp —spec-draft-n-max 2

Using vulkan llama.cpp