unsloth

/

Magistral-Small-2506-GGUF

like 68

Follow

Unsloth AI 6.34k

Running with recommended parameters (sans -ngl 99) on llama.cpp results in 85% GPU saturation tops. While Qwen3, Gemma3 and pretty much any other model use >= 97%

Is there something special about its architecture that might prevent full GPU utilisation? I am using ROCm backend.