This only uses llama correct? So the output should be the same as if you were only using llama.cpp. Am I the only one who doesn't get nearly the same quality of

This only uses llama correct? So the output should be the same as if you were only using llama.cpp. Am I the only one who doesn't get nearly the same quality of output using a quantized model compared to GPU? Some models I tried get astounding results when running on a GPU, but create only "garbage" when running on a CPU. Even when not quantized down to 4bit llama.cpp just doesn't compare for me. Am I alone with this?