Even a 4-bit quant version of the MoE 8x22 is going eat ~80GB of VRAM. No way you're running this on a 4090 without setting it up as a GGUF to split between VRAM and regular RAM, and then you're going to have to deal with the low token rate as a result.
Running WizardLM-2 70B or lower WizardLM-2 7B is much more feasible however.
Look into Ollama: