Add cpu offload for Qwen/Qwen2-VL-72B-Instruct-AWQ

[nguyen-brat]

Hello, I want to run Qwen/Qwen2-VL-72B-Instruct-AWQ on my local computer, currently, I have 2xrtx 3090 but it has trouble OOM. Then I see in your vision.py 's options there --max-memory option to offload on CPU. Can you please implement it also for Qwen/Qwen2-VL-72B-Instruct-AWQ

[matatonic]

I'd like to do this, and for other models also, but it's awkward and vastly complicates the loading as far as I know it. I'd like to enable this and have had it back of mind for a while.