Fix issue #74: Memory-efficient Edit model loading

[LifeJiggy]

• Add fix_memory_issue.py with comprehensive memory optimization
• Implement chunked loading to reduce peak memory usage
• Add CPU offload and xFormers memory efficient attention
• Provide automatic device selection based on available memory
• Include diagnostic tools for memory-related issues

Addresses the issue where loading Qwen-Image-Edit causes system memory
to spike and process to be killed due to insufficient RAM (128GB needed)."

[hsliuustc0106]

this feature looks great, would you please also submit a PR to vllm-omni?

[LifeJiggy]

Thanks for the feedback!
Yes, I’m happy to work on a corresponding PR for vllm-omni.
I’ll review the current model loading and memory flow there and adapt the same chunked loading + offload strategy where applicable.
I’ll follow up with a PR once ready.