Can I reproduce this paper using 4x RTX 4090 cards?

[yifan0414]

Or perform fine-tuning for specific domains

[chenjoya]

Use smaller VLM, e.g. Qwen2.5VL-2B, with LoRA, I think you can successfully tune