How to Fine-Tune Qwen2-VL for Blockchain Graph Classification on AMD MI300X

A developer experiment explored using Qwen2-VL, a 7B vision-language model, to classify blockchain transaction graphs by treating them as images rather than text sequences. The approach leverages the model's ability to detect visual topology patterns — such as star-shaped hub-and-wallet structures and layered mixing flows — that are difficult to reconstruct reliably from raw transaction data alone. Qwen2-VL was chosen over graph neural networks for its open weights, native high-resolution image support, and compatibility with existing fine-tuning frameworks like LLaMA-Factory and ms-swift. However, the project ran on AMD MI300X GPUs with ROCm 6.x, exposing significant documentation gaps since virtually all fine-tuning guides assume NVIDIA CUDA environments. Key ROCm-specific hurdles included flash attention compilation, PyTorch build differences, and environment variables that become critical but are absent from standard setup guides.