Fine-Tuning Qwen2-VL for Blockchain Graph Classification on AMD MI300X: What the Docs Don't Tell You

Short summary

A detailed development log on fine-tuning Qwen2-VL (7B) for blockchain transaction graph classification on AMD MI300X with ROCm 6.x. The author argues that rendered graph images preserve topology signals (hub-and-spoke patterns, mixing layers) that text serialization loses, making vision-language models a practical alternative to GNNs. The real value is in documenting ROCm-specific friction points — flash attention build issues, silent CPU fallbacks, and environment variable workarounds — that CUDA-centric documentation completely omits.

• •Blockchain transaction graphs rendered as images preserve topology signals that text serialization loses
• •Qwen2-VL chosen over GNNs for practical reasons: open weights, working fine-tuning toolchain, native high-res image input
• •ROCm on MI300X has significant documentation gaps vs CUDA — silent CPU fallbacks and flash attention build friction are the main blockers

Generated with AI, which can make mistakes.