Rockchip rknn-llm demo for Qwen2-VL-2B-Instruct runs only on transformers 4.51.3

The root of the trouble is that rknn-llm is not internally consistent. The export script follows an older contract, reading model.visual directly, while the quantization script is already written for the 5.x module layout (model.model.language_model). Each generation shows why: 5.8.0 moved visual to model.model.visual, 4.57.0 made VisionAttention unpack position_embeddings unconditionally, and 4.54.x had already removed the fallback that derived it from rotary_pos_emb. Only 4.51.3 satisfies both conditions, confirmed by reading the source and by an actual export.

Rather than patch around it, the author kept two rules: no patches to upstream scripts or the model, and verification from source rather than guesswork. An alias for .visual was tried first, but even when the attribute error was bypassed, the attention contract still failed, and altering the embedding flow risked silently wrong embeddings that would contaminate later benchmarks. Pinning versions solved it, at the cost of running 4.51.3 for the Vision export and 5.8.0 for quantization embeddings in parallel.

With the Vision tower exported (a 53MB ONNX graph plus 2,661,085,184 bytes of weights, passing onnx.checker), the remaining work is quantization and RKLLM export, then board measurements of per-call Vision encoding time and prefill/decode tokens/s, and a 24-question quality check that includes Chinese signage and OCR cases.