Ablation: move red block to bottom-right to test saliency vs localization

Configure 3 contrasting ablation studies on unified spatial image

Fix UnboundLocalError by removing invalid del statement

Fix internal Accelerate OOM by strictly limiting image resolution to 384x384

Fix OOM by CPU offloading and dual GPU setup

Fix SDPA error and output.language_model structure issue

Fix absolute path to /root/models/llava-ov-7b and add path check

Fix local model path to /models/llava-ov-7b

Add layer head probing script for spatial attention analysis

Fix dummy image for pure text and align prompt format with MMBench

Add scripts for color analysis, binding analysis and spatial hypothesis testing

feat: 更新纯色图测试 - 问题改为颜色识别+HEX/RGB输出, 增加2x2色块图测试

feat: 添加纯色图测试脚本，对比LLaVA-OneVision 0.5B vs 7B

Update test script to auto-load .env file

Add DeepSeek API connectivity test script