Result narration (Qwen-powered) + UI polish for upload area + dark-mode fixes

SQL self-correction: 3 retries with error context + GROUP BY rule in prompt

SVG: standalone download (XML prolog, white bg, fixed dims) + Plotly first

Better SQL prompt: column-name mapping rules + few-shot + distinct value hints

Load LoRA via PeftModel on top of standard base models (fixes r=16 vs r=8 mismatch)

Load LoRA via PeftModel on top of standard base models (fixes r=16 vs r=8 mismatch)

Load LoRA via PeftModel on top of standard base models (fixes r=16 vs r=8 mismatch)

ZeroGPU best practice: load models at module level (cuda), inference only inside @spaces.GPU

ZeroGPU best practice: load models at module level (cuda), inference only inside @spaces.GPU

ZeroGPU best practice: load models at module level (cuda), inference only inside @spaces.GPU

Sequential model load/unload: keep peak VRAM at ~14GB

Initial deploy: Apple/Claude design, DuckDB, 3 trained LoRAs