vlm/config.json

{
  "model_type": "qwen2vl",
  "quantization": "8-bit",
  "architectures": [
    "Qwen2VLForConditionalGeneration"
  ],
  "torch_dtype": "bfloat16",
  "precision": "8-bit",
  "base_model": "Qwen/Qwen2.5-VL-3B-Instruct"
}