{
  "exported_from": "pittsburghese-merged-prompt-completion",
  "output_dir": "pittsburghese-web",
  "onnx_task": "text-generation-with-past",
  "quantization": "dynamic-int8",
  "preprocess": {
    "symbolic_shape_inference": true,
    "optimization": false,
    "onnx_shape_inference": true
  },
  "per_channel": true,
  "op_types_to_quantize": [
    "MatMul"
  ],
  "note": "Accuracy-first q8 export using ORT preprocess + dynamic INT8 quantization. Quantized models are re-saved with external data location forced to *.onnx_data."
}