mbart-large-onnx / quantize_config.json
Varosa's picture
upload json files
82541f4
{
"per_channel": true,
"reduce_range": true,
"per_model_config": {
"encoder_model": {
"op_types": [
"Div",
"Transpose",
"Mul",
"Range",
"Shape",
"Softmax",
"Cast",
"Reshape",
"Constant",
"ReduceMean",
"Gather",
"ConstantOfShape",
"Unsqueeze",
"Expand",
"Sub",
"MatMul",
"Add",
"Sqrt",
"Equal",
"Where",
"Concat",
"Relu",
"Pow"
],
"weight_type": "QInt8"
},
"decoder_with_past_model": {
"op_types": [
"Div",
"Transpose",
"Mul",
"Range",
"Shape",
"Softmax",
"Cast",
"Reshape",
"Constant",
"ReduceMean",
"Gather",
"ConstantOfShape",
"Unsqueeze",
"Expand",
"Sub",
"MatMul",
"Add",
"Sqrt",
"Equal",
"Where",
"Concat",
"Relu",
"Pow"
],
"weight_type": "QInt8"
},
"decoder_model_merged": {
"op_types": [
"Div",
"Less",
"Transpose",
"Mul",
"Range",
"Shape",
"Softmax",
"Cast",
"Reshape",
"Constant",
"ReduceMean",
"If",
"Squeeze",
"Gather",
"ConstantOfShape",
"Unsqueeze",
"Expand",
"Sub",
"MatMul",
"Add",
"Sqrt",
"Equal",
"Where",
"Concat",
"Slice",
"Relu",
"Pow"
],
"weight_type": "QInt8"
},
"decoder_model": {
"op_types": [
"Div",
"Less",
"Transpose",
"Mul",
"Range",
"Shape",
"Softmax",
"Cast",
"Reshape",
"Constant",
"ReduceMean",
"Squeeze",
"Gather",
"ConstantOfShape",
"Unsqueeze",
"Expand",
"Sub",
"MatMul",
"Add",
"Sqrt",
"Equal",
"Where",
"Concat",
"Slice",
"Relu",
"Pow"
],
"weight_type": "QInt8"
}
}
}