prem-1B-chat-ONNX / quantize_config.json
ucalyptus's picture
Upload folder using huggingface_hub
1c1d1b6 verified
{
"per_channel": false,
"reduce_range": false,
"per_model_config": {
"decoder_model": {
"op_types": [
"ConstantOfShape",
"Mul",
"Transpose",
"Expand",
"Shape",
"Reshape",
"Slice",
"Cast",
"Neg",
"Concat",
"Unsqueeze",
"MatMul",
"Div",
"Constant",
"Gather",
"Less",
"Sqrt",
"Equal",
"Sub",
"Range",
"Sigmoid",
"Softmax",
"Pow",
"Add",
"Squeeze",
"ReduceMean",
"Where",
"Identity"
],
"weight_type": "QInt8"
},
"model": {
"op_types": [
"ConstantOfShape",
"Mul",
"Transpose",
"Greater",
"ScatterND",
"Expand",
"Shape",
"Reshape",
"Slice",
"Cast",
"Neg",
"Concat",
"Unsqueeze",
"MatMul",
"Div",
"Constant",
"Cos",
"Gather",
"Sqrt",
"Trilu",
"And",
"Equal",
"Range",
"Sigmoid",
"Sin",
"Softmax",
"Pow",
"Add",
"ReduceMean",
"Where",
"Identity"
],
"weight_type": "QInt8"
}
}
}