t5-small / quantize_config.json
tuanluong105's picture
Upload folder using huggingface_hub
dc36e21 verified
{
"per_channel": true,
"reduce_range": true,
"per_model_config": {
"decoder_model_merged": {
"op_types": [
"Gather",
"ReduceMean",
"Add",
"Unsqueeze",
"Expand",
"Sqrt",
"Sub",
"Min",
"Log",
"Softmax",
"Transpose",
"Cast",
"Relu",
"Neg",
"If",
"MatMul",
"ConstantOfShape",
"Div",
"Less",
"Pow",
"Mul",
"Slice",
"Constant",
"LessOrEqual",
"Range",
"Where",
"Shape",
"Tile",
"Reshape",
"Concat"
],
"weight_type": "QInt8"
},
"decoder_with_past_model": {
"op_types": [
"Gather",
"ReduceMean",
"Add",
"Unsqueeze",
"Expand",
"Sqrt",
"Sub",
"Min",
"Log",
"Softmax",
"Transpose",
"Cast",
"Relu",
"Neg",
"MatMul",
"ConstantOfShape",
"Div",
"Less",
"Pow",
"Mul",
"Slice",
"Constant",
"LessOrEqual",
"Range",
"Where",
"Shape",
"Tile",
"Reshape",
"Concat"
],
"weight_type": "QInt8"
},
"encoder_model": {
"op_types": [
"Gather",
"ReduceMean",
"Add",
"Unsqueeze",
"Sqrt",
"Sub",
"Log",
"Min",
"Softmax",
"Transpose",
"Cast",
"Greater",
"Relu",
"MatMul",
"ConstantOfShape",
"Div",
"Less",
"Pow",
"Mul",
"Constant",
"Where",
"Range",
"Shape",
"Reshape",
"Abs",
"Concat"
],
"weight_type": "QInt8"
},
"decoder_model": {
"op_types": [
"Gather",
"ReduceMean",
"Add",
"Unsqueeze",
"Expand",
"Sqrt",
"Sub",
"Min",
"Log",
"Softmax",
"Transpose",
"Cast",
"Relu",
"Neg",
"MatMul",
"ConstantOfShape",
"Div",
"Less",
"Pow",
"Mul",
"Constant",
"LessOrEqual",
"Range",
"Where",
"Shape",
"Tile",
"Reshape",
"Concat"
],
"weight_type": "QInt8"
}
}
}