flan-t5-small-onnx-js-quantized / quantize_config.json
dmmagdal's picture
Uploaded model files (quantized)
e8d20a9
{
"per_channel": true,
"reduce_range": true,
"per_model_config": {
"decoder_with_past_model": {
"op_types": [
"Cast",
"Less",
"Sub",
"Range",
"Sqrt",
"Tanh",
"Transpose",
"Gather",
"Softmax",
"ReduceMean",
"Constant",
"Where",
"Min",
"Shape",
"LessOrEqual",
"Pow",
"Concat",
"Expand",
"Add",
"Slice",
"Tile",
"Reshape",
"ConstantOfShape",
"MatMul",
"Log",
"Neg",
"Unsqueeze",
"Mul",
"Div"
],
"weight_type": "QInt8"
},
"decoder_model": {
"op_types": [
"Cast",
"Less",
"Sub",
"Range",
"Sqrt",
"Tanh",
"Transpose",
"Gather",
"Softmax",
"ReduceMean",
"Constant",
"Where",
"Min",
"Shape",
"LessOrEqual",
"Pow",
"Concat",
"Expand",
"Add",
"Tile",
"Reshape",
"ConstantOfShape",
"MatMul",
"Log",
"Neg",
"Unsqueeze",
"Mul",
"Div"
],
"weight_type": "QInt8"
},
"encoder_model": {
"op_types": [
"Cast",
"Less",
"Sub",
"Range",
"Sqrt",
"Tanh",
"Transpose",
"Gather",
"Softmax",
"ReduceMean",
"Constant",
"Where",
"Min",
"Shape",
"Abs",
"Pow",
"Concat",
"Add",
"Reshape",
"MatMul",
"ConstantOfShape",
"Greater",
"Log",
"Unsqueeze",
"Mul",
"Div"
],
"weight_type": "QInt8"
},
"decoder_model_merged": {
"op_types": [
"Cast",
"Less",
"Sub",
"Range",
"Sqrt",
"Tanh",
"Transpose",
"Gather",
"Softmax",
"ReduceMean",
"Constant",
"Where",
"Min",
"Shape",
"LessOrEqual",
"Pow",
"Concat",
"Expand",
"Add",
"Slice",
"Tile",
"Reshape",
"ConstantOfShape",
"MatMul",
"Log",
"Neg",
"Unsqueeze",
"If",
"Mul",
"Div"
],
"weight_type": "QInt8"
}
}
}