flan-t5-base / quantize_config.json
Xenova's picture
Xenova HF staff
Upload folder using huggingface_hub
ddb2d21
{
"per_channel": true,
"reduce_range": true,
"per_model_config": {
"decoder_model": {
"op_types": [
"Mul",
"MatMul",
"Neg",
"LessOrEqual",
"Expand",
"Constant",
"Gather",
"Min",
"Where",
"Tanh",
"Reshape",
"Range",
"Less",
"Transpose",
"Shape",
"Log",
"ConstantOfShape",
"Sqrt",
"ReduceMean",
"Sub",
"Add",
"Unsqueeze",
"Div",
"Pow",
"Concat",
"Tile",
"Softmax",
"Cast"
],
"weight_type": "QInt8"
},
"encoder_model": {
"op_types": [
"Mul",
"MatMul",
"Constant",
"Gather",
"Greater",
"Min",
"Abs",
"Where",
"Tanh",
"Reshape",
"Range",
"Less",
"Transpose",
"Shape",
"Log",
"Sqrt",
"ConstantOfShape",
"ReduceMean",
"Sub",
"Add",
"Unsqueeze",
"Div",
"Pow",
"Concat",
"Softmax",
"Cast"
],
"weight_type": "QInt8"
},
"decoder_model_merged": {
"op_types": [
"Mul",
"MatMul",
"Neg",
"LessOrEqual",
"Expand",
"Constant",
"Gather",
"Slice",
"Min",
"Where",
"Tanh",
"Reshape",
"If",
"Range",
"Less",
"Transpose",
"Shape",
"Log",
"ConstantOfShape",
"Sqrt",
"ReduceMean",
"Sub",
"Add",
"Unsqueeze",
"Div",
"Pow",
"Concat",
"Tile",
"Softmax",
"Cast"
],
"weight_type": "QInt8"
},
"decoder_with_past_model": {
"op_types": [
"Mul",
"MatMul",
"Neg",
"LessOrEqual",
"Expand",
"Constant",
"Gather",
"Slice",
"Min",
"Where",
"Tanh",
"Reshape",
"Range",
"Less",
"Transpose",
"Shape",
"Log",
"ConstantOfShape",
"Sqrt",
"ReduceMean",
"Add",
"Sub",
"Unsqueeze",
"Div",
"Pow",
"Concat",
"Tile",
"Softmax",
"Cast"
],
"weight_type": "QInt8"
}
}
}