kogpt-j-350m / quantize_config.json
Xenova's picture
Xenova HF staff
Upload folder using huggingface_hub
fbe0a88
{
"per_channel": true,
"reduce_range": true,
"per_model_config": {
"decoder_model": {
"op_types": [
"Equal",
"Cast",
"SequenceAt",
"Tile",
"SplitToSequence",
"Div",
"Split",
"Range",
"Where",
"Constant",
"Expand",
"Add",
"Gather",
"Sqrt",
"Pow",
"Tanh",
"Reshape",
"Slice",
"Loop",
"SequenceEmpty",
"MatMul",
"Mul",
"ConcatFromSequence",
"SequenceInsert",
"ReduceMean",
"Transpose",
"GatherElements",
"Shape",
"Squeeze",
"Sub",
"Unsqueeze",
"Concat",
"Neg",
"Softmax",
"ConstantOfShape"
],
"weight_type": "QInt8"
},
"decoder_model_merged": {
"op_types": [
"Equal",
"Cast",
"SequenceAt",
"Tile",
"SplitToSequence",
"Div",
"Split",
"Range",
"Where",
"Constant",
"Expand",
"Add",
"Gather",
"Sqrt",
"Pow",
"Tanh",
"Reshape",
"Slice",
"Loop",
"SequenceEmpty",
"MatMul",
"Mul",
"ConcatFromSequence",
"SequenceInsert",
"If",
"ReduceMean",
"Transpose",
"GatherElements",
"Shape",
"Squeeze",
"Sub",
"Unsqueeze",
"Concat",
"Neg",
"Softmax",
"ConstantOfShape"
],
"weight_type": "QInt8"
},
"decoder_with_past_model": {
"op_types": [
"Equal",
"Cast",
"SequenceAt",
"Tile",
"SplitToSequence",
"Div",
"Split",
"Range",
"Where",
"Constant",
"Expand",
"Add",
"Gather",
"Sqrt",
"Pow",
"Tanh",
"Reshape",
"Slice",
"Loop",
"SequenceEmpty",
"MatMul",
"Mul",
"ConcatFromSequence",
"SequenceInsert",
"ReduceMean",
"Transpose",
"GatherElements",
"Shape",
"Squeeze",
"Sub",
"Unsqueeze",
"Concat",
"Neg",
"Softmax",
"ConstantOfShape"
],
"weight_type": "QInt8"
}
}
}