whisper-small-bemba2 / quantize_config.json
chiyo123's picture
Upload 19 files
b21de1a
{
"per_channel": false,
"reduce_range": false,
"per_model_config": {
"decoder_model": {
"op_types": [
"Softmax",
"Where",
"Concat",
"Constant",
"ConstantOfShape",
"Less",
"ReduceMean",
"Cast",
"Unsqueeze",
"Gather",
"Equal",
"Mul",
"Sub",
"Reshape",
"Transpose",
"Erf",
"Div",
"Range",
"Sqrt",
"Squeeze",
"Add",
"Shape",
"MatMul",
"Expand",
"Pow",
"Slice"
],
"weight_type": "QInt8"
},
"decoder_model_merged": {
"op_types": [
"Softmax",
"Where",
"Concat",
"Constant",
"ConstantOfShape",
"Less",
"ReduceMean",
"Cast",
"Unsqueeze",
"Gather",
"Equal",
"Mul",
"Sub",
"Reshape",
"Transpose",
"Erf",
"Div",
"If",
"Range",
"Sqrt",
"Squeeze",
"Add",
"Shape",
"MatMul",
"Expand",
"Pow",
"Slice"
],
"weight_type": "QInt8"
},
"decoder_with_past_model": {
"op_types": [
"Softmax",
"Reshape",
"Sqrt",
"Transpose",
"Erf",
"Add",
"Shape",
"MatMul",
"Unsqueeze",
"Concat",
"Constant",
"Gather",
"ReduceMean",
"Div",
"Mul",
"Pow",
"Sub",
"Slice"
],
"weight_type": "QInt8"
},
"encoder_model": {
"op_types": [
"Conv",
"Softmax",
"Transpose",
"Sqrt",
"Erf",
"Reshape",
"Add",
"Shape",
"MatMul",
"Unsqueeze",
"Concat",
"Constant",
"Div",
"Gather",
"Mul",
"Pow",
"Sub",
"ReduceMean"
],
"weight_type": "QUInt8"
}
}
}