pythia-410m-deduped / quantize_config.json
Xenova's picture
Xenova HF staff
Upload folder using huggingface_hub
b87db57
{
"per_channel": false,
"reduce_range": false,
"per_model_config": {
"decoder_model": {
"op_types": [
"MatMul",
"Slice",
"ConstantOfShape",
"Neg",
"ReduceMean",
"Squeeze",
"Erf",
"Add",
"Where",
"Range",
"Softmax",
"Cast",
"Transpose",
"Mul",
"Tile",
"Reshape",
"Sqrt",
"Expand",
"GatherElements",
"Concat",
"Div",
"Gather",
"Pow",
"Constant",
"Shape",
"Sub",
"Unsqueeze"
],
"weight_type": "QInt8"
},
"decoder_model_merged": {
"op_types": [
"MatMul",
"Slice",
"ConstantOfShape",
"Neg",
"ReduceMean",
"Squeeze",
"Erf",
"Add",
"Where",
"If",
"Range",
"Softmax",
"Cast",
"Transpose",
"Mul",
"Tile",
"Reshape",
"Sqrt",
"Expand",
"GatherElements",
"Concat",
"Div",
"Gather",
"Pow",
"Constant",
"Shape",
"Sub",
"Unsqueeze"
],
"weight_type": "QInt8"
},
"decoder_with_past_model": {
"op_types": [
"MatMul",
"Slice",
"ConstantOfShape",
"Neg",
"ReduceMean",
"Squeeze",
"Erf",
"Add",
"Where",
"Range",
"Softmax",
"Cast",
"Transpose",
"Mul",
"Tile",
"Reshape",
"Sqrt",
"Expand",
"GatherElements",
"Concat",
"Div",
"Gather",
"Pow",
"Constant",
"Shape",
"Sub",
"Unsqueeze"
],
"weight_type": "QInt8"
}
}
}