distilgpt2 / quantize_config.json
Xenova's picture
Xenova HF staff
Upload folder using huggingface_hub
0aa7002
{
"per_channel": false,
"reduce_range": false,
"per_model_config": {
"decoder_model": {
"op_types": [
"Squeeze",
"MatMul",
"Reshape",
"Slice",
"Range",
"Where",
"Constant",
"ConstantOfShape",
"Split",
"Div",
"Mul",
"Concat",
"ReduceMean",
"Shape",
"Gather",
"Sqrt",
"Tanh",
"Cast",
"Transpose",
"Sub",
"Softmax",
"Unsqueeze",
"Gemm",
"Pow",
"Add"
],
"weight_type": "QInt8"
},
"decoder_model_merged": {
"op_types": [
"Squeeze",
"MatMul",
"Reshape",
"Slice",
"Range",
"Where",
"Constant",
"ConstantOfShape",
"Split",
"Div",
"Mul",
"Concat",
"If",
"ReduceMean",
"Shape",
"Gather",
"Sqrt",
"Tanh",
"Cast",
"Transpose",
"Sub",
"Softmax",
"Unsqueeze",
"Gemm",
"Pow",
"Add"
],
"weight_type": "QInt8"
},
"decoder_with_past_model": {
"op_types": [
"Squeeze",
"MatMul",
"Reshape",
"Slice",
"Range",
"Where",
"Constant",
"ConstantOfShape",
"Split",
"Div",
"Mul",
"Concat",
"ReduceMean",
"Shape",
"Gather",
"Sqrt",
"Tanh",
"Cast",
"Transpose",
"Sub",
"Softmax",
"Unsqueeze",
"Gemm",
"Pow",
"Add"
],
"weight_type": "QInt8"
}
}
}