Model size check doesn't seem to work

[kleiti]

Even with my model that is less than 250KB in size, I get the onnx_data file after quantization.

neural-compressor/onnx_neural_compressor/onnx_model.py

Line 245 in aabbf96

save_as_external_data=True,

[mengniwang95]

Hi, is your original model format .onnx file + weight file?

neural-compressor/onnx_neural_compressor/onnx_model.py

Line 116 in aabbf96

if init.HasField("data_location") and init.data_location == onnx.TensorProto.EXTERNAL:

If the original format is .onnx file + weight file, saving operation will follow it.

[kleiti]

My model is just the .onnx file, no external weight file.

[mengniwang95]

Do you quantize the model with this API

neural-compressor/onnx_neural_compressor/quantization/quantize.py

Line 28 in aabbf96

def quantize(

?
If so, during optimization it will save external data if model size > min_size:

neural-compressor/onnx_neural_compressor/quantization/quantize.py

Line 43 in aabbf96

"session.optimized_model_external_initializers_min_size_in_bytes", "1024"

[kleiti]

Yes, I do quantize using that API. Why is the size limit so low?

[mengniwang95]

Since onnxruntime can load external data automatically and we find model with external data format can be processed faster. But according to you case, the size limitation may need optimization.
Do you have some limitations to use external data format?