HUGE disk space usage in ORT_COREML

[yuygfgg]

The coreml models compiled (or translated) from onnx is taking over 180GB in /private/var/folders/gb/m6gky8t10yx68drv06ncbj_h0000gn/T/ folder. I assume there is some missing cleaning up steps in vsort but can't confirm that. As a workaround, I added

import subprocess; subprocess.run('rm -r /private/var/folders/gb/m6gky8t10yx68drv06ncbj_h0000gn/T/onnxruntime*', shell=True, check=True) 

at the end of my VS script.

• I think the path may vary on different devices, but in the past weeks only this folder is polluted.

The picture shows running a small model (Waifu2xModel.upconv_7_anime_style_art_rgb) for several times.

If larger model is used, larger mlmodel(c) folder (yes, although they look like files, they are folders in fact) is generated. For example, DPIR takes 130.1MB per run per tile.

[WolframRhodium]

Thanks for the information, I'll check it tomorrow.

[WolframRhodium]

I think this is a limitation of ORT itself: https://github.com/microsoft/onnxruntime/blob/0f1f3b7705ddc2fe4f371f78a8a8b6a0428a68de/onnxruntime/core/providers/coreml/builders/model_builder.cc#L996-L1008 so your post cleaning script is required unfortunately.

[yuygfgg]

I think this is a limitation of ORT itself: https://github.com/microsoft/onnxruntime/blob/0f1f3b7705ddc2fe4f371f78a8a8b6a0428a68de/onnxruntime/core/providers/coreml/builders/model_builder.cc#L996-L1008 so your post cleaning script is required unfortunately.

I see. I'm now kind of worrying about my SSD...