Converted checkpoint models redownload after clearing files from cache

[bzlibby]

After removing the safetensors files for my already converted checkpoint models from my cache folder, they were all re-downloaded the next time I started my local server. Once the safetensor file was re-downloaded to cache, logs indicate that the already existing converted model was recognized and conversion was skipped:

[2023-07-10 16:50:51,349] INFO: MainProcess MainThread __main__: downloading model from Civitai: https://civitai.com/api/download/models/97123 -> D:\onnx-web\models\.cache\diffusion-occidentalmix.safetensors
100%|████████████████████████████████████████████████████████████████████████████████████████| 1.99G/1.99G [08:45<00:00, 4.06MB/s]
[2023-07-10 16:59:47,852] INFO: MainProcess MainThread onnx_web.convert.diffusion.diffusers: converting Stable Diffusion model diffusion-occidentalmix: D:\onnx-web\models\.cache\diffusion-occidentalmix.safetensors -> D:\onnx-web\models\diffusion-occidentalmix/
[2023-07-10 16:59:47,853] INFO: MainProcess MainThread onnx_web.convert.diffusion.diffusers: ONNX model already exists, skipping

The models work as expected once the server is running, but it does mean that my cache folder is taking up a significant amount of space on my drive, which was my reason for deleting the files in the first place. Removing the already-converted models from the extras JSON file gets around this.

I haven't tested this with LoRas or inversions yet, but I'll update once I get the chance to with what I find.

[ssube]

Once diffusion models have been converted, you should not need the original safetensors file or the temporary -torch folder, so that seems like a bug between the download code and conversion code.

You do need the safetensors for LoRAs and TIs, because they are never converted, but those are much smaller.

[ssube]

This should be fixed and will be in the v0.11 release. If the ONNX model directory already exists, the conversion code will not fetch the source tensor.

There is one case where that can still break: if you run out of disk while converting a model but after the ONNX model's folder has been created, you can end up with a partial model. In that case, delete the failed conversion's folder and run it again after ensuring you have enough space for the torch intermediates and ONNX output.