refactor: convert prepare-test-weights.sh to python

[brycedrennan]

This does feel a lot nicer to run but it is admittedly a mostly useless refactor (in that the previous version was mostly working).

This is in preparation for wanting to re-use components of this for a similar script that outputs all the needed mapping to do load-time conversion.

Ran successfully to completion. But on a repeat run convert_unclip didn't pass the hash check for some reason.

Features

• fixes inpainting model download urls. some of the urls for the bash script were incorrect and the errors ignored
• skips downloading inpainting safety model and text encoder since we don't need to convert those
• shows a progress bar for downloads
• skips downloading existing files
• uses a temporary file to prevent partial downloads
• can do a dry run to check if url is valid DRY_RUN=1 python scripts/prepare_test_weights.py
• displays the downloaded file hash
• checks if converted weights exist, skips conversion if hash matches

Further refactoring could be done:

• simplify the downloads to just a list of files and destinations (and possibly hashes)
• use commit references instead of main branch names in urls for better reliability
• put the converted weights in a separate folder

But I wanted to leave the script in a format that would still look somewhat recognizable.

Screenshots

[catwell]

check if converted weights exist and match hash. skip if thats the case

This one would be useful (in combination with skipping the downloads if files exists, it will make re-running the script very cheap). I suggest you add it.

I tried the script and conversion failed for SAM (I did not yet investigate if this is due to the script or not):

You will also need to change CONTRIBUTING.md which references the old script.

[catwell]

I tried the script and conversion failed for SAM (I did not yet investigate if this is due to the script or not):

This is unrelated with your changes, so you can ignore it (it fails on a specific machine, the difference is probably due to something related to torch).

[brycedrennan]

check if converted weights exist and match hash. skip if thats the case

Done.

[catwell]

There are still things that could be improved, for instance: why shell out to curl in convert_preprocessors instead of using download_file?

However in its current state it works and it is undeniably an improvement, so let's merge and change it later if needed!

[brycedrennan]

I used chatgpt to do the first pass at the conversion to python and then went and cleaned it up a lot. Missed that curl call- my bad!