Feature request

Add progress bars for large model loading from cache files.

Motivation

Most of the time, model loading time will be dominated by download speed. However, for very large models we will often first download the checkpoints, then during runtime simply load them from cache. For models like Bloom however, it can take upwards of 100 minutes to load the model into RAM. During this time, there is no feedback to the user, even with verbosity set to debug. This can be frustrating as the only way to check progress is by checking system utilisation through top.

Your contribution

Happy to help if I am pointed to the relevant file or files! I don't think the progress bar would need to be extremely accurate, just some indication that something is happening.