Add options to the "delete-cache" command

[Wauplin]

A CLI tool has been introduced in #1025. It allow to scan and delete the HF cache directory. This is especially useful when hard drive gets full.

huggingface-cli delete-cache

Currently a list of repos is printed with details like revision name, size and last modified date. The user can select which revisions to delete. It can be done either via a Terminal UI (if huggingface_hub[cli] is installed) or via a temporary file to edit (if TUI not supported).

At the moment the selection is entirely manual to let the user decide what to do. To ease the process, we discussed about implementing new CLI options:

• --filter to filter repo names
• --sort to sort by age, alphabetically, size,...
• --limit to display only the top X repos
• --keep-last to keep only the last revision of each repo

Each option can be implemented separately in a different PR. CLI implementation can be found in ./commands/delete_cache.py while the cache scan tool itself is in ./utils/_cache_manager.py

[jantrienes]

We are interested in cleaning cached models which were not accessed in the past N days. Would it make sense to provide an option like --last_access N which returns all repos with last access before today - N?

For the time being, I think this can be done with the Python API, but the CLI would be more convenient.

from datetime import datetime
from huggingface_hub import scan_cache_dir

expiry = 21  # in days
now = datetime.now()
cache_info = scan_cache_dir()

to_clean = []
for repo in cache_info.repos:
    delta = now - datetime.fromtimestamp(repo.last_accessed)
    if delta.days >= expiry:
        print(f"{repo.size_on_disk_str:>8}", f"{delta.days:>4} days", repo.repo_id)
        to_clean += [revision.commit_hash for revision in repo.revisions]

delete_strategy = cache_info.delete_revisions(*to_clean)
print(f"Will free {delete_strategy.expected_freed_size_str}.")
# delete_strategy.execute()

[sifisKoen]

@Wauplin is this issue still open? Dose this needs a solution?

[sealad886]

I'll end up doing one or several of these while fixing #2219 as well.

The one that most bothers me right now is that the sort order for delete-cache (with the tui tool) is date-sorted. It's way more useful, imo, to sort by name because I intuitively know approximately where to find what I'm looking for with that sort. A date sort...requires me to look at every. single. line. every. single. time. It's so slow.