Fix estimate-memory for timm>=1.0.29 by adding the hf-hub: prefix

[iamsharduld]

What does this PR do?

accelerate estimate-memory timm/resnet50.a1_in1k --library_name timm fails with timm 1.0.29:

ValueError: Model name 'timm/resnet50.a1_in1k' has no source prefix but looks like a Hugging Face Hub repo id or a local path. Use 'hf-hub:timm/resnet50.a1_in1k' to load from the Hub or 'local-dir:timm/resnet50.a1_in1k' to load from a local folder.

Since huggingface/pytorch-image-models#2727 (shipped in timm 1.0.29), timm needs a source prefix for anything that looks like a repo id. Older versions dropped the repo owner and looked the rest up in the registry, which only worked because timm names its Hub repos after the architecture. The two timm tests in tests/test_cli.py fail on main because of this, see https://github.com/huggingface/accelerate/actions/runs/33763859075/job/100676635381.

I now add the hf-hub: prefix before calling timm.create_model when the name looks like a repo id, and leave bare architecture names like resnet50 alone. With the prefix, timm reads config.json from the Hub and raises its own errors when the file is missing or belongs to another library. I map those to the existing "Tried to load ... with timm but ..." message, so the user sees the same error as with timm 1.0.28.

Note that timm names now resolve through the Hub config.json rather than the registry entry with the same name. That means user-owned timm repos work as well.

I tested with timm 1.0.28 and 1.0.29. I added unit tests for the prefix rule and the error mapping, and one end-to-end test for a repo whose config.json belongs to another library. No version pin is needed, timm has supported the hf-hub: prefix for a long time.

Fixes #4212

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Nothing user-facing changed, the documented timm/resnet50.a1_in1k example works as before.
• Did you write any new necessary tests?

Who can review?

@SunMarc (Command Line Interface)

[SunMarc]

Thanks, just a couple of nits. Also, does this still works with older version of timm ?

[SunMarc]

Suggested change

	@require_timm
	def test_check_has_model_timm(self):
	# Unknown architecture in the `timm` registry
	assert check_has_model(RuntimeError("Unknown model (dummy)")) == "timm"
	# Hub repo without a `timm` `config.json`, raised by `timm.create_model("hf-hub:...")`
	assert check_has_model(EntryNotFoundError("Entry Not Found for url: .../config.json.")) == "timm"
	# Hub repo whose `config.json` belongs to another library, raised by `timm.create_model("hf-hub:...")`
	assert check_has_model(KeyError("architecture")) == "timm"

[SunMarc]

Suggested change

	# `timm.create_model("hf-hub:...")` reads `config.json` from the Hub without wrapping the errors: a missing file
	# raises `EntryNotFoundError` and a `config.json` from another library has no `architecture` key (`KeyError`)

[SunMarc]

keep it short

[iamsharduld]

Applied all three suggestions: dropped the unit test and the comment, and shortened the docstring.

On older timm versions: yes, it still works. The hf-hub: prefix has been accepted since timm 0.6 (the hf_hub: spelling even earlier), and I ran the estimator tests with timm 1.0.28 as well as 1.0.29 after this change.