Even more test data cached

[ydshieh]

What does this PR do?

Even more test data cached

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[ydshieh]

First, the name url_to_local_path is misleading, sorry.

It actually check if the file exists locally ( the root of transformers), return local path (i.e. its filename) if it is there, otherwise return the original url.

I will change the name to make it more clear.

But how we need to handle both cases in a better way, but not if (url ) ... else ... (local path) everywhere.

@zucchini-nlp Do we have some excising methods that could handle this well?

(chat template could handle both cases, but here we need some more primitive functions)

[zucchini-nlp]

yeah, i remember looking at how url_to_local_path works. Makes sense to me

@zucchini-nlp Do we have some excising methods that could handle this well?

Not sure I got this. Do you mean a method to do url_to_local_path in chat templates?

[ydshieh]

No, What I mean is:

this block

                requests.get(
                    url_to_local_path(
                        "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
                    )

but url_to_local_path may return the original url or the local path (if it finds).

Therefore we want to have a simple function to handle both case like here: as requests.get doesn't work with local path.

So I am wondering if you remember if we have already utility functions for doing things like this (and likely used by chat template ..?)

[zucchini-nlp]

ahh, no, I don't remember any existing utility we have. Maybe we could put all images/videos/audios in hub and use hf-hub download. AFAIK it should cache everything automatically?

[ydshieh]

OK, I find

load_image

src/transformers/image_utils.py

[ydshieh]

will use it and merge the PR

[zucchini-nlp]

it is the same utility I use in chat templates, and it has no caching mechanism. It will keep sending requests to download every time

[ydshieh]

I don't mean simply use it will fix the issue. But it combined with url_to_local_path will (if we have a previous step of downloading the files before the test run step).

i.e.

load_image(url_to_local_path( ... url ...))

will work well (assuming already downloaded)

while

        cls.image1 = Image.open(
            BytesIO(
                requests.get(
                    url_to_local_path(
                        "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
                    )
             )
        )

will fail as it can't handle the local path.

(We are talking 2 things different here. My question is more about how to handle (i.e. load things into memory) both cases in a simple way, assuming the files are already downloaded)

[zucchini-nlp]

ahh sorry, I misunderstood the comment probably. Right, load_image handles all input types

[zucchini-nlp]

yeah, i remember looking at how url_to_local_path works. Makes sense to me

@zucchini-nlp Do we have some excising methods that could handle this well?

Not sure I got this. Do you mean a method to do url_to_local_path in chat templates?

[ydshieh]

TODO: explore of using hf_hub_download instead of url_to_local_path. If not fit, use a better name than url_to_local_path which is confusing.

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: aria, aya_vision, bridgetower, cohere2_vision, deepseek_vl, deepseek_vl_hybrid, eomt, fuyu, gemma3, gemma3n, glm4v, idefics2, idefics3, internvl, janus, kosmos2

[ydshieh]

hf_hub_download indeed could work as it returns

    Returns:
        `str`: Local path of file or if networking is off, last version of file cached on disk.

However:

1. can't work with a simple url string
2. can't work with data hosted outside HF hub

For 2., we could always upload to a HF hub repo. For 1., the changes need to be done is a bit annoying.

For now, I will simply rename url_to_local_path to be less confusing, and maybe allow to specify the root output directory.

This is intended to be used only on a CI environment (CI runners) and not for local development. For now, let's not spend to much time on engineering it