[feat] raise exceptions on all network errors

[psychedelicious]

This is not a bug report - please do not ask for updates on a HF outage here. This is a feature request for the diffusers library to change how to handles 50X network issues.

The May 9 HF outage is a known issue. Suggest checking Twitter for status updates https://twitter.com/huggingface or visit HF's discord

Is your feature request related to a problem? Please describe.
There was a minor outage today in which calls to from_pretrained() resulted in 504 and 502 errors.

Instead of raising an exception, a message was logged.

Though the message was helpful, because it did not raise an exception, it was difficult to handle programmatically.

Describe the solution you'd like
All network errors raise exceptions.

Describe alternatives you've considered
I don't think there really is any.

[patrickvonplaten]

Hey @psychedelicious,

Very sorry for this outage - it's due to a HF Hub outage: https://status.huggingface.co/
We're working hard at resolving this.

[alexblattner]

@patrickvonplaten good luck on resolving it. Is there a way to use controlnet with only the safetensors and stable diffusion 1.5 without using the auto download or downloading the entire repo (the repo is way too heavy)?

[psychedelicious]

No worries re: the outage - it happens!

Hoping we can get exceptions raised on all network errors so that consumers of diffusers can handle the situation gracefully.

[Noyii]

I met the same problem.

huggingface_hub.utils._errors.HfHubHTTPError: 404 Client Error: Not Found for url: https://huggingface.co/api/models/runwayml/stable-diffusion-v1-5 (Request ID: Root=1-645a3ec3-3088a9cb6ee7bb3836022b22)

[ASH1998]

Why are my local models not working, I have stored them in my cache dir, why do i need to connect to hub when i have the resources in local. This needs to be changed.

[evancasey]

Are there any workarounds yet? It's preventing me from running models that are cached locally for some reason :(

[psychedelicious]

@Noyii @ASH1998 @evancasey
This is not a bug report - this is a feature request for the diffusers library.

The outage is a known issue. Suggest checking Twitter for status updates: https://twitter.com/huggingface

[aycaecemgul]

I can download models with git lfs, but it's a pain in the butt, it takes too long since there is a different versions of weights.
For example i do not need .ckpt or .safetensors version of the same model but i have to wait for it since it downloads the whole repo.
(Also before i found i can download models with git lfs, i tried snapshot_download and hf_hub_download functions to download specific files but examples and documentation is not clear and i couldn't make them work.)

tl,dr: there needs to be a clear and efficient way to download models without using from_pretrained.

[pansanity666]

I have also got 504 Server Error and 404 client error when loading instructpix2pix with from pretrained. Is there any workaround like downloading the ckpt to local machine?

[alexblattner]

@aycaecemgul agreed, it's a bad implementation philosophy regardless to be dependent on a server. Why can't diffusers just allow to use a ckpt/safetensors file and be done with it?

[patrickvonplaten]

Very sorry for the problems that were created two yesterday! It's the first time something similar happened in years.

BTW you can always run everything locally by defining local_files_only=True.

Nevertheless, we should have better code that detects if the Hub is down and catches a huggingface_hub.utils._errors.HfHubHTTPError error.

@Wauplin any idea how we could gracefully throw error messages or automatically switch to local_files_only=True when the Hub is down?

[Wauplin]

@patrickvonplaten If I remember correctly the first thing diffusers do when instantiating a pipeline is downloading the config file + make a call to huggingface_hub.model_info() to list the files in the repo. What I would do is to make this model_info call first. If it fails with a HTTPError, print a warning to the user + default back to "local_files_only=True" (either the user has no internet or the Hub is down).

So in this file I would rewrite the logic to:

        if not local_files_only:
            try:
                info = model_info(
                    pretrained_model_name,
                    use_auth_token=use_auth_token,
                    revision=revision,
                )
            except RepoNotFoundError, EntryNotFoundError:
                raise # maybe still raise if Hub is up but repo/file do not exist ?
            except HTTPError as e:
                warnings.warn(f"Couldn't connect to the Hub: {e}.\nWill try to load from local cache.")
                local_files_only = True 

        if not local_files_only:
            config_file = hf_hub_download(
                pretrained_model_name,
                cls.config_name,
                cache_dir=cache_dir,
                revision=revision,
                proxies=proxies,
                force_download=force_download,
                resume_download=resume_download,
                use_auth_token=use_auth_token,
            )
           (...) # same logic afterwards

Warning: I did not try it myself. The snippet above is just there to give an idea.