Attempt to access socket despite HF_HUB_OFFLINE = 1 if cache warmed outside current process

[fr1ll]

System Info

• transformers version: 4.57.1
• Platform: Linux-6.6.84.1-microsoft-standard-WSL2-x86_64-with-glibc2.35
• Python version: 3.13.0
• Huggingface_hub version: 0.36.0
• Safetensors version: 0.6.2
• Accelerate version: not installed
• Accelerate config: not found
• DeepSpeed version: not installed
• PyTorch version (accelerator?): 2.9.1+cpu (NA)
• Tensorflow version (GPU?): not installed (NA)
• Flax version (CPU?/GPU?/TPU?): not installed (NA)
• Jax version: not installed
• JaxLib version: not installed
• Using distributed or parallel set-up in script?: No

Who can help?

@ydshieh I have created a reproducible example of the issue I mentioned in #41311 (comment).

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

Reproducible example: https://github.com/fr1ll/HF_HUB_OFFLINE

Warming the cache in a subprocess, then disabling sockets, then loading the same model should work.
However, it fails with an attempt to access a socket and then "Can't load" errors.

The script named subprocess-warm_then_offline-load.py reproduces this error.

Interestingly, warming the cache in process, then disabling sockets, then loading the same model works.
This is reproduced in inprocess-warm_then_offline-load.py in the repo above.

Expected behavior

When a model has already been loaded into the cache ("warm cache"), if HF_OFFLINE_MODE = "1", a Transformers pipeline should be able to load the model without accessing any network sockets.

[ydshieh]

Thank you @fr1ll

It's indeed valid and I can reproduce. Even with inprocess-warm_then_offline-load.py, if I comment out the line that remove the cache dir. Run it twice, I will get failure.

Quite mysterious, but maybe related to the slow/fast image process issues.

I will need to dive into this more. But let me call for contribution/help from the community first.

[Aaraviitkgp]

@ydshieh I have made a pr on this issue.

[fr1ll]

#42308

[fr1ll]

Closed with explanation in #42308 by @Wauplin:

HF_HUB_OFFLINE environment variable is evaluated once at import time. If you really want to update it at runtime, you need to patch huggingface_hub.constant.HF_HUB_OFFLINE

tl;dr: When possible, set the environment variable before importing transformers.

Note: Possible that local_files_only argument can be used to disable hub access at runtime for some actions. #42269 indicates this may not work for transformers.pipeline; #42318 offers a fix.

[fr1ll]

@ydshieh sorry if I closed this with the wrong reason- probably it should be something besides completed. I don't have access to reopen and amend the reason.

[Wauplin]

closing as "completed" is fine here, considering #42197 (comment) contains the solution (set the environment variable before importing transformers)