hf_hub_download doesn't retry on ReadTimeout

[alexmojaki]

I got a test failure in https://app.circleci.com/pipelines/github/huggingface/transformers/66984/workflows/0a4f87b2-f6eb-4a8f-ba6f-8edead4065fd/jobs/835240 that came down to a ReadTimeout. Here's part of the traceback:

../.pyenv/versions/3.8.12/lib/python3.8/site-packages/huggingface_hub/file_download.py:1364: in hf_hub_download
    http_get(
../.pyenv/versions/3.8.12/lib/python3.8/site-packages/huggingface_hub/file_download.py:505: in http_get
    r = _request_wrapper(
../.pyenv/versions/3.8.12/lib/python3.8/site-packages/huggingface_hub/file_download.py:442: in _request_wrapper
    return http_backoff(
../.pyenv/versions/3.8.12/lib/python3.8/site-packages/huggingface_hub/utils/_http.py:212: in http_backoff
    response = session.request(method=method, url=url, **kwargs)

In particular in _request_wrapper is this line:

huggingface_hub/src/huggingface_hub/file_download.py

Line 448 in 3832f67

retry_on_exceptions=(ConnectTimeout, ProxyError),

I don't know the reasoning behind this choice of exceptions to retry, but failures in CI where the network is presumably good seems like a data point in favour of adding ReadTimeout to that tuple.

[Wauplin]

Thanks for opening the issue @alexmojaki. I opened a PR (#1529) to fix this. No strong choice on the (ConnectTimeout, ProxyError) tuple in my opinion. I think ReadTimeout has been forgotten just because it's a quite rare issue.

[Wauplin]

Fix by #1529.