Hugging Face transformers: 'OSError: We couldn't connect to 'https://huggingface.co' to load the files, and couldn't find them in the cached files' (or LocalEntryNotFoundError) on an offline, firewalled or air-gapped machine, in Docker or CI. How do I run models offline?

What it is: from_pretrained() and pipeline() normally contact the Hugging Face Hub to find and download a model's files, then cache them (by default under ~/.cache/huggingface/hub). If the machine can't reach the Hub (no internet, a firewall or proxy, an air-gapped server, a container with no network, a Hub outage or a slow connection timing out) and the needed files aren't already in the cache that this process looks at, transformers raises this OSError; huggingface_hub raises LocalEntryNotFoundError in the same situation. A common hidden cause is that the files were downloaded into a different cache: another user (root vs your user), another HF_HOME or HF_HUB_CACHE value, or a container that doesn't mount the host cache. Am I affected: - The error appears only on machines or containers without Hub access, or only for some users. - Check: echo $HF_HOME $HF_HUB_CACHE, and look for a models--<org>--<name> folder in the cache directory; hf cache scan (the hf CLI) lists cached repos. Fix: 1. Pre-download on a machine with access: snapshot_download(repo_id="org/model") in Python, or the hf download CLI command, ideally pinned to a revision (commit hash). Copy the cache folder to the offline machine, or download straight into a folder (local_dir) you ship with your app or image. 2. On the offline machine, point to the same cache (set HF_HOME or HF_HUB_CACHE to where the files are; in Docker, mount that folder or bake it into the image at the same path). 3. Tell the libraries not to try the network: set HF_HUB_OFFLINE=1, or pass local_files_only=True to from_pretrained(). You can also load from a local directory path: from_pretrained("/models/my-model"). 4. Make sure every file the model needs was downloaded (config, tokenizer files, weights, and any custom code or adapter repos it references); offline loads fail on whatever file is missing. 5. On a slow but working connection, raise HF_HUB_DOWNLOAD_TIMEOUT (default 10 seconds) instead of going offline; HF_HUB_ETAG_TIMEOUT controls how long it waits for metadata before falling back to cached files. 6. Behind a corporate proxy, configure HTTPS_PROXY / the proxy's CA bundle properly rather than disabling TLS verification. Verify: with HF_HUB_OFFLINE=1 set, from_pretrained() loads the model with no network access, and the same works inside the container or CI job. Sources: https://huggingface.co/docs/transformers/installation https://huggingface.co/docs/huggingface_hub/package_reference/environment_variables https://huggingface.co/docs/huggingface_hub/guides/download