Offline and air-gapped installation¶

OpenMed can package the Python wheels, a small PII model, documentation, and a portable Hugging Face cache into one integrity-checked installation kit. Build the kit on a connected machine with the same target architecture and Python version as the clinic server, then transfer it by USB or another approved offline medium.

The default bundle includes OpenMed's core package, the cli and required model-runtime dependencies, and OpenMed/OpenMed-PII-SuperClinical-Small-44M-v1. The builder refuses to create the default bundle if its exact payload exceeds 900,000,000 bytes.

Build on a connected machine¶

Install OpenMed with its Hugging Face support, then create a directory bundle:

python -m pip install "openmed[cli,hf]"
openmed airgap bundle ./openmed-airgap

Use --archive or a .tar.gz output path for a single transfer file:

openmed airgap bundle ./openmed-airgap.tar.gz --archive

For a 64-bit Raspberry Pi-class Linux target running CPython 3.11, select the manylinux platform and Python version explicitly:

openmed airgap bundle ./openmed-pi-cp311 \
  --platform manylinux2014_aarch64 \
  --python-version cp311

linux/aarch64 is accepted as an alias for manylinux2014_aarch64. Cross-target downloads are binary-only: the command fails instead of silently including a source distribution that would need build tools or internet access at the clinic.

The cli extra is selected by default. Repeat --extra to choose another supported profile:

openmed airgap bundle ./openmed-onnx-kit \
  --extra cli \
  --extra onnx \
  --model OpenMed/OpenMed-PII-SuperClinical-Small-44M-v1

Repeat --model to include multiple snapshots. The model runtime is always included so the bundled PII model can run after installation.

Verify before and after transfer¶

Every payload file is listed in bundle-manifest.json with its exact byte size and SHA-256 digest. The manifest also records the target tags, package spec, model snapshot paths, artifact count, and exact total payload size. The manifest itself is excluded because a file cannot contain its own stable digest.

Verify either a directory or archive:

openmed airgap verify ./openmed-airgap
openmed airgap verify ./openmed-airgap.tar.gz

The command exits non-zero for a changed, missing, unexpected, unreadable, or duplicate file. Verify once on the connected machine and again after copying the kit to local storage on the offline machine. Do not install directly from a USB drive that has not passed verification.

Install on the offline machine¶

Extract an archive if necessary, enter the bundle directory, and run:

./install.sh

The installer runs pip with only the local wheelhouse:

python3 -m pip install --no-index --find-links wheels/ "openmed[cli,hf]==<version>"

It then copies the bundled model repositories into HF_HUB_CACHE, $HF_HOME/hub, or the standard $HOME/.cache/huggingface/hub location. Set PYTHON=/path/to/python when a different interpreter should receive the installation. Set HF_HOME or HF_HUB_CACHE before running the script when the clinic uses a shared or non-default Hugging Face cache. The installer also links the cached model repositories into OpenMed's default $HOME/.cache/openmed location without duplicating model blobs. If the OpenMed configuration already uses another cache_dir, set OPENMED_AIRGAP_CACHE to that directory while running install.sh.

Confirm local-only operation¶

Keep offline mode enabled whenever the installation must not open outbound connections:

export OPENMED_OFFLINE=1
openmed doctor
printf '%s\n' 'Patient Jane Example called 555-0100 on 2025-01-20.' | \
  openmed deid --model OpenMed/OpenMed-PII-SuperClinical-Small-44M-v1

Use only synthetic text for this first check. OPENMED_OFFLINE=1 enables the Hugging Face and Transformers offline flags, forces cache-only model loading, and activates OpenMed's socket guard around local-only inference.

Warning

An air-gapped network does not make a model clinically validated. Review local policies, test direct-identifier recall with synthetic fixtures, and never place raw PHI in bundle manifests, transfer logs, or support tickets.

Troubleshooting¶

If pip download reports no compatible wheel, confirm the manylinux platform, Python version, implementation, and ABI supported by the clinic server.
If installation reports a missing distribution, rebuild the complete bundle; do not download individual wheels on the offline machine.
If a model is not found in offline mode, confirm that the cache variables used during installation match those present when OpenMed runs.
If verification fails, discard or recopy the kit. Do not edit the manifest to match an unexplained file change.