Bundled offline model bootstrap¶

OpenMed exposes one small PII model as an explicit, registry-backed offline bundle. The package carries a versioned manifest containing the registry key, exact model id, reproducibility checksum, and license. The bundle is opt-in; importing OpenMed or constructing ModelLoader does not select it.

The model snapshot must already be present in the configured OpenMed cache (for example, in an application or air-gap bundle built during deployment). The bootstrap never downloads a missing snapshot. It resolves the model through the ordinary registry, passes local_files_only=True, requires the exact cached artifact-integrity sidecar for the pinned registry revision, and blocks sockets for the complete load. Missing, stale, skipped, or tampered integrity evidence fails closed with a content-free bootstrap error. An integrity-required load rebuilds the model and tokenizer from that verified path instead of trusting a model object cached earlier under a permissive policy.

The socket guard is process-wide. Overlapping guarded loads keep it active until the final guarded scope exits, so one thread cannot reopen egress while another offline load is still running. Other threads in the same process are also prevented from opening sockets during that interval.

from openmed.core import OpenMedConfig
from openmed.models.bundled import (
    get_bundled_model_manifest,
    load_bundled_model,
)

manifest = get_bundled_model_manifest()
print(manifest.version, manifest.model_id, manifest.license)

model = load_bundled_model(
    config=OpenMedConfig(
        cache_dir="/opt/openmed/model-cache",
        device="cpu",
    )
)

The returned value has the same shape as ModelLoader.load_model, so callers can use the normal tokenizer/model pipeline construction. If the cache does not contain the pinned snapshot and its integrity sidecar, prepare both through the normal verified model-loading path on a connected build host before the offline hand-off:

from openmed.core import OpenMedConfig
from openmed.core.models import ModelLoader

config = OpenMedConfig(cache_dir="/opt/openmed/model-cache")
ModelLoader(config).load_model("pii_detection", require_integrity=True)

Transfer the complete cache, including its integrity/ sidecars, into the offline deployment. If preparation is not possible, surface the bootstrap error to the operator; do not enable a remote fallback for clinical text. An explicitly injected custom loader receives both local_files_only=True and require_integrity=True and is responsible for honoring that contract.

This is a local inference bootstrap, not a medical device or a clinical decision guarantee. Examples and tests use synthetic, non-PHI data only.