OpenMed v1.8.1 — architecture-safe attention hotfix

Bug fix: Resolved the DebertaV2ForTokenClassification scaled-dot-product-attention loading error. Automatic attention selection now uses Transformers architecture compatibility checks instead of forcing PyTorch SDPA from runtime availability alone.

OpenMed 1.8.1 fixes a model-loading regression for DeBERTa-v2 token-classification models, including OpenMed/OpenMed-NER-DiseaseDetect-SuperClinical-434M, across NVIDIA, AMD, and CPU environments.

Fixed

• torch_attention_backend="auto" no longer forces SDPA from runtime availability alone. Transformers now chooses an implementation that the model architecture supports.
• Unavailable explicit accelerated attention backends fall back to the architecture-independent eager implementation.
• Direct model loading and pipeline construction now omit the attention override when automatic selection is requested.
• Android tag publishing now skips unchanged artifacts cleanly and includes embedded model-catalog inputs when deciding whether publication is required.

Verification

• Full hosted Python matrix passed on Linux, macOS, and Windows with Python 3.10–3.12.
• DeBERTa-v2 attention regression tests, Android, Swift, Helm, security, package build, and container smoke checks passed.
• PyPI distributions were built, checked, attested, verified, and published.
• GHCR v1.8.1 is a verified linux/amd64 and linux/arm64 manifest.
• Package, image, and CycloneDX SBOM evidence is attached to this release.

Install with:

:::bash
pip install --upgrade "openmed[hf]==1.8.1"

• PyPI 1.8.1
• Container package
• Attention hotfix PR [#1545]
• Release workflow fix PR [#1546]

Full Changelog: https://github.com/maziyarpanahi/openmed/compare/v1.8.0...v1.8.1