Summary

• add a Bridge-local, PEP 440-aware Transformers compatibility helper with actionable version and missing-symbol errors
• attach optional minimum-version and required-symbol metadata to registered bridge classes, enforcing it lazily after model selection and before provider/model loading or conversion
• narrowly translate known guarded AutoConfig architecture-recognition failures without masking network, authentication, malformed-config, or unrelated errors
• migrate GLM-5, Qwen3.5 text/VL, and DeepSeek V4 registrations and provider imports to the new mechanism
• add focused coverage for version parsing, dotted symbols, registration metadata, guard timing, config error handling, and fresh-process registration
• update the adding-model-support skill with compatibility discovery, registration, lazy-import, and boundary-test requirements

Motivation

This is the guard and metadata foundation proposed in #4314. It allows package and registry discovery to remain independent of newer model-specific Transformers APIs while preserving the current global dependency range. This PR does not loosen dependency bounds.

min_transformers_version=None intentionally means that compatibility research is incomplete. It is not treated as an implicit minimum; compatibility metadata is added only where the version boundary is supported by direct evidence.

Compatibility evidence

• GLM-5 and Qwen3.5 required symbols are absent in Transformers 5.1.0 and present in 5.2.0+, so those bridges declare >=5.2.0 with exact symbols.
• DeepSeek V4's native config and model symbols are available and its bridge constructs with Transformers 5.8.0, while 5.7.0 produces the exact lazy guard. It therefore declares >=5.8.0.
• The 26.06 container baseline uses Transformers 5.8.1. Disposable wheel overlays exercised exact Transformers 5.7.0 and 5.6.0 without changing project dependency bounds.
• Every registered model family reached Bridge's loading progress path in all eligible environments. DeepSeek V4 is the only family with a newly observed version-specific pre-load incompatibility.

Validation

• uv run pre-commit run --all-files in nvcr.io/nvidia/nemo:26.06
• focused impacted unit tests: 87 passed on Transformers 5.8.1; 58 passed on 5.7.0; 58 passed on 5.6.0
• scoped tests/unit_tests/models: 2783 passed on 5.8.1; 2753 passed on 5.7.0; 2753 passed on 5.6.0
• the same three Qwen3-VL processor setup errors occurred in each broad run because of process-global auto-doc registration order; the full Qwen3-VL test file passed 10/10 in a fresh process on every version
• public tiny-model plus repository toy/synthetic import probes covered every registered model-family directory and reached loading progress on 5.8.1, 5.7.0, and 5.6.0 where compatible
• exact DeepSeek V4 boundary probes passed at 5.8.0 and produced the expected lazy version error at 5.7.0
• DeepSeek V4 on the 5.8.1 baseline proceeds past Transformers compatibility, then encounters an independent pinned-MCore dsv4_hybrid layer-type mismatch
• quick_validate.py skills/adding-model-support
• git diff --check origin/main...HEAD