Transformers errors: separate imports, API changes and model downloads

Find the first failing stage

Transformers documents installation, caching and offline use as separate concerns.1 An installed library does not prove the model's local files are complete.

python -m pip show transformers huggingface-hub

Use the intended interpreter and compare the results with the node's version requirements. Our proposed split is import discovery, API compatibility, model initialization, and downloading. Preserve the repository, revision and requested filename for download errors; preserve the missing class or function for import errors.

Do not upgrade to a new major library version merely because a download failed. For offline use, document which files are present and which steps still attempt a request. Avoid deleting a whole cache as an exploratory first step.

Retest one controlled change with the original model identity and a small input. Record import, file parsing and inference independently. Settings that execute custom code from a model repository require a separate trust decision, not automatic approval to make a warning disappear.

Related checks

• aiohttp and requests: a missing library is not a connection failure
• Missing safetensors package or unreadable weights? Diagnose them separately

Research draft. No local/GPU reproduction or independent editorial sign-off. Upstream statements and our recommendations are distinguished.