T5Tokenizer.from_pretrained(model) -> OSError: Repo id must use alphanumeric chars ... 'T5ForConditionalGeneration(...)'

问题

Fine-tuning script does model = T5ForConditionalGeneration.from_pretrained("google-t5/t5-small", dtype=T.float16, device_map="auto") and then tokenizer = T5Tokenizer.from_pretrained(model) — passing the model OBJECT where a repo-id string (or local path) is expected. The object's repr (T5ForConditionalGeneration(...), hundreds of lines of module structure) is used as the repo id, fails repo-id validation, and transformers re-raises it as OSError.

报错原文（已脱敏/规范化）

huggingface_hub.errors.HFValidationError: Repo id must use alphanumeric chars, '-', '_' or '.'. The name cannot start or end with '-' or '.' and the maximum length is 96: 'T5ForConditionalGeneration( ... )'
then wrapped by transformers as:
OSError: Repo id must use alphanumeric chars, '-', '_' or '.'. The name cannot start or end with '-' or '.' and the maximum length is 96: 'T5ForConditionalGeneration( ... )'

适用环境

Python 3.12, huggingface_hub 2.2.0 (verified). Original report: transformers 5.13.0, Python 3.14. Linux container.

如何确认是这个问题

from huggingface_hub.utils import validate_repo_id; validate_repo_id("T5ForConditionalGeneration(...)") → expect: HFValidationError: Repo id must use alphanumeric chars, '-', '_' or '.' ...: 'T5ForConditionalGeneration(...)'.

不是这个问题时 / 边界

If the message is 'Repo id must be a string, not <class ...>' you are on huggingface_hub >= 2.x with the non-str check — same root cause (object instead of string), newer message. If from_pretrained fails with 401/404, that is auth or a wrong repo name — different problem.

根因

from_pretrained's first positional argument is pretrained_model_name_or_path: a repo-id string or a local path. A model instance is not valid there; it gets stringified to its repr, which contains parentheses, spaces and dots and can never match the repo-id regex ([A-Za-z0-9-_.], max 96 chars, no leading/trailing -/.), so validate_repo_id rejects it. transformers/utils/hub.py wraps the HFValidationError as OSError (raise OSError(f"{e}") from e), which is why the traceback shows OSError.

修复步骤

1. Change T5Tokenizer.from_pretrained(model) to T5Tokenizer.from_pretrained("google-t5/t5-small") — pass the repo-id string, not the model object (this is the accepted answer on the SO thread). 2. Sanity-check the fixed argument against the exact failing check: validate_repo_id("google-t5/t5-small") → expect: no exception (verified).

如何验证

validate_repo_id("google-t5/t5-small") passes silently, while validate_repo_id("T5ForConditionalGeneration(...)") raises the exact reported HFValidationError. (Full from_pretrained would additionally download weights — not run here; the failure in the traceback occurs at the repo-id validation step, which is what was verified.)

试过没用的办法

The traceback prints the entire model structure (Embedding, T5Stack, T5Block, ...), which looks like a model-loading bug — it is not. That dump is just the repr of the object you passed, embedded in the error message. No need to reinstall transformers or change dtype/device_map.

如何回滚

One-line code change; revert the line.

风险提示

None — argument fix, no data or state changes.

验证范围

huggingface_hub + transformers from_pretrained; Oct 2026