Bug: Broken error messages in token normalization—`{token!r}` shown literally instead of actual token value

[markknoffler]

Description

In _normalize_token, when a stop_token or forbidden_token string maps to multiple token IDs, the code raises a ValueError. The error message uses {token!r} in a regular string instead of an f-string, so users see the literal {token!r} instead of the actual invalid token value. This makes it harder to debug misconfigured stop/forbidden tokens.

Example error shown:

ValueError: Invalid token: {token!r}. `stop_token`s and `forbidden_token`s must map to single token ids in the vocab.

Instead of:

ValueError: Invalid token: 'hello world'. `stop_token`s and `forbidden_token`s must map to single token ids in the vocab.

Locations

1. gemma/gm/text/_sampler.py

2. gemma/research/t5gemma/sampling.py

Steps to reproduce

1. Create a tokenizer (e.g. Tokenizer.from_version(3)).
2. Call _normalize_token(tokenizer, "hello world") (or any string that tokenizes to more than one token).
3. The ValueError message will contain the literal {token!r} instead of the actual token string.

[stanley1208]

This is fixed in my PR #542 — the f prefix was missing from the error message string in _normalize_token().