[qgallouedec]

Small low priority contribution:

Using the same token in forced_eos_token_id/forced_bos_token_id and suppress_tokens is contradictory: at the forcing step every logit but the forced token is set to -inf, then the shared token is suppressed to -inf too — leaving all logits -inf, producing nan probabilities and the cryptic RuntimeError: probability tensor contains inf/nan.

GenerationConfig.validate() now rejects this with an actionable error. It fires only when every forced token is also suppressed (a guaranteed crash); partial overlap leaves a surviving forced token and is allowed. This also covers generate() (via validate()) and save_pretrained.

Before

model.generate(**inputs, forced_eos_token_id=eos, suppress_tokens=[eos])
# RuntimeError: probability tensor contains either `inf`, `nan` or element < 0

After

model.generate(**inputs, forced_eos_token_id=eos, suppress_tokens=[eos])
# ValueError: Every token in `forced_eos_token_id` ([2]) is also in `suppress_tokens`.
# Forcing a token while suppressing it sets all logits to `-inf` at the forcing step, which
# produces `nan` probabilities and crashes generation. Remove the overlapping token(s) from
# either `forced_eos_token_id` or `suppress_tokens` (if you meant to prevent an early EOS
# token, use `min_new_tokens` instead).

Still allowed — partial overlap leaves a usable forced token, so no crash:

GenerationConfig(forced_eos_token_id=[2, 5], suppress_tokens=[2])  # ok

Closes #24099.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[github-actions]

---

CI recap

Dashboard: View test results in Grafana
Latest run: 30142558377:1
Result: success | Jobs: 16 | Tests: 165,742 | Failures: 0 | Duration: 14h 22m