[ydshieh]

Summary

Continuation of #47995 ([Whisper] Fix integration test failures on A10G).

Fixes test_whisper_longform_single_batch, one of the 5 tests left in Group 5 of that PR as requiring changes outside the test file.

---

Why the test passed before PR #40936 (commit 05c0e1d390)

Before 05c0e1d390, the base class PreTrainedTokenizerBase.batch_decode iterated per-sequence, calling decode() once per sequence with a single 1D list of token ids. decode() in WhisperTokenizer therefore always received a 1D input, so both decode_with_timestamps=True and False worked correctly for both single and batch inputs.

Root cause

PR #40936 (commit 05c0e1d390, "rm slow tokenizers", Nov 27 2025) changed the base class batch_decode to pass the full 2D batch (list of token id lists) to decode() in one call instead of iterating per-sequence. decode() in WhisperTokenizer was partially updated to handle the 2D input:

• ✅ The else branch (non-timestamp path) got isinstance(text, list) handling for _filter_timestamp_ids
• ❌ The decode_with_timestamps=True branch was missed — it still calls _decode_with_timestamps(filtered_ids, ...) unconditionally

When batch_decode(decode_with_timestamps=True) is called with a batch, filtered_ids is now a list of lists. _decode_with_timestamps iterates over it and immediately hits token >= timestamp_begin where token is a list → TypeError: '>=' not supported between instances of 'list' and 'int'.

Bisect confirmed: parent commit 01c51596ec passes, 05c0e1d390 fails. First seen in CI on 2025-11-28 (the day after the commit landed).

Why only this one test is affected

decode_with_timestamps=True is used in exactly one integration test in the entire whisper test file — test_whisper_longform_single_batch (line 2348). All other tests use either:

• output_offsets=True — the newer structured return format, a different code path entirely
• Plain batch_decode(skip_special_tokens=True) — no timestamp decoding

The other 4 tests remaining in Group 5 of #47995 have completely unrelated root causes and are not addressed here.

Fix

Add the same isinstance batch check to the decode_with_timestamps branch that was already applied to the else branch: detect a 2D input and process each sequence individually.

if isinstance(filtered_ids, list) and filtered_ids and isinstance(filtered_ids[0], list):
    # batched input: process each sequence individually
    text = [
        self._decode_with_timestamps(seq, ...)
        for seq in filtered_ids
    ]
else:
    text = self._decode_with_timestamps(filtered_ids, ...)

Test

WhisperModelIntegrationTests::test_whisper_longform_single_batch — verified passing on runner (A10G, torch 2.13 / CUDA 13.0).

🤖 Generated with Claude Code

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: whisper

[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: 31956662948:1
Result: success | Jobs: 16 | Tests: 181,622 | Failures: 0 | Duration: 17h 47m

[ydshieh]

Despite this touching a file outside the tests directory, the analysis shows the root cause and the fix is clearly necessary and correct. Merging directly.