Fix IndexError in SequenceFeatureExtractor.pad on all-empty batch

[Osamaali313]

What does this PR do?

SequenceFeatureExtractor.pad (src/transformers/feature_extraction_sequence_utils.py) — the padding entry point for every audio feature extractor (Wav2Vec2, Whisper, CLAP, EnCodec, Gemma3n, … 30 subclasses) — crashes with IndexError when the whole batch is empty.

When the first element of the batch is an empty list/tuple, pad scans for the first non-empty element to infer the input type:

first_element = required_input[0]
if isinstance(first_element, (list, tuple)):
    # first_element might be an empty list/tuple in some edge cases so we grab the first non empty element.
    index = 0
    while len(required_input[index]) == 0:   # unbounded
        index += 1
    if index < len(required_input):
        first_element = required_input[index][0]

The while loop has no upper bound, so if every element in the batch is empty it increments index past the end and raises IndexError: list index out of range — before the following if index < len(required_input): guard (which was clearly meant to handle exactly this case) can ever run.

Evidence this is unintended

The analogous code in PreTrainedTokenizerBase.pad (tokenization_utils_base.py) does the same "grab the first non-empty element" with a bounded loop, and its comment explicitly notes that an all-empty batch should just be left alone:

first_element = required_input[0]
if isinstance(first_element, (list, tuple)):
    # first_element might be an empty list/tuple in some edge cases so we grab the first non empty element.
    for item in required_input:
        if len(item) > 0:
            first_element = item[0]
            break
    # At this state, if `first_element` is still a list/tuple, it's an empty one so there is nothing to do.

Fix

Bound the loop so the existing guard is reached and an all-empty batch is handled gracefully instead of crashing:

while index < len(required_input) and len(required_input[index]) == 0:
    index += 1

Reproduction

from transformers import Wav2Vec2FeatureExtractor
from transformers.feature_extraction_utils import BatchFeature

fe = Wav2Vec2FeatureExtractor(feature_size=1, sampling_rate=16000, padding_value=0.0, return_attention_mask=False)
fe.pad(BatchFeature({"input_values": [[], []]}), padding="longest")

	result
before	IndexError: list index out of range ❌
after	returns a BatchFeature with 2 (empty) rows ✅

A regression test is added in tests/utils/test_feature_extraction_utils.py.

Before submitting

• This PR fixes a bug.
• Added a regression test (SequenceFeatureExtractorPadTester.test_pad_all_empty_batch_does_not_raise).

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[github-actions]

---

CI recap

Dashboard: View test results in Grafana
Latest run: 29606181675:2
Result: success | Jobs: 15 | Tests: 169,057 | Failures: 0 | Duration: 14h 46m

[Rocketknight1]

Hey, can you stop making these PRs with your code agent? They're 'theoretical bugs' found by static analysis that never really trigger in practice (this one only fails on invalid inputs, so it's unclear what we should even return in this case). Telling a code agent to find bugs and fix them almost always creates PRs like this which spam maintainers and make open-source maintenance more difficult without improving the code in any real way.

[Rocketknight1]

Also we are ML engineers at a leading AI company so we do not need users to prompt Copilot for us. We are able to do that ourselves if we want to 😅

[Osamaali313]

Thanks, and point taken — this is fair feedback.

To be transparent: this came from an automated code-maintenance/bug-fixing pipeline I've been testing for long-running sessions, not a person hand-picking your repo. You've surfaced two real problems with it: (1) it's surfacing low-value, largely theoretical findings that aren't worth a maintainer's time, and (2) it re-pushed to a PR after it was closed, which it should never do. Both are going straight into the pipeline's guardrails.

Concretely, I'm removing the HuggingFace organizations from the pipeline's target list so you won't get more of these, and I'll close this PR now. Apologies for the noise, and thanks for taking the time to push back — that's the most useful signal I could've gotten.