VideoCaptionTopologyBench is the smallest Video → Frame → Video reference. Every video expands into ordered frames, READY frames from different videos can share Caption actor batches, and completed captions reduce by source video and frame order into one summary record.

Topology

Run

from rayorch.benchmark import VideoCaptionTopologyBench

report = VideoCaptionTopologyBench(
    output_dir="./results",
    video_count=8,
    frames_per_video=16,
    workers=4,
    batch_size=8,
    input_batch_size=1,
    max_active_input_batches=4,
).run(profile=False)

The case generates logical videos and frames; it reads no media and loads no VLM.

Outputs and metrics

Each output contains the video name, ordered frame IDs, and deterministic caption strings. The report adds total frames.

What performance it demonstrates

Observation	Experiment	Question answered
frame scheduling throughput	frames / measured_wall_s	control-plane rate for a 1 → M → 1 workload
cross-video batching	increase video_count and active input batches	whether more inputs fill Caption batches more effectively
ordered-reduction cost	increase frames_per_video	how reduction cost changes with larger child groups
Worker/RPC tradeoff	scan workers and batch_size	whether more actors help when UDF work is lightweight

Use it to validate data shape and scheduler parameters before integrating a real video model. Its numbers do not represent decode or VLM inference performance.

Contributors: Sunnyhaze

Previous pageNested Document

Next pageVideo Multimodal