vLLM caps Qwen-VL video sampling knobs a client could raise to force decoding every source frame

GitHub·medium signalXBlueskyLinkedInCopy link

Merged 2026-09-14, PR #56729 notes that Qwen2-VL and Qwen3-VL video samplers ignore num_frames and bound sampling on request-controlled max_frames and fps, which a client can raise through media_io_kwargs. The fix applies the class-level ceilings already used by the GLMGA sampler, so a request may lower the sample count but cannot push past 768 frames or 30 fps. The sink is compute_frames_index_to_sample, which OpenCV, TorchCodec, PyNvVideoCodec and DeepStream all call, and chat completions, embeddings, responses, tokenize, invocations, batch and offline LLM all reach it; the Rust frontend rejects media_io_kwargs and is unaffected.