A 35-second audio clip killed the vLLM engine for every request that followed
vLLM PR #57769 (merged 2026-09-21) fixes offline `LLM.generate` with Whisper crashing on clips longer than 30 s with `RuntimeError: The size of tensor a (1750) must match the size of tensor b (1500)`, after which every later request returns `EngineDeadError`. The cause chains back to #53275, which passed `truncation=False` to HF processors so placeholder text would not be cut; `WhisperProcessor` forwards that to `WhisperFeatureExtractor`, which then stops truncating audio to its 30 s window, so a 35 s clip becomes 3500 frames and 1750 post-conv positions against a 1500-position encoder. v0.26.0 handled the same input. The server transcription endpoints are unaffected because they chunk to 30 s first, so this only bites offline users.
Source
↳ Follow the thread