Issue evidence · Serve · vllm-project/vllm

/v1/realtime drops session.update prompt and language, so Qwen3-ASR realtime always uses a fixed prompt

GitHub issue: https://github.com/vllm-project/vllm/issues/61026

Reproduced

A prompt or language sent in session.update on /v1/realtime is silently ignored for Qwen3-ASR, though /v1/audio/transcriptions honors both.

What Badgr ran

Served Qwen/Qwen3-ASR-1.7B with the issue's realtime override on an RTX 4090 through Badgr (vLLM 0.31.1rc1.dev273, cu134 image) and ran the issue's script: four sessions with no options, language, prompt, and both, over the same 4.9 s clip.

What came back

  • prompt_tokens was 74 in all four sessions, so the server built the same prompt whether or not a prompt or language was sent.
  • No error came back for the ignored options. Every session returned a transcript beginning "language English<asr_text>".

One run on one host. This run used a vLLM nightly image, not the 0.30.0 wheel the issue also names.

Checked 2026-10-11 in Badgr’s local development environment.

← All issue evidence