From 4bca522f01ca49f04bb9a6cfa14c7c8839013b0c Mon Sep 17 00:00:00 2001 From: ChenWenjing <54166744+Shirley125@users.noreply.github.com> Date: Tue, 12 May 2026 01:09:10 +0800 Subject: [PATCH] [bugfix][ci] avoid Whisper transcript deduplication in realtime audio test (#3417) Signed-off-by: CHEN <116010019@link.cuhk.edu.cn> --- .../openai_api/test_qwen3_omni_realtime_websocket.py | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) diff --git a/tests/entrypoints/openai_api/test_qwen3_omni_realtime_websocket.py b/tests/entrypoints/openai_api/test_qwen3_omni_realtime_websocket.py index 90f8897c5..00a6565d9 100644 --- a/tests/entrypoints/openai_api/test_qwen3_omni_realtime_websocket.py +++ b/tests/entrypoints/openai_api/test_qwen3_omni_realtime_websocket.py @@ -30,7 +30,9 @@ os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn" MODEL = "Qwen/Qwen3-Omni-30B-A3B-Instruct" # Synthetic input for realtime E2E (``generate_synthetic_audio``); distinct cache file per phrase. -REALTIME_SYNTH_PHRASE_TEXT = "Translate into Chinese: Beijing is the Capital of China" +REALTIME_SYNTH_PHRASE_TEXT = ( + "Translate into Chinese: Beijing is the Capital of China. It is the center of culture and politics" +) # The new-schema CI overlay bakes in async_chunk: False and covers CUDA/ROCm/XPU # via its ``platforms:`` section, so one path serves all three. @@ -201,7 +203,7 @@ class TestQwen3OmniRealtimeWebSocket: assert whisper_text, "Whisper returned empty string for synthesized output audio" sim = cosine_similarity_text(whisper_text.lower(), final_text.lower()) - assert sim > 0.9, ( + assert sim > 0.8, ( f"Output audio transcript should match model text (sim={sim:.3f}): " f"whisper={whisper_text!r}, model_text={final_text!r}" )