[CI/Build] Stabilize MiniCPM-o mixed-input audio consistency test (#7710)

Signed-off-by: 0z5a <dezhen.lu@student.uni-tuebingen.de>

tests/e2e/online_serving/test_minicpmo_4_5.py

@@ -361,13 +361,20 @@ def test_mix_to_text_audio_001(omni_server, openai_client) -> None:
     """
     video_data_url = f"data:video/mp4;base64,{generate_synthetic_video(24, 24, 20)['base64']}"
     image_data_url = f"data:image/jpeg;base64,{generate_synthetic_image(24, 24)['base64']}"
-    audio_data_url = f"data:audio/wav;base64,{generate_synthetic_audio(5, 1)['base64']}"
+    # A loop of the default single word "test" can make the answer repeat
+    # to the token limit, leaving TTS/ASR to count a long run of identical words
+    # (#7630). Exercise speech perception with a meaningful phrase instead.
+    audio = generate_synthetic_audio(5, 1, phrase_text="The weather is sunny today.")
+    audio_data_url = f"data:audio/wav;base64,{audio['base64']}"
     messages = dummy_messages_from_mix_data(
         system_prompt=get_system_prompt(),
         video_data_url=video_data_url,
         image_data_url=image_data_url,
         audio_data_url=audio_data_url,
-        content_text=get_prompt("mix"),
+        content_text=(
+            "Briefly describe the image and video, and summarize what the voice says. "
+            "Answer in one short sentence without repeating the spoken phrase."
+        ),
     )
 
     request_config = {

tests/helpers/tests/test_assertions.py

@@ -328,6 +328,21 @@ def test_bounded_tail_after_complete_answer_passes():
     )
 
 
+def test_minicpmo_mix_runaway_repetition_still_fails():
+    # #7630, Buildkite 15299 lines 1472-1475: 49 copies in the answer,
+    # 112 in the transcript. This is not a bounded noise tail and must not
+    # be hidden by relaxing the shared audio/text similarity gate.
+    prefix = "A black background with some colorful patterns appears, accompanied by a voice saying "
+    expected = prefix + '"' + " ".join(["test"] * 49)
+    transcript = prefix + ", ".join(["test"] * 112)
+    response = OmniResponse(success=True, text_content=expected, audio_content=transcript)
+
+    assert assertions.cosine_similarity_text(transcript, expected) == pytest.approx(0.675451892050252)
+    assert not assertions._transcript_has_bounded_tail(transcript, expected)
+    with pytest.raises(AssertionError, match=assertions.AUDIO_MISMATCH_MESSAGE):
+        assertions.assert_omni_response(response, {"modalities": ["text", "audio"]}, "advanced_model")
+
+
 def test_bounded_tail_exact_match_passes():
     assert assertions._transcript_has_bounded_tail(
         "The squares in this image are black.",