[ROCm][CI] Stabilize ROCm audio streaming test (#50163)

Signed-off-by: Andreas Karatzas <Andreas.Karatzas@amd.com>

tests/entrypoints/multimodal/openai/chat_completion/test_audio.py

@@ -10,6 +10,7 @@ import pytest_asyncio
 from tests.utils import RemoteOpenAIServer
 from vllm.assets.audio import AudioAsset
 from vllm.multimodal.utils import encode_audio_base64, encode_audio_url, fetch_audio
+from vllm.platforms import current_platform
 
 MODEL_NAME = "fixie-ai/ultravox-v0_5-llama-3_2-1b"
 TEST_AUDIO_URLS = [
@@ -18,6 +19,10 @@ TEST_AUDIO_URLS = [
 ]
 MAXIMUM_AUDIOS = 2
 
+# Disable prefix caching on ROCm to reduce non-determinism in
+# streaming-vs-non-streaming comparisons.
+_ROCM_ARGS = ["--no-enable-prefix-caching"] if current_platform.is_rocm() else []
+
 
 @pytest.fixture(scope="module")
 def server():
@@ -32,6 +37,7 @@ def server():
         "--trust-remote-code",
         "--limit-mm-per-prompt",
         json.dumps({"audio": MAXIMUM_AUDIOS}),
+        *_ROCM_ARGS,
     ]
 
     with RemoteOpenAIServer(MODEL_NAME, args) as remote_server: