[ROCm][CI] Stabilize ROCm audio streaming test (#50163)
Signed-off-by: Andreas Karatzas <Andreas.Karatzas@amd.com>
Andreas Karatzas committed
Jul 28, 2026 at 18:58 UTC
176256b9628d2d8d22db41d7a72c15540f438026
1 file changed
+6
tests/entrypoints/multimodal/openai/chat_completion/test_audio.py
+6
@@ -10,6 +10,7 @@ import pytest_asyncio
10
from tests.utils import RemoteOpenAIServer
11
from vllm.assets.audio import AudioAsset
12
from vllm.multimodal.utils import encode_audio_base64, encode_audio_url, fetch_audio
13
+from vllm.platforms import current_platform
14
15
MODEL_NAME = "fixie-ai/ultravox-v0_5-llama-3_2-1b"
16
TEST_AUDIO_URLS = [
@@ -18,6 +19,10 @@ TEST_AUDIO_URLS = [
19
]
20
MAXIMUM_AUDIOS = 2
21
22
+# Disable prefix caching on ROCm to reduce non-determinism in
23
+# streaming-vs-non-streaming comparisons.
24
+_ROCM_ARGS = ["--no-enable-prefix-caching"] if current_platform.is_rocm() else []
25
+
26
27
@pytest.fixture(scope="module")
28
def server():
@@ -32,6 +37,7 @@ def server():
37
"--trust-remote-code",
38
"--limit-mm-per-prompt",
39
json.dumps({"audio": MAXIMUM_AUDIOS}),
40
+ *_ROCM_ARGS,
41
]
42
43
with RemoteOpenAIServer(MODEL_NAME, args) as remote_server: