A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 7, 02:45 UTC
[ROCm][CI] Making entrypoints more deterministic on ROCm (#36293)
Andreas Karatzas committed
Mar 6, 2026 at 21:04 UTC
58928475e4c1910df28548849734ba30d3ef4580
1 file changed
+8
-4
tests/entrypoints/openai/test_realtime_validation.py
+8
-4
@@ -13,7 +13,7 @@ import websockets
13
14
from vllm.assets.audio import AudioAsset
15
16
-from ...utils import RemoteOpenAIServer
16
+from ...utils import ROCM_ENV_OVERRIDES, ROCM_EXTRA_ARGS, RemoteOpenAIServer
17
from .conftest import add_attention_backend
18
19
MISTRAL_FORMAT_ARGS = [
@@ -23,7 +23,7 @@ MISTRAL_FORMAT_ARGS = [
23
"mistral",
24
"--load_format",
25
"mistral",
26
-]
26
+] + ROCM_EXTRA_ARGS
27
28
MODEL_NAME = "mistralai/Voxtral-Mini-4B-Realtime-2602"
29
@@ -77,7 +77,9 @@ async def test_multi_chunk_streaming(
77
78
add_attention_backend(server_args, rocm_aiter_fa_attention)
79
80
- with RemoteOpenAIServer(model_name, server_args) as remote_server:
80
+ with RemoteOpenAIServer(
81
+ model_name, server_args, env_dict=ROCM_ENV_OVERRIDES
82
+ ) as remote_server:
83
ws_url = _get_websocket_url(remote_server)
84
async with websockets.connect(ws_url) as ws:
85
# Receive session.created
@@ -178,7 +180,9 @@ async def test_empty_commit_does_not_crash_engine(
180
181
add_attention_backend(server_args, rocm_aiter_fa_attention)
182
181
- with RemoteOpenAIServer(model_name, server_args) as remote_server:
183
+ with RemoteOpenAIServer(
184
+ model_name, server_args, env_dict=ROCM_ENV_OVERRIDES
185
+ ) as remote_server:
186
ws_url = _get_websocket_url(remote_server)
187
188
# --- First connection: empty commit (no audio appended) ----------