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 8, 08:45 UTC
[CI/Build] Fix CUDA re-initialization error in distributed model tests (#34491)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Cyrus Leung committed
Feb 13, 2026 at 14:43 UTC
1b4e8e53f87b2c6f5cd30d0eace501d7d2192236
1 file changed
+3
-1
tests/models/multimodal/generation/test_voxtral_realtime.py
+3
-1
@@ -14,7 +14,6 @@ from mistral_common.tokens.tokenizers.mistral import MistralTokenizer
14
from vllm import LLM, EngineArgs, SamplingParams
15
from vllm.assets.audio import AudioAsset
16
from vllm.engine.arg_utils import AsyncEngineArgs
17
-from vllm.model_executor.models.voxtral_realtime import VoxtralRealtimeBuffer
17
from vllm.v1.engine.async_llm import AsyncLLM
18
19
MODEL_NAME = "mistralai/Voxtral-Mini-4B-Realtime-2602"
@@ -114,6 +113,9 @@ def test_voxtral_realtime_forward(audio_assets, tokenizer, engine):
113
114
@pytest.mark.asyncio
115
async def test_voxtral_realtime_generator(audio_assets, tokenizer, async_engine):
116
+ # Lazy import to avoid CUDA-reinitialization error
117
+ from vllm.model_executor.models.voxtral_realtime import VoxtralRealtimeBuffer
118
+
119
sampling_params = SamplingParams(temperature=0.0, max_tokens=1)
120
audio_config = tokenizer.instruct_tokenizer.audio_encoder.audio_config
121