From 1b4e8e53f87b2c6f5cd30d0eace501d7d2192236 Mon Sep 17 00:00:00 2001 From: Cyrus Leung Date: Fri, 13 Feb 2026 14:43:53 +0800 Subject: [PATCH] [CI/Build] Fix CUDA re-initialization error in distributed model tests (#34491) Signed-off-by: DarkLight1337 --- tests/models/multimodal/generation/test_voxtral_realtime.py | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/tests/models/multimodal/generation/test_voxtral_realtime.py b/tests/models/multimodal/generation/test_voxtral_realtime.py index 2b769e3ed3..ebd979ddb6 100644 --- a/tests/models/multimodal/generation/test_voxtral_realtime.py +++ b/tests/models/multimodal/generation/test_voxtral_realtime.py @@ -14,7 +14,6 @@ from mistral_common.tokens.tokenizers.mistral import MistralTokenizer from vllm import LLM, EngineArgs, SamplingParams from vllm.assets.audio import AudioAsset from vllm.engine.arg_utils import AsyncEngineArgs -from vllm.model_executor.models.voxtral_realtime import VoxtralRealtimeBuffer from vllm.v1.engine.async_llm import AsyncLLM MODEL_NAME = "mistralai/Voxtral-Mini-4B-Realtime-2602" @@ -114,6 +113,9 @@ def test_voxtral_realtime_forward(audio_assets, tokenizer, engine): @pytest.mark.asyncio async def test_voxtral_realtime_generator(audio_assets, tokenizer, async_engine): + # Lazy import to avoid CUDA-reinitialization error + from vllm.model_executor.models.voxtral_realtime import VoxtralRealtimeBuffer + sampling_params = SamplingParams(temperature=0.0, max_tokens=1) audio_config = tokenizer.instruct_tokenizer.audio_encoder.audio_config