[Bugfix][Voxtral Realtime] Fix token feedback timeout silent hang (#44461)
Signed-off-by: Ting Sun <suntcrick@gmail.com>
Ting SUN committed
Jul 5, 2026 at 19:42 UTC
8974ed89cd0b17615ae48cf09ef824cd8e3ec521
1 file changed
+3
-6
vllm/model_executor/models/voxtral_realtime.py
+3
-6
@@ -17,7 +17,6 @@ from vllm.compilation.decorators import support_torch_compile
17
from vllm.config import ModelConfig, SpeechToTextConfig, VllmConfig
18
from vllm.config.speech_to_text import SpeechToTextParams
19
from vllm.engine.protocol import StreamingInput
20
-from vllm.envs import VLLM_ENGINE_ITERATION_TIMEOUT_S
20
from vllm.inputs import PromptType, TokensPrompt
21
from vllm.logger import init_logger
22
from vllm.model_executor.models.interfaces import MultiModalEmbeddings, SupportsRealtime
@@ -265,13 +264,11 @@ class VoxtralRealtimeGeneration(VoxtralForConditionalGeneration, SupportsRealtim
264
await buffer.append_audio(right_pad.audio_array)
265
await buffer.append_audio(None) # signal end
266
268
- # Feed output tokens back into buffer in background
267
+ # Feed output tokens back into the buffer. Idle waits are normal here;
268
+ # request cleanup still cancels this task through the finally block.
269
async def feed_tokens():
270
while True:
271
- all_outputs = await asyncio.wait_for(
272
- input_stream.get(),
273
- timeout=VLLM_ENGINE_ITERATION_TIMEOUT_S,
274
- )
271
+ all_outputs = await input_stream.get()
272
await buffer.append_tokens(all_outputs[-1:])
273
274
audio_task = asyncio.create_task(feed_audio())