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, 11:15 UTC
[Bugfix] send None sentinel on final commit so server properly sends transcription.done (#33963)
Signed-off-by: pjs102793 <pjs102793@naver.com> Co-authored-by: Nick Hill <nickhill123@gmail.com>
Junseo Park committed
Feb 12, 2026 at 06:01 UTC
5458eb835d66323a11d4a252ad551d001ce00ac8
2 files changed
+2
-8
tests/entrypoints/openai/test_realtime_validation.py
+1
-1
@@ -129,5 +129,5 @@ async def test_multi_chunk_streaming(
129
" First words I spoke in the original phonograph."
130
" A little piece of practical poetry. Mary had a little lamb,"
131
" it sleeps with quite a flow, and everywhere that Mary went,"
132
- " the lamb was sure to go"
132
+ " the lamb was sure to go."
133
)
vllm/entrypoints/openai/realtime/connection.py
+1
-7
@@ -48,7 +48,6 @@ class RealtimeConnection:
48
self.generation_task: asyncio.Task | None = None
49
50
self._is_connected = False
51
- self._is_input_finished = False
51
self._is_model_validated = False
52
53
self._max_audio_filesize_mb = envs.VLLM_MAX_AUDIO_CLIP_FILESIZE_MB
@@ -145,7 +144,7 @@ class RealtimeConnection:
144
commit_event = InputAudioBufferCommit(**event)
145
# final signals that the audio is finished
146
if commit_event.final:
148
- self._is_input_finished = True
147
+ self.audio_queue.put_nowait(None)
148
else:
149
await self.start_generation()
150
else:
@@ -239,11 +238,6 @@ class RealtimeConnection:
238
# finish because websocket connection was killed
239
break
240
242
- if self.audio_queue.empty() and self._is_input_finished:
243
- # finish because client signals that audio input
244
- # is finished
245
- break
246
-
241
usage = UsageInfo(
242
prompt_tokens=prompt_token_ids_len,
243
completion_tokens=completion_tokens_len,