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
[Voxstral Realtime] Enable tests (#33803)
Signed-off-by: Patrick von Platen <patrick.v.platen@gmail.com>
Patrick von Platen committed
Feb 12, 2026 at 18:43 UTC
1100a97621ebbf226488268f47d0252b789276e6
5 files changed
+22
-18
tests/entrypoints/openai/test_realtime_validation.py
+1
-11
@@ -27,15 +27,6 @@ MISTRAL_FORMAT_ARGS = [
27
MODEL_NAME = "mistralai/Voxtral-Mini-4B-Realtime-2602"
28
29
30
-def _audio_to_base64_pcm16(path: str, target_sr: int = 16000) -> str:
31
- """Load audio file, convert to PCM16 @ target sample rate, base64 encode."""
32
- audio, _ = librosa.load(path, sr=target_sr, mono=True)
33
- # Convert float32 [-1, 1] to int16 [-32768, 32767]
34
- audio_int16 = (audio * 32767).astype(np.int16)
35
- audio_bytes = audio_int16.tobytes()
36
- return base64.b64encode(audio_bytes).decode("utf-8")
37
-
38
-
30
def _get_websocket_url(server: RemoteOpenAIServer) -> str:
31
"""Convert HTTP URL to WebSocket URL for realtime endpoint."""
32
http_url = server.url_root
@@ -74,12 +65,11 @@ def mary_had_lamb_audio_chunks() -> list[str]:
65
66
@pytest.mark.asyncio
67
@pytest.mark.parametrize("model_name", [MODEL_NAME])
77
-@pytest.mark.skip(reason="Voxtral streaming is not yet public")
68
async def test_multi_chunk_streaming(
69
model_name, mary_had_lamb_audio_chunks, rocm_aiter_fa_attention
70
):
71
"""Test streaming multiple audio chunks before committing."""
82
- server_args = ["--enforce-eager"]
72
+ server_args = ["--enforce-eager", "--max-model-len", "2048"]
73
74
if model_name.startswith("mistralai"):
75
server_args += MISTRAL_FORMAT_ARGS
tests/models/multimodal/generation/test_voxtral_realtime.py
-2
@@ -73,7 +73,6 @@ def async_engine() -> AsyncLLM:
73
return AsyncLLM.from_engine_args(engine_args)
74
75
76
-@pytest.mark.skip(reason="Voxtral streaming is not yet public")
76
def test_voxtral_realtime_forward(audio_assets, tokenizer, engine):
77
audio_config = tokenizer.instruct_tokenizer.tokenizer.audio
78
@@ -218,7 +217,6 @@ class RealTimeAudioInput:
217
218
219
@pytest.mark.asyncio
221
-@pytest.mark.skip(reason="Voxtral streaming is not yet public")
220
async def test_voxtral_realtime_generator(audio_assets, tokenizer, async_engine):
221
sampling_params = SamplingParams(temperature=0.0, max_tokens=1)
222
tests/models/multimodal/processing/test_common.py
+7
@@ -441,6 +441,13 @@ def test_processing_correctness(
441
"Qwen-VL tokenizer requires downloading a font file from "
442
"servers that often refuse connections in CI"
443
)
444
+ if model_id == "mistralai/Voxtral-Mini-4B-Realtime-2602":
445
+ pytest.skip(
446
+ "Voxtral Realtime doesn't make use of any place-holder"
447
+ "tokens and hence cannot pass the processing "
448
+ "correctness test as is. Let's revisit adapting this "
449
+ "test once more realtime models exist."
450
+ )
451
if model_id == "internlm/Intern-S1-Pro":
452
# FIXME(Isotr0py): Fix later.
453
pytest.skip("Tokenization issue. Fix later")
tests/models/registry.py
+4
-5
@@ -1031,13 +1031,12 @@ _MULTIMODAL_EXAMPLE_MODELS = {
1031
),
1032
"VoxtralForConditionalGeneration": _HfExamplesInfo(
1033
"mistralai/Voxtral-Mini-3B-2507",
1034
- # disable this temporarily until we support HF format
1035
- is_available_online=False,
1034
+ tokenizer_mode="mistral",
1035
),
1036
"VoxtralRealtimeGeneration": _HfExamplesInfo(
1038
- "<place-holder>",
1039
- # disable this temporarily until we support HF format
1040
- is_available_online=False,
1037
+ "mistralai/Voxtral-Mini-4B-Realtime-2602",
1038
+ enforce_eager=True,
1039
+ tokenizer_mode="mistral",
1040
),
1041
# [Encoder-decoder]
1042
"NemotronParseForConditionalGeneration": _HfExamplesInfo(
vllm/model_executor/models/voxtral.py
+10
@@ -54,6 +54,7 @@ from vllm.multimodal.processing.processor import (
54
BaseMultiModalProcessor,
55
BaseProcessingInfo,
56
MultiModalProcessingInfo,
57
+ PlaceholderFeaturesInfo,
58
PromptReplacement,
59
PromptUpdate,
60
)
@@ -283,6 +284,15 @@ class VoxtralMultiModalProcessor(BaseMultiModalProcessor[VoxtralProcessingInfo])
284
) -> Mapping[str, MultiModalFieldConfig]:
285
return dict(audio_arrays=MultiModalFieldConfig.batched("audio"))
286
287
+ def _validate_mm_placeholders(
288
+ self,
289
+ mm_placeholders: Mapping[str, list[PlaceholderFeaturesInfo]],
290
+ mm_item_counts: Mapping[str, int],
291
+ ) -> None:
292
+ # mistral_common's tokenizer's does not follow HF's placeholder norms
293
+ # skip validation here
294
+ ...
295
+
296
def _get_prompt_updates(
297
self,
298
mm_items: MultiModalDataItems,