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 9, 00:00 UTC
[Voxtral models] Skip warm-up to skip confusing error message in warm-up (#33576)
Signed-off-by: Patrick von Platen <patrick.v.platen@gmail.com> Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk>
Patrick von Platen committed
Feb 3, 2026 at 16:22 UTC
f0d525171557e3fe74e8e6df52257f9d66831d3f
3 files changed
+10
-3
vllm/entrypoints/openai/translations/speech_to_text.py
+4
-3
@@ -138,6 +138,9 @@ class OpenAISpeechToText(OpenAIServing):
138
if not supports_transcription(self.model_cls):
139
return
140
141
+ if getattr(self.model_cls, "skip_warmup_audio_preprocessing", False):
142
+ return
143
+
144
try:
145
warmup_start = time.perf_counter()
146
logger.info("Warming up audio preprocessing libraries...")
@@ -150,9 +153,7 @@ class OpenAISpeechToText(OpenAIServing):
153
_ = librosa.get_duration(y=dummy_audio, sr=self.asr_config.sample_rate)
154
155
# Warm up mel-spectrogram computation with model-specific parameters
153
- from vllm.transformers_utils.processor import (
154
- cached_processor_from_config,
155
- )
156
+ from vllm.transformers_utils.processor import cached_processor_from_config
157
158
processor = cached_processor_from_config(self.model_config)
159
feature_extractor = None
vllm/model_executor/models/voxtral.py
+3
@@ -335,6 +335,9 @@ class VoxtralForConditionalGeneration(
335
nn.Module, SupportsMultiModal, SupportsPP, SupportsLoRA, SupportsTranscription
336
):
337
supported_languages = ISO639_1_SUPPORTED_LANGS
338
+ # transformers' currently has limited support for MistralCommon backend
339
+ # and cached_get_processor. Let's skip until fixed
340
+ skip_warmup_audio_preprocessing = True
341
342
packed_modules_mapping = {
343
"qkv_proj": ["q_proj", "k_proj", "v_proj"],
vllm/model_executor/models/voxtral_realtime.py
+3
@@ -218,6 +218,9 @@ class VoxtralRealtimeBuffer:
218
@support_torch_compile
219
class VoxtralRealtimeGeneration(VoxtralForConditionalGeneration, SupportsRealtime):
220
requires_raw_input_tokens = True
221
+ # transformers' currently has limited support for MistralCommon backend
222
+ # and cached_get_processor. Let's skip until fixed
223
+ skip_warmup_audio_preprocessing = True
224
225
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
226
super().__init__(vllm_config=vllm_config, prefix=prefix)