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, 10:00 UTC
[bugfix] refactor FunASR's _get_data_parser (#34397)
Signed-off-by: zixiao <shunli.dsl@alibaba-inc.com> Co-authored-by: zixiao <shunli.dsl@alibaba-inc.com>
AllenDou committed
Feb 12, 2026 at 15:26 UTC
386bfe5d08103f570a3aa03055372cbd33cf41ca
1 file changed
+7
-11
vllm/model_executor/models/funasr.py
+7
-11
@@ -714,10 +714,6 @@ class FunASRProcessingInfo(BaseProcessingInfo):
714
def get_hf_config(self) -> Qwen3Config:
715
return self.ctx.get_hf_config(Qwen3Config)
716
717
- @property
718
- def skip_prompt_length_check(self) -> bool:
719
- return True # Because the encoder prompt is padded
720
-
717
def get_supported_mm_limits(self) -> Mapping[str, int | None]:
718
return {"audio": 1}
719
@@ -727,6 +723,13 @@ class FunASRProcessingInfo(BaseProcessingInfo):
723
assert isinstance(feature_extractor, FunASRFeatureExtractor)
724
return feature_extractor
725
726
+ def get_data_parser(self) -> MultiModalDataParser:
727
+ feature_extractor = self.get_feature_extractor()
728
+ return MultiModalDataParser(
729
+ target_sr=feature_extractor.sampling_rate,
730
+ target_channels=self.get_target_channels(),
731
+ )
732
+
733
def get_target_channels(self) -> int:
734
return 1
735
@@ -765,13 +768,6 @@ class FunASRDummyInputsBuilder(BaseDummyInputsBuilder[FunASRProcessingInfo]):
768
769
770
class FunASRMultiModalProcessor(BaseMultiModalProcessor[FunASRProcessingInfo]):
768
- def _get_data_parser(self) -> MultiModalDataParser:
769
- feature_extractor = self.info.get_feature_extractor()
770
- return MultiModalDataParser(
771
- target_sr=feature_extractor.sampling_rate,
772
- target_channels=self.info.get_target_channels(),
773
- )
774
-
771
def _call_hf_processor(
772
self,
773
prompt: str,