biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Renderer] Move InputPreprocessor into Renderer (2/2) (#34560)

Browse Source 
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2026-02-17 21:29:01 +08:00
committed by GitHub
parent c61a98f529
commit 574fe75245
32 changed files with 984 additions and 1054 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
vllm/model_executor/models/qwen2_audio.py
View File

@@ -59,7 +59,6 @@ from vllm.multimodal.processing import (
BaseProcessingInfo,
PromptReplacement,
PromptUpdate,
PromptUpdateDetails,
)
from vllm.sequence import IntermediateTensors
from vllm.utils.tensor_schema import TensorSchema, TensorShape
@@ -187,8 +186,10 @@ class Qwen2AudioDummyInputsBuilder(BaseDummyInputsBuilder[Qwen2AudioProcessingIn
hf_processor = self.info.get_hf_processor()
audio_token = hf_processor.audio_token
audio_bos_token = hf_processor.audio_bos_token
audio_eos_token = hf_processor.audio_eos_token
return audio_token * num_audios
return (audio_bos_token + audio_token + audio_eos_token) * num_audios
def get_dummy_mm_data(
self,
@@ -262,17 +263,7 @@ class Qwen2AudioMultiModalProcessor(BaseMultiModalProcessor[Qwen2AudioProcessing
out_mm_kwargs: MultiModalKwargsItems,
) -> Sequence[PromptUpdate]:
processor = self.info.get_hf_processor(**hf_processor_mm_kwargs)
tokenizer = self.info.get_tokenizer()
vocab = tokenizer.get_vocab()
# Use getattr with default to be compatible with transformers<4.48
audio_token = getattr(processor, "audio_token", "<|AUDIO|>")
audio_bos_token = getattr(processor, "audio_bos_token", "<|audio_bos|>")
audio_eos_token = getattr(processor, "audio_eos_token", "<|audio_eos|>")
audio_token_id = vocab[audio_token]
audio_bos_id = vocab[audio_bos_token]
audio_eos_id = vocab[audio_eos_token]
audio_token_id = processor.audio_token_id
out_mm_data = out_mm_kwargs.get_data()
feature_attention_mask = out_mm_data.get("feature_attention_mask")
@@ -303,17 +294,12 @@ class Qwen2AudioMultiModalProcessor(BaseMultiModalProcessor[Qwen2AudioProcessing
"to be represented inside the model"
)
audio_tokens = [audio_token_id] * num_features
return PromptUpdateDetails.select_token_id(
[audio_bos_id] + audio_tokens + [audio_eos_id],
embed_token_id=audio_token_id,
)
return [audio_token_id] * num_features
return [
PromptReplacement(
modality="audio",
target=audio_token,
target=[audio_token_id],
replacement=get_replacement_qwen2_audio,
)
]