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

[VLM] Avoid unnecessary dummy multimodal data during processing (#16416)

Browse Source 
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2025-04-11 03:32:14 +08:00
committed by GitHub
parent dd143ef541
commit 56d4aefa33
33 changed files with 436 additions and 394 deletions
Download Patch File Download Diff File Expand all files Collapse all files

27
vllm/model_executor/models/deepseek_vl2.py
View File

@@ -19,14 +19,14 @@ from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler
from vllm.model_executor.model_loader.utils import set_default_torch_dtype
from vllm.multimodal import MULTIMODAL_REGISTRY
from vllm.multimodal.inputs import (MultiModalFieldConfig, MultiModalKwargs,
NestedTensors)
from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig,
MultiModalKwargs, NestedTensors)
from vllm.multimodal.parse import (ImageEmbeddingItems, ImageProcessorItems,
ImageSize, MultiModalDataItems)
from vllm.multimodal.processing import (BaseMultiModalProcessor,
BaseProcessingInfo, PromptReplacement,
PromptUpdate)
from vllm.multimodal.profiling import BaseDummyInputsBuilder, ProcessorInputs
from vllm.multimodal.profiling import BaseDummyInputsBuilder
from vllm.sequence import IntermediateTensors
from vllm.transformers_utils.configs.deepseek_vl2 import (DeepseekVLV2Config,
MlpProjectorConfig,
@@ -172,29 +172,30 @@ class DeepseekVL2ProcessingInfo(BaseProcessingInfo):
class DeepseekVL2DummyInputsBuilder(
BaseDummyInputsBuilder[DeepseekVL2ProcessingInfo]):
def get_dummy_processor_inputs(
def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
num_images = mm_counts.get("image", 0)
processor = self.info.get_hf_processor()
image_token = processor.image_token
return image_token * num_images
def get_dummy_mm_data(
self,
seq_len: int,
mm_counts: Mapping[str, int],
) -> ProcessorInputs:
) -> MultiModalDataDict:
num_images = mm_counts.get("image", 0)
hf_processor = self.info.get_hf_processor()
image_token: str = hf_processor.image_token
max_image_size = self.info.get_image_size_with_most_features()
mm_data = {
return {
"image":
self._get_dummy_images(width=max_image_size.width,
height=max_image_size.height,
num_images=num_images)
}
return ProcessorInputs(
prompt_text=image_token * num_images,
mm_data=mm_data,
)
class DeepseekVL2MultiModalProcessor(
BaseMultiModalProcessor[DeepseekVL2ProcessingInfo]):