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

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2025-04-10 19:32:14 +00:00
committed by GitHub
parent dd143ef541
commit 56d4aefa33
33 changed files with 436 additions and 394 deletions
+10 -11
View File
@@ -21,13 +21,14 @@ from vllm.model_executor.models.bart import (BartDecoder, BartEncoder,
BartScaledWordEmbedding)
from vllm.model_executor.sampling_metadata import SamplingMetadata
from vllm.multimodal import MULTIMODAL_REGISTRY
from vllm.multimodal.inputs import MultiModalFieldConfig, MultiModalKwargs
from vllm.multimodal.parse import MultiModalDataDict, MultiModalDataItems
from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig,
MultiModalKwargs)
from vllm.multimodal.parse import MultiModalDataItems
from vllm.multimodal.processing import (BaseProcessingInfo,
EncDecMultiModalProcessor,
PromptIndexTargets, PromptInsertion,
PromptUpdate)
from vllm.multimodal.profiling import BaseDummyInputsBuilder, ProcessorInputs
from vllm.multimodal.profiling import BaseDummyInputsBuilder
from vllm.sequence import IntermediateTensors
from .interfaces import (MultiModalEmbeddings, SupportsMultiModal,
@@ -772,27 +773,25 @@ class Florence2ProcessingInfo(BaseProcessingInfo):
class Florence2DummyInputsBuilder(
BaseDummyInputsBuilder[Florence2ProcessingInfo]):
def get_dummy_processor_inputs(
def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
return ""
def get_dummy_mm_data(
self,
seq_len: int,
mm_counts: Mapping[str, int],
) -> ProcessorInputs:
) -> MultiModalDataDict:
num_images = mm_counts.get("image", 0)
target_width = target_height = self.info.get_hf_config().projection_dim
mm_data = {
return {
"image":
self._get_dummy_images(width=target_width,
height=target_height,
num_images=num_images)
}
return ProcessorInputs(
prompt_text="",
mm_data=mm_data,
)
class Florence2MultiModalProcessor(
EncDecMultiModalProcessor[Florence2ProcessingInfo]):