[MM] Simplify _apply_hf_processor_main (#53275)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2026-08-22 01:49:37 -07:00
committed by GitHub
parent 7ca49fbe4b
commit 0b19ebcacd
90 changed files with 1848 additions and 1706 deletions
+11 -17
View File
@@ -43,7 +43,6 @@ from vllm.multimodal.processing.processor import (
)
from vllm.sequence import IntermediateTensors
from vllm.tokenizers import cached_tokenizer_from_config
from vllm.tokenizers.hf import HfTokenizer
from vllm.transformers_utils.configs.deepseek_vl2 import (
DeepseekVLV2Config,
MlpProjectorConfig,
@@ -236,28 +235,23 @@ class DeepseekVL2DummyInputsBuilder(BaseDummyInputsBuilder[DeepseekVL2Processing
class DeepseekVL2MultiModalProcessor(
BaseMultiModalProcessor[DeepseekVL2ProcessingInfo]
):
def _call_hf_processor(
def _get_hf_processor_text(self, mm_counts: Mapping[str, int]) -> str:
return self.dummy_inputs.get_dummy_text(mm_counts)
def _postprocess_hf_mm_data(
self,
prompt: str,
mm_data: Mapping[str, object],
mm_kwargs: Mapping[str, object],
hf_processor_mm_kwargs: Mapping[str, object],
processed_data: BatchFeature,
) -> BatchFeature:
if not mm_data:
tokenizer = self.info.get_tokenizer()
assert isinstance(tokenizer, HfTokenizer)
return tokenizer(prompt, add_special_tokens=True, return_tensors="pt")
return processed_data
processed_outputs = super()._call_hf_processor(
prompt=prompt,
mm_data=mm_data,
mm_kwargs=mm_kwargs,
processed_data["num_patches"] = (
processed_data["images_spatial_crop"].prod(-1) + 1
)
processed_outputs["num_patches"] = (
processed_outputs["images_spatial_crop"].prod(-1) + 1
)
return processed_outputs
return processed_data
def _get_mm_fields_config(
self,
@@ -315,7 +309,7 @@ class DeepseekVL2MultiModalProcessor(
self,
inputs: ProcessorInputs,
timing_ctx: TimingContext,
) -> tuple[list[int], MultiModalProcessingInfo, bool]:
) -> MultiModalProcessingInfo:
# The processor logic is different for len(images) <= 2 vs > 2
# Since the processing cache assumes that the processor output is
# invariant of how many images are passed per prompt, we only