[MM] Further cleanup _apply_hf_processor_main (#53610)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2026-09-16 10:21:00 -07:00
committed by GitHub
parent 6ca2b23e22
commit 0b9e018fb9
74 changed files with 1048 additions and 1323 deletions
+21 -21
View File
@@ -52,6 +52,7 @@ from vllm.multimodal.processing import (
PromptUpdateDetails,
cached_encode,
)
from vllm.multimodal.processing.processor import HFMultiModalInputs
from vllm.sequence import IntermediateTensors
from vllm.utils.gpu_sync_debug import gpu_sync_allowed
from vllm.utils.tensor_schema import TensorSchema, TensorShape
@@ -240,31 +241,32 @@ class Idefics3DummyInputsBuilder(BaseDummyInputsBuilder[Idefics3ProcessingInfo])
class Idefics3MultiModalProcessor(BaseMultiModalProcessor[Idefics3ProcessingInfo]):
def _apply_hf_processor_main(
def _get_hf_mm_inputs(
self,
mm_items: MultiModalDataItems,
hf_processor_mm_kwargs: Mapping[str, object],
) -> BatchFeature:
valid_mm_items = mm_items.select(
{k for k, c in mm_items.get_all_counts().items() if c > 0}
)
mm_data, passthrough_data = self._get_hf_mm_data(valid_mm_items)
if not mm_data:
return BatchFeature(dict(passthrough_data))
hf_kwargs: Mapping[str, object],
) -> HFMultiModalInputs:
hf_inputs = super()._get_hf_mm_inputs(mm_items, hf_kwargs)
image_processor = self.info.get_hf_processor().image_processor
if getattr(image_processor, "backend", "pil") == "pil":
hf_processor_mm_kwargs = {
"input_data_format": "channels_last",
**hf_processor_mm_kwargs,
}
hf_inputs = hf_inputs._replace(
hf_kwargs={
"input_data_format": "channels_last",
**hf_inputs.hf_kwargs,
}
)
processed_data = self.info.ctx.call_hf_processor(
self.info.get_hf_processor(**hf_processor_mm_kwargs),
mm_data,
hf_processor_mm_kwargs,
)
return hf_inputs
def _postprocess_hf_mm_data(
self,
mm_data: Mapping[str, object],
hf_processor_mm_kwargs: Mapping[str, object],
processed_data: BatchFeature,
) -> BatchFeature:
if not mm_data:
return processed_data
images = mm_data.get("images", [])
mm_items = self.info.parse_mm_data({"image": images}, validate=False)
@@ -289,8 +291,6 @@ class Idefics3MultiModalProcessor(BaseMultiModalProcessor[Idefics3ProcessingInfo
processed_data["pixel_values"].squeeze_(0)
processed_data["pixel_attention_mask"].squeeze_(0)
processed_data.update(passthrough_data)
return processed_data
def _get_mm_fields_config(