[MM] Simplify _apply_hf_processor_main (#53275)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
@@ -51,12 +51,10 @@ from vllm.multimodal.processing import BaseDummyInputsBuilder
|
||||
from vllm.multimodal.processing.processor import (
|
||||
BaseMultiModalProcessor,
|
||||
BaseProcessingInfo,
|
||||
MultiModalProcessingInfo,
|
||||
ProcessorInputs,
|
||||
PromptReplacement,
|
||||
PromptUpdate,
|
||||
PromptUpdateDetails,
|
||||
TimingContext,
|
||||
)
|
||||
from vllm.platforms import current_platform
|
||||
from vllm.sequence import IntermediateTensors
|
||||
@@ -213,6 +211,9 @@ class PixtralDummyInputsBuilder(BaseDummyInputsBuilder[PixtralProcessingInfo]):
|
||||
|
||||
|
||||
class PixtralMultiModalProcessor(BaseMultiModalProcessor[PixtralProcessingInfo]):
|
||||
# The tokens are already inserted by the chat template
|
||||
hf_processor_applies_updates = True
|
||||
|
||||
def _get_mm_fields_config(
|
||||
self,
|
||||
hf_inputs: Mapping[str, NestedTensors],
|
||||
@@ -220,24 +221,35 @@ class PixtralMultiModalProcessor(BaseMultiModalProcessor[PixtralProcessingInfo])
|
||||
) -> Mapping[str, MultiModalFieldConfig]:
|
||||
return dict(images=MultiModalFieldConfig.batched("image"))
|
||||
|
||||
def _call_hf_processor(
|
||||
def _apply_hf_processor_main(
|
||||
self,
|
||||
prompt: str,
|
||||
mm_data: Mapping[str, object],
|
||||
mm_kwargs: Mapping[str, object],
|
||||
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))
|
||||
|
||||
prompt_text = self.dummy_inputs.get_dummy_text(mm_items.get_all_counts())
|
||||
|
||||
outputs = self.info.ctx.call_hf_processor(
|
||||
self.info.get_hf_processor(**mm_kwargs),
|
||||
dict(text=prompt, **mm_data),
|
||||
self.info.get_hf_processor(**hf_processor_mm_kwargs),
|
||||
dict(text=prompt_text, **mm_data),
|
||||
# Avoid padding issue
|
||||
dict(**mm_kwargs, return_tensors=None),
|
||||
dict(**hf_processor_mm_kwargs, return_tensors=None),
|
||||
)
|
||||
|
||||
# Missing batch dimension
|
||||
if is_list_of(outputs["input_ids"], int):
|
||||
outputs["input_ids"] = [outputs["input_ids"]]
|
||||
|
||||
return outputs
|
||||
processed_data = outputs
|
||||
processed_data.update(passthrough_data)
|
||||
return processed_data
|
||||
|
||||
def _get_prompt_updates(
|
||||
self,
|
||||
@@ -273,16 +285,6 @@ class PixtralMultiModalProcessor(BaseMultiModalProcessor[PixtralProcessingInfo])
|
||||
),
|
||||
]
|
||||
|
||||
def _cached_apply_hf_processor(
|
||||
self,
|
||||
inputs: ProcessorInputs,
|
||||
timing_ctx: TimingContext,
|
||||
) -> tuple[list[int], MultiModalProcessingInfo, bool]:
|
||||
prompt_ids, mm_info, _ = super()._cached_apply_hf_processor(inputs, timing_ctx)
|
||||
|
||||
# NOTE: The tokens are already inserted by the chat template
|
||||
return prompt_ids, mm_info, True
|
||||
|
||||
|
||||
@MULTIMODAL_REGISTRY.register_processor(
|
||||
PixtralMultiModalProcessor,
|
||||
|
||||
Reference in New Issue
Block a user