[Misc] Clean up renderers (#36770)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2026-03-11 16:39:29 +00:00
committed by GitHub
parent c84b519cf3
commit 196802dfa6
12 changed files with 136 additions and 220 deletions
-12
View File
@@ -63,12 +63,10 @@ from vllm.multimodal.processing import (
BaseDummyInputsBuilder,
BaseMultiModalProcessor,
BaseProcessingInfo,
InputProcessingContext,
PromptReplacement,
PromptUpdate,
PromptUpdateDetails,
)
from vllm.renderers import TokenizeParams
from vllm.sequence import IntermediateTensors
from vllm.utils.tensor_schema import TensorSchema, TensorShape
@@ -546,9 +544,6 @@ class Llama4VisionModel(nn.Module):
class Mllama4ProcessingInfo(BaseProcessingInfo):
def __init__(self, ctx: InputProcessingContext) -> None:
super().__init__(ctx)
def get_hf_config(self) -> Llama4Config:
return self.ctx.get_hf_config(Llama4Config)
@@ -557,9 +552,6 @@ class Mllama4ProcessingInfo(BaseProcessingInfo):
Llama4Processor, use_fast=kwargs.pop("use_fast", True), **kwargs
)
def get_default_tok_params(self) -> TokenizeParams:
return super().get_default_tok_params().with_kwargs(add_special_tokens=False)
def get_supported_mm_limits(self) -> Mapping[str, int | None]:
# Although vLLM can support more images from an infra capability
# perspective, we do not recommend using >10 images in practice.
@@ -597,10 +589,6 @@ class Mllama4MultiModalProcessor(BaseMultiModalProcessor[Mllama4ProcessingInfo])
mm_kwargs: Mapping[str, object],
tok_kwargs: Mapping[str, object],
) -> BatchFeature:
tokenizer = self.info.get_tokenizer()
if mm_data is None:
return tokenizer(prompt, add_special_tokens=False) # exclude bos
processed_outputs = super()._call_hf_processor(
prompt=prompt,
mm_data=mm_data,