[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
+18 -28
View File
@@ -8,7 +8,7 @@ from typing import Any, ClassVar
import torch
import torch.nn.functional as F
from torch import nn
from transformers import BatchFeature, PretrainedConfig
from transformers import PretrainedConfig
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, ModelConfig, SpeechToTextConfig, VllmConfig
@@ -50,7 +50,10 @@ from vllm.multimodal.processing import (
PromptUpdate,
TimingContext,
)
from vllm.multimodal.processing.processor import MultiModalProcessingInfo
from vllm.multimodal.processing.processor import (
HFMultiModalInputs,
MultiModalProcessingResult,
)
from vllm.renderers import TokenizeParams
from vllm.tokenizers import cached_tokenizer_from_config
from vllm.transformers_utils.processors.cohere_asr import (
@@ -1962,42 +1965,29 @@ class CohereASRMultiModalProcessor(EncDecMultiModalProcessor[CohereASRProcessing
) -> list[int]:
return [0]
def _get_hf_processor_text(self, mm_counts: Mapping[str, int]) -> str:
def _get_hf_mm_text(self, mm_counts: Mapping[str, int]) -> str:
return self.dummy_inputs.get_dummy_text(mm_counts)
def _preprocess_hf_mm_data(
def _get_hf_mm_inputs(
self,
mm_data: Mapping[str, object],
hf_processor_mm_kwargs: Mapping[str, object],
) -> tuple[Mapping[str, object], Mapping[str, object]]:
feature_extractor = self.info.get_feature_extractor(**hf_processor_mm_kwargs)
mm_items: MultiModalDataItems,
hf_kwargs: Mapping[str, object],
) -> HFMultiModalInputs:
hf_inputs = super()._get_hf_mm_inputs(mm_items, hf_kwargs)
mm_data = dict(mm_data)
mm_data["audio"] = mm_data.pop("audios")
hf_processor_mm_kwargs = dict(
**hf_processor_mm_kwargs,
sampling_rate=feature_extractor.sampling_rate,
feature_extractor = self.info.get_feature_extractor(**hf_kwargs)
return hf_inputs._replace(
hf_kwargs=dict(
hf_inputs.hf_kwargs,
sampling_rate=feature_extractor.sampling_rate,
)
)
return mm_data, hf_processor_mm_kwargs
def _postprocess_hf_mm_data(
self,
mm_data: Mapping[str, object],
hf_processor_mm_kwargs: Mapping[str, object],
processed_data: BatchFeature,
) -> BatchFeature:
if "labels" in processed_data:
processed_data["input_ids"] = processed_data.pop("labels")
return processed_data
def _cached_apply_hf_processor(
self,
inputs: ProcessorInputs,
timing_ctx: TimingContext,
) -> MultiModalProcessingInfo:
) -> MultiModalProcessingResult:
# Dithering injects noise into the extracted features, so the
# feature extractor is not a pure function of its input. Since the
# processing cache assumes that processor outputs are invariant