[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
+24 -16
View File
@@ -28,6 +28,7 @@ from vllm.model_executor.layers.pooler.tokwise import pooler_for_token_embed
from vllm.model_executor.layers.pooler.tokwise.heads import TokenPoolerHead
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
from vllm.multimodal import MULTIMODAL_REGISTRY
from vllm.multimodal.parse import MultiModalDataItems
from .interfaces import SupportsLateInteraction
from .interfaces_base import default_pooling_type
@@ -59,24 +60,31 @@ class ColPaliProcessingInfo(PaliGemmaProcessingInfo):
class ColPaliMultiModalProcessor(PaliGemmaMultiModalProcessor):
"""Multimodal processor for ColPali."""
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:
if mm_data:
# The ColPali tokenizer_config.json ships with a small default
# max_length (50) that truncates the 1024 image tokens inserted
# by PaliGemmaProcessor, causing a token-count mismatch.
# vLLM enforces its own max_model_len, so we disable HF
# truncation to keep all image + text tokens intact.
return self.info.ctx.call_hf_processor(
self.info.get_hf_processor(**mm_kwargs),
dict(text=prompt, **mm_data),
dict(**mm_kwargs, truncation=False),
)
return super()._call_hf_processor(prompt, mm_data, mm_kwargs)
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))
# The ColPali tokenizer_config.json ships with a small default
# max_length (50) that truncates the 1024 image tokens inserted
# by PaliGemmaProcessor, causing a token-count mismatch.
# vLLM enforces its own max_model_len, so we disable HF
# truncation to keep all image + text tokens intact.
processed_data = self.info.ctx.call_hf_processor(
self.info.get_hf_processor(**hf_processor_mm_kwargs),
mm_data,
dict(**hf_processor_mm_kwargs, truncation=False),
)
processed_data.update(passthrough_data)
return processed_data
@default_pooling_type(seq_pooling_type="CLS", tok_pooling_type="ALL")