[Model] Support Unlimited OCR (#46564)

Signed-off-by: Tianyu Guo <guoty9@mail2.sysu.edu.cn>
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Co-authored-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Co-authored-by: Roger Wang <hey@rogerw.io>
This commit is contained in:
Tianyu Guo
2026-06-27 23:09:18 -07:00
committed by GitHub
co-authored by Isotr0py Roger Wang
parent a65f93fb2e
commit c6741b2ad4
35 changed files with 1084 additions and 49 deletions
@@ -161,10 +161,12 @@ class DeepseekOCRProcessor(ProcessorMixin):
image_size: int = IMAGE_SIZE,
base_size: int = BASE_SIZE,
strategy: Literal["v1", "v2"] = "v1",
max_crops: int = MAX_CROPS,
**kwargs,
):
self.image_size = image_size
self.base_size = base_size
self.max_crops = max_crops
# image token calculation strategy for
# Deepseek-OCR and Deepseek-OCR-2
@@ -332,7 +334,7 @@ class DeepseekOCRProcessor(ProcessorMixin):
crop_ratio = [1, 1]
elif cropping:
images_crop_raw, crop_ratio = dynamic_preprocess(
image, image_size=self.image_size
image, image_size=self.image_size, max_num=self.max_crops
)
else:
crop_ratio = [1, 1]
@@ -0,0 +1,46 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Image processor for Unlimited-OCR (baidu/Unlimited-OCR)."""
from PIL import Image
from vllm.logger import init_logger
from vllm.transformers_utils.processors.deepseek_ocr import DeepseekOCRProcessor
logger = init_logger(__name__)
class UnlimitedOCRProcessor(DeepseekOCRProcessor):
"""DeepseekOCRProcessor variant for Unlimited-OCR.
The only behavioural difference from the base processor is a multi-image
safeguard: when more than one image is present, crop ("gundam") mode is
disabled.
Because the effective crop flag then depends on *how many* images are in the
request, the per-item processing output is no longer invariant of sibling
images. ``UnlimitedOCRMultiModalProcessor`` accounts for this by bypassing
the multimodal processing cache for multi-image requests (see its
``_cached_apply_hf_processor``), so the two paths stay consistent.
DeepSeek-OCR does *not* have this restriction because its ``max_crops=6`` is
small enough to be safe for multi-image use.
"""
def tokenize_with_images(
self,
conversation: str,
images: list[Image.Image],
bos: bool = True,
eos: bool = True,
cropping: bool = True,
):
if len(images) > 1 and cropping:
logger.warning_once(
"Unlimited-OCR: crop mode is not supported for multi-image "
"input. Falling back to cropping=False."
)
cropping = False
return super().tokenize_with_images(
conversation, images, bos=bos, eos=eos, cropping=cropping
)