biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Bugfix] Fix deepseek-ocr multi-image inference and add merge_by_field_config=True with tensor schema support (#27361)

Browse Source 
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
This commit is contained in:
Isotr0py
2025-10-23 08:15:38 +08:00
committed by GitHub
parent b4fda58a2d
commit 2566dca2a9
4 changed files with 112 additions and 66 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
vllm/transformers_utils/processors/deepseek_ocr.py
View File

@@ -411,20 +411,16 @@ class DeepseekOCRProcessor(ProcessorMixin):
images_seq_mask = images_seq_mask[:-1]
if len(images_list) == 0:
pixel_values = torch.zeros((1, 3, self.base_size, self.base_size))
images_spatial_crop = torch.zeros((1, 1), dtype=torch.long)
images_crop = torch.zeros(
(1, 3, self.image_size, self.image_size)
).unsqueeze(0)
pixel_values = torch.zeros((0, 3, self.base_size, self.base_size))
images_spatial_crop = torch.zeros((0, 2), dtype=torch.long)
images_crop = torch.zeros((0, 3, self.image_size, self.image_size))
else:
pixel_values = torch.stack(images_list, dim=0)
images_spatial_crop = torch.tensor(images_spatial_crop, dtype=torch.long)
if images_crop_list:
images_crop = torch.stack(images_crop_list, dim=0).unsqueeze(0)
images_crop = torch.stack(images_crop_list, dim=0)
else:
images_crop = torch.zeros(
(1, 3, self.image_size, self.image_size)
).unsqueeze(0)
images_crop = torch.zeros((0, 3, self.image_size, self.image_size))
input_ids = input_ids.unsqueeze(0)