xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-15 07:14:59 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Bugfix] Fix max image size for PaddleOCR-VL (#28442)

Browse Source 
Signed-off-by: Roger Wang <hey@rogerw.io>
This commit is contained in:
Roger Wang 2025-11-11 00:07:24 -08:00 committed by GitHub
parent cc079763c5
commit 4fd4b743a2
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 21 additions and 15 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

20
vllm/model_executor/models/paddleocr_vl.py
View File

@ -198,13 +198,10 @@ class PaddleOCRVLProcessingInfo(BaseProcessingInfo):
if image_processor is None: if image_processor is None:
image_processor = self.get_image_processor() image_processor = self.get_image_processor()
do_resize = True
hf_config = self.get_hf_config() hf_config = self.get_hf_config()
vision_config = hf_config.vision_config vision_config = hf_config.vision_config
patch_size = vision_config.patch_size patch_size = vision_config.patch_size
merge_size = vision_config.spatial_merge_size merge_size = vision_config.spatial_merge_size
if do_resize:
resized_height, resized_width = smart_resize( resized_height, resized_width = smart_resize(
height=image_height, height=image_height,
width=image_width, width=image_width,
@ -213,8 +210,6 @@ class PaddleOCRVLProcessingInfo(BaseProcessingInfo):
max_pixels=image_processor.max_pixels, max_pixels=image_processor.max_pixels,
) )
preprocessed_size = ImageSize(width=resized_width, height=resized_height) preprocessed_size = ImageSize(width=resized_width, height=resized_height)
else:
preprocessed_size = ImageSize(width=image_width, height=image_height)
grid_t = 1 grid_t = 1
grid_h = preprocessed_size.height // patch_size grid_h = preprocessed_size.height // patch_size
@ -227,8 +222,19 @@ class PaddleOCRVLProcessingInfo(BaseProcessingInfo):
def get_image_size_with_most_features(self) -> ImageSize: def get_image_size_with_most_features(self) -> ImageSize:
hf_config = self.get_hf_config() hf_config = self.get_hf_config()
image_size = hf_config.vision_config.image_size
return ImageSize(height=image_size, width=image_size) # See `smart_resize` for the calculation of the image size.
merge_size = hf_config.vision_config.spatial_merge_size
patch_size = hf_config.vision_config.patch_size
factor = merge_size * patch_size
max_num_tokens = self.get_image_processor().max_pixels // (factor**2)
# Find factors of max_num_tokens close to its square root
# to create a dummy image with a reasonable aspect ratio.
h_patches = int(math.sqrt(max_num_tokens))
while max_num_tokens % h_patches != 0:
h_patches -= 1
w_patches = max_num_tokens // h_patches
return ImageSize(height=h_patches * factor, width=w_patches * factor)
class PaddleOCRVLDummyInputsBuilder(BaseDummyInputsBuilder[PaddleOCRVLProcessingInfo]): class PaddleOCRVLDummyInputsBuilder(BaseDummyInputsBuilder[PaddleOCRVLProcessingInfo]):