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

[Fix] handle PaddleOCR-VL image processor max_pixels across Transformers v4/v5 (#38629)

Browse Source 
Signed-off-by: zhangyue66 <zhangyue66@baidu.com>
This commit is contained in:
zhang-prog
2026-03-31 23:50:41 +08:00
committed by GitHub
parent f1ff50c86c
commit b6e636c12c
1 changed files with 7 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
vllm/model_executor/models/paddleocr_vl.py
View File

@@ -200,7 +200,13 @@ class PaddleOCRVLProcessingInfo(BaseProcessingInfo):
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 = image_processor.max_pixels // (factor**2)
if self.ctx.model_config.trust_remote_code:
# Defined in HF Hub repo
max_pixels = image_processor.max_pixels
else:
# Defined in Transformers library (requires v5.0 or above)
max_pixels = image_processor.size.longest_edge
max_num_tokens = 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))