A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 5, 18:30 UTC
[Fix] handle PaddleOCR-VL image processor max_pixels across Transformers v4/v5 (#38629)
Signed-off-by: zhangyue66 <zhangyue66@baidu.com>
zhang-prog committed
Mar 31, 2026 at 23:50 UTC
b6e636c12c28c9ae83e22c65b4fca64a5571483b
1 file changed
+7
-1
vllm/model_executor/models/paddleocr_vl.py
+7
-1
@@ -200,7 +200,13 @@ class PaddleOCRVLProcessingInfo(BaseProcessingInfo):
200
merge_size = hf_config.vision_config.spatial_merge_size
201
patch_size = hf_config.vision_config.patch_size
202
factor = merge_size * patch_size
203
- max_num_tokens = image_processor.max_pixels // (factor**2)
203
+ if self.ctx.model_config.trust_remote_code:
204
+ # Defined in HF Hub repo
205
+ max_pixels = image_processor.max_pixels
206
+ else:
207
+ # Defined in Transformers library (requires v5.0 or above)
208
+ max_pixels = image_processor.size.longest_edge
209
+ max_num_tokens = max_pixels // (factor**2)
210
# Find factors of max_num_tokens close to its square root
211
# to create a dummy image with a reasonable aspect ratio.
212
h_patches = int(math.sqrt(max_num_tokens))