PaddleOCR-VL serving crashes with shape mismatch

[byeongjaeSon]

Bug Description

When serving a PaddleOCR-VL model using the FastDeploy framework, the worker process crashes with a shape mismatch error during inference. The error occurs inside fastdeploy/model_executor/models/paddleocr_vl/paddleocr_vl.py when assigning image_features to input_embeddings using image_mask.
Is there a known issue with image_mask / image token length handling in paddleocr_vl.py?

Environment

paddlepaddle-gpu 3.2.1
fastdeploy-gpu 2.3.3

Error Message

Non compatible API. Please refer to https://www.paddlepaddle.org.cn/documentation/docs/en/develop/guides/model_convert/convert_from_pytorch/api_difference/torch/torch.split.html first.
warnings.warn(

Traceback (most recent call last):
  File ".../fastdeploy/worker/worker_process.py", line 868, in <module>
    run_worker_proc()
  ...
  File ".../fastdeploy/model_executor/models/paddleocr_vl/paddleocr_vl.py", line 251, in get_input_embeddings
    input_embeddings[image_mask] = image_features.cast(self._dtype)
ValueError: (InvalidArgument) shape mismatch: value tensor of shape [5532, 1024] cannot be broadcast to indexing result of shape [1, 1024].
[Hint: Expected phi::funcs::CheckIsDimsMatchBool(...) == true, but received false]

[TBD1]

You can check the shape of image_features==[5532, 1024]. If that’s correct, please then check the shape of image_mask—there might be an error in how it’s being computed or generated.