GitHub Repo
Transformers Confirms PaddleOCR-VL Loading Path Incompatibility, Awaiting Model-Side Configuration Update
A report on November 15 said PaddleOCR-VL-1.6 may crash when loaded with a specified Auto class because the native model and remote configuration can be mixed. The next day, a Transformers contributor confirmed this resolution path and said they were coordinating an update with the model authors.

On November 16, a Transformers project contributor confirmed that one loading path for PaddleOCR-VL-1.6 combines the native model implementation with a remote configuration, causing initialization to fail. The report, submitted the previous day, used Transformers 5.17.0, Windows 11, and PyTorch 2.12.1. Loading through AutoModelForImageTextToText raised an exception about the missing rope_parameters, even with trust_remote_code=True. Issue report and contributor response
The key is that the checkpoint’s auto_map does not list the task class above. According to the report and the project contributor’s explanation, the loader therefore selects Transformers’ built-in PaddleOCR-VL model, while the configuration still comes from remote code; the two are incompatible. This causes the error during model construction and makes it difficult for users to tell from the call which implementation is actually running. Explanation of the resolution path
This mismatch directly affects the public usage example: the model card still demonstrates AutoProcessor with AutoModelForImageTextToText and requires Transformers 5. The card also says this example supports element-level recognition and text localization, while recommending the official inference method for page-level document parsing. Therefore, this report does not establish that all PaddleOCR deployments are affected. Official model card
The reporter said that using AutoModelForCausalLM, which is listed in auto_map, and allowing remote code enabled OCR to complete successfully on the same page. This is the temporary workaround verified by the reporter. The project contributor said they had contacted the authors and were waiting to learn whether they would adopt the native integration or add the missing task-class mapping. As of this check, the discussion had not announced a completed fix. Issue status
For document recognition services, this is a reminder for engineering teams to include the model checkpoint version, package version, and Auto class in loading validation. Teams should watch for updates to the model-side configuration and check whether the documentation examples match the actual resolution path. Current evidence is not sufficient to establish the scope of impact on other versions.