XGrammar falling to Outlines

[michele-aindo]

I’ve been using XgrammarLogitsProcessor with the Hugging Face transformers library to generate JSON output, starting with Pydantic BaseModels.

However, when I try using the same Pydantic models in GuidedDecodingParams, I get a warning from has_xgrammar_unsupported_json_features (in vllm/v1/structured_output/utils.py), and it falls back to using the outlines engine. I suspect this is happening because of the enum field, but I’m not entirely sure.

Is there any difference between the features provided by Xgrammar in LogitsProcessor and the vLLM backend?

Thanks in advance!