Evaluating VLMs on text-only datasets.

[sbdzdz]

Hi, I'd like to evaluate a few checkpoints of InternVL2.5 and InternVL3 from HuggingFace (e.g. this one) on text-only benchmarks like WinoGrande. I noticed a model config was added recently in #2068 that uses TurboMindModelwithChatTemplate, but I assume this is limited to checkpoints supported in LMDeploy? Is there a way to use the HuggingFace class with a VLM or do I need to write a custom Model class like shown here?