Can vLLM be used with non-HuggingFace models?

[vadimcn]

On the surface, is seems that the answer is yes, however upon trying to follow this path, one quickly discovers that even before instantiating the model itself, vLLM creates a model config object and goes straight to HF AutoConfig.from_pretrained(), which supports only known HF models...