""" Lazy loader for FrozenLLaVAExtractor. Reference-cache inference does not need LLaVA weights at runtime. The extractor is imported only when on-the-fly feature extraction is requested. """ from __future__ import annotations import inspect from typing import Any def create_frozen_llava_extractor(**kwargs: Any): """Construct FrozenLLaVAExtractor; imports LLaVA stack on first use.""" from llava.model.frozen_llava_extractor import FrozenLLaVAExtractor sig = inspect.signature(FrozenLLaVAExtractor.__init__) if "merge_lora" not in sig.parameters: kwargs.pop("merge_lora", None) return FrozenLLaVAExtractor(**kwargs)