# auto-load.py
# This script forces transformers to use the real Qwen2VLConfig
from transformers import AutoConfig, Qwen2VLForConditionalGeneration, AutoTokenizer, AutoProcessor
def load_model(repo_path):
    # Use trust_remote_code so the hub config class is rehydrated
    config = AutoConfig.from_pretrained(repo_path, trust_remote_code=True)
    model  = Qwen2VLForConditionalGeneration.from_pretrained(repo_path, config=config, torch_dtype=torch.float16)
    tokenizer = AutoTokenizer.from_pretrained(repo_path, trust_remote_code=True)
    processor = AutoProcessor.from_pretrained(repo_path, trust_remote_code=True)
    return model, tokenizer, processor