Function _init_llm

eval/run_predict_qwen2_5VL.py:67–75

(model_name)

_tokenizer = None

def _init_llm(model_name):
    global _llm,_tokenizer
    if _llm is None:
        _llm = Qwen2_5_VLForConditionalGeneration.from_pretrained(model_name, trust_remote_code=True,
                                                                  torch_dtype=torch.bfloat16,
                                                                  attn_implementation="flash_attention_2",
                                                                  )
    if _tokenizer is None:
        _tokenizer = AutoProcessor.from_pretrained(model_name, trust_remote_code=True)

def move_to(device):
    global _llm,_tokenizer