Method __init__

inference/app.py:27–46

(self, models_root_path, device_map="auto", torch_dtype="bfloat16")

class PromptEnhancerV2:
    def __init__(self, models_root_path, device_map="auto", torch_dtype="bfloat16"):
        if not logging.getLogger(__name__).handlers:
            logging.basicConfig(level=logging.INFO)
        self.logger = logging.getLogger(__name__)

        # dtype 兼容处理
        if torch_dtype == "bfloat16":
            dtype = torch.bfloat16
        elif torch_dtype == "float16":
            dtype = torch.float16
        else:
            dtype = torch.float32

        self.model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
            models_root_path,
            torch_dtype=dtype,
            attn_implementation="flash_attention_2",
            device_map=device_map,
        )
        self.processor = AutoProcessor.from_pretrained(models_root_path)

    @torch.inference_mode()
    def predict(