from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
from app.config import MODEL_NAME, DEVICE

class DeepSeekModel:
    def __init__(self):
        self.tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, token=os.getenv("HF_TOKEN"))
        self.model = AutoModelForCausalLM.from_pretrained(
            MODEL_NAME, torch_dtype=torch.float16, device_map="auto"
        )

    def generate(self, prompt, max_length=256):
        inputs = self.tokenizer(prompt, return_tensors="pt").to(DEVICE)
        with torch.no_grad():
            outputs = self.model.generate(**inputs, max_length=max_length)
        return self.tokenizer.decode(outputs[0], skip_special_tokens=True)

model = DeepSeekModel()