Loading Data Gemma from HuggingFace

model.py

class DataGemma:
   def __init__(self, model_id: str = "bartowski/datagemma-rag-27b-it-GGUF", model_file: str =  "datagemma-rag-27b-it-Q2_K.gguf"):
      self.generation_kwargs = {
         "max_tokens": 4096, # Max number of new tokens to generate
      }
      self.model_path = hf_hub_download(model_id, model_file)
      self.llm = Llama(
            self.model_path
      )
      self.name = "DataGemma"
   
   def complete(self, question: str) -> str:
      llm_resp = self.llm(question, **self.generation_kwargs)

      return llm_resp["choices"][0]["text"]