Update app.py

app.py

@@ -24,8 +24,17 @@ from transformers import AutoModelForCausalLM, AutoTokenizer
 from airllm import AirLLMLlama2
 
 MAX_LENGTH = 128
+
+
+hf_hub_download(
+    repo_id="CohereForAI/c4ai-command-r-plus-4bit",
+    # filename="Meta-Llama-3-70B-Instruct-Q3_K_M.gguf",
+    local_dir = "./models",
+    token= token
+)
+
 # could use hugging face model repo id:
-model = AirLLMLlama2("CohereForAI/c4ai-command-r-plus-4bit", token= token)
+model = AirLLMLlama2("./models", )
 
 
 # tokenizer = AutoTokenizer.from_pretrained(model_id, token= token)