Update app.py

app.py

@@ -94,7 +94,10 @@ db = Chroma(persist_directory=CHROMA_PATH, embedding_function=embeddings)
 MODEL_NAME = "gpt2"
 
 model = AutoModelForCausalLM.from_pretrained(
-    MODEL_NAME, trust_remote_code=True, device_map="auto",offload_folder="offload"
+"gpt2",
+device_map="auto",
+low_cpu_mem_usage=True,
+torch_dtype=torch.float16  # Use float16 to reduce memory usage
 )
 model = model.eval()