Update app.py

app.py

@@ -7,6 +7,7 @@ import spaces
 import torch
 from transformers import (
     AutoModelForCausalLM,
+    BitsAndBytesConfig,
     GemmaTokenizerFast,
     TextIteratorStreamer,
 )
@@ -33,6 +34,7 @@ model = AutoModelForCausalLM.from_pretrained(
     model_id,
     device_map="auto",
     torch_dtype=torch.bfloat16,
+    quantization_config=BitsAndBytesConfig(load_in_8bit=True),
 )
 model.config.sliding_window = 4096
 model.eval()