upd

app.py

@@ -68,14 +68,16 @@ def evaluate(
 def load_lora(lora_path, base_model="decapoda-research/llama-7b-hf"):
     model = LlamaForCausalLM.from_pretrained(
         base_model,
-        # load_in_8bit=True,
+        load_in_8bit=True,
         # device_map=device_map,
         low_cpu_mem_usage=True,
+        torch_type=torch.float16,
     )
     print("Loading LoRA...")
     lora = PeftModel.from_pretrained(
         model,
         lora_path,
+        torch_type=torch.float16,
         # device_map=device_map,
     )
     return lora