Update app.py

app.py

@@ -10,8 +10,10 @@ adapter_path = "jalonso24/lora-lateblight-v3"
 base_model = AutoModelForCausalLM.from_pretrained(
     base_model_id,
     torch_dtype=torch.float32,
-    device_map="auto"
+    device_map="auto",
+    offload_folder="./offload"  # 👈 tells it where to put overflow weights
 )
+
 model = PeftModel.from_pretrained(base_model, adapter_path)
 
 tokenizer = AutoTokenizer.from_pretrained(base_model_id, use_fast=False)