Update app.py

app.py

@@ -10,11 +10,17 @@ tokenizer = AutoTokenizer.from_pretrained(base_model)
 # Load base model in CPU-only mode
 model = AutoModelForCausalLM.from_pretrained(
     base_model,
-    device_map="cpu"
+    device_map="auto",  # Use 'auto' or manually move later
+    torch_dtype=torch.float32,
+    low_cpu_mem_usage=True  # Ensures meta device usage
 )
 
-# Load the LoRA adapter
-model = PeftModel.from_pretrained(model, "lora_adapter")
+# Move the model to CPU safely
+model = model.to_empty(device=torch.device("cpu"))
+
+# Now load the LoRA adapter
+from peft import PeftModel
+model = PeftModel.from_pretrained(model, "lora_adapter", device_map="cpu")
 model.eval()
 
 def format_prompt(instruction):