Update app.py

app.py

@@ -27,21 +27,16 @@ def load_model():
         trust_remote_code=True
     )
 
-    base_model = AutoModelForCausalLM.from_pretrained(
+    model = AutoModelForCausalLM.from_pretrained(
         MODEL_ID,
-        torch_dtype=torch.float32,
-        device_map=None
+        dtype=torch.float32,
+        device_map="cpu"   # explicit
     )
 
-    # Load LoRA if present
-    try:
-        model = PeftModel.from_pretrained(base_model, MODEL_ID)
-    except Exception:
-        model = base_model
+    # DO NOT wrap with PeftModel again
+    if hasattr(model, "peft_config"):
+        print("LoRA detected and loaded once ✅")
 
-    print(model.peft_config)
-
-    model.to("cpu")
     model.eval()
     return tokenizer, model