fix use after free

app.py

@@ -102,12 +102,15 @@ def completion(prompt: str, model_id: str):
         temperature=None,
     )
 
+    # decode response
+    resp = tokenizer.decode(outputs[0][input_ids.shape[-1]:], skip_special_tokens=True)
+
     # cleanup memory
     del model, tokenizer
     torch.cuda.empty_cache()
     gc.collect()
 
-    return tokenizer.decode(outputs[0][input_ids.shape[-1] :], skip_special_tokens=True)
+    return resp
 
 
 def completion_openrouter(prompt: str, model_id: str):