Update app.py

app.py

@@ -101,12 +101,13 @@ def run_mmlu_evaluation(all_subjects, num_subjects, num_shots, all_questions, nu
 
     # Run evaluation with timing
     start_time = time.time()  # Start timing
-    results = evaluate_mmlu(
+    results = evaluate_mmlu_batched(
         model, 
         tokenizer,
         num_subjects=num_subjects,
         num_questions=num_questions,
         num_shots=num_shots, 
+        batch_size=32
         auto_batch_size=True
     )
     elapsed_time = time.time() - start_time  # Calculate elapsed time