fix(utils): disable bench_net in gputest.py (#421)

internlm/utils/gputest.py

@@ -39,7 +39,9 @@ def empty_cache_and_diag(batch_count, interval=50):
             with torch.no_grad():
                 timer_diagnosis()
                 bench_gpu()
-                bench_net()
+                # FIXME: Runtime benchmark diagnosis can easily cause the training process
+                # to exit due to NCCL errors.
+                # bench_net()
         # do empty_cache after the bench
         torch.cuda.empty_cache()
         # do garbage collection