unset CUDA_VISIBLE_DEVICES before train

tests/test.sh

@@ -248,6 +248,7 @@ for gpu in ${gpu_list[*]}; do
                 cmd="${python} -m paddle.distributed.launch --ips=${ips} --gpus=${gpu} ${run_train} ${save_model_key}=${save_log} ${set_pretrain} ${set_epoch} ${set_autocast} ${set_batchsize} ${set_train_params1}"
             fi
             # run train
+            eval "unset CUDA_VISIBLE_DEVICES"
             eval $cmd
             status_check $? "${cmd}" "${status_log}"