diff --git a/tests/test_tipc/static/auto_parallel/baichuan2/N4C32/meta-llama-Llama-2-13b_pretrain_dy2st_bs32_bf16_DP1_MP1_PP4_1F1B_Sharding4_Stage1.sh b/tests/test_tipc/static/auto_parallel/baichuan2/N4C32/meta-llama-Llama-2-13b_pretrain_dy2st_bs32_bf16_DP1_MP1_PP4_1F1B_Sharding4_Stage1.sh
new file mode 100644
index 000000000000..4adc579b79de
--- /dev/null
+++ b/tests/test_tipc/static/auto_parallel/baichuan2/N4C32/meta-llama-Llama-2-13b_pretrain_dy2st_bs32_bf16_DP1_MP1_PP4_1F1B_Sharding4_Stage1.sh
@@ -0,0 +1,25 @@
+# Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
+# 
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+# 
+#     http://www.apache.org/licenses/LICENSE-2.0
+# 
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+param="model_item=baichuan-inc-baichaun-2-13b_pretrain "
+param+="run_mode=DP1_MP2_PP4_1F1B_Sharding8_Stage2 "
+param+="device_num=N4C32 "
+param+="global_batch_size=32 "
+param+="nnodes=4 "
+param+="model_type=baichuan2_13b "
+
+cd ./tests
+bash ./test_tipc/static/auto_parallel/baichuan2/benchmark_common/prepare.sh
+
+bash -c "${param} bash ./test_tipc/static/auto_parallel/baichuan2/benchmark_common/run_benchmark.sh"
diff --git a/tests/test_tipc/static/auto_parallel/baichuan2/benchmark_common/prepare.sh b/tests/test_tipc/static/auto_parallel/baichuan2/benchmark_common/prepare.sh
new file mode 100644
index 000000000000..90f4ce0363eb
--- /dev/null
+++ b/tests/test_tipc/static/auto_parallel/baichuan2/benchmark_common/prepare.sh
@@ -0,0 +1,36 @@
+# Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
+# 
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+# 
+#     http://www.apache.org/licenses/LICENSE-2.0
+# 
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+python -m pip install -r ../requirements.txt
+python -m pip install -r ../requirements-dev.txt
+
+# install fused_ln custom ops
+cd ../slm/model_zoo/gpt-3/external_ops/
+python setup.py install
+cd -
+
+# install fast_dataindex
+cd ../llm/auto_parallel/llama
+python -m pip install fast_dataindex
+
+# download data
+wget https://bj.bcebos.com/paddlenlp/models/transformers/llama/data/llama_openwebtext_100k_ids.npy
+wget https://bj.bcebos.com/paddlenlp/models/transformers/llama/data/llama_openwebtext_100k_idx.npz
+mkdir data
+mv llama_openwebtext_100k_ids.npy ./data
+mv llama_openwebtext_100k_idx.npz ./data
+
+# mv pretrain_config
+rm -rf pretrain_config_*
+cp -r ../../../tests/test_tipc/static/auto_parallel/baichuan2/pretrain_config_* ./
diff --git a/tests/test_tipc/static/auto_parallel/baichuan2/benchmark_common/run_benchmark.sh b/tests/test_tipc/static/auto_parallel/baichuan2/benchmark_common/run_benchmark.sh
new file mode 100644
index 000000000000..e550b7256f7f
--- /dev/null
+++ b/tests/test_tipc/static/auto_parallel/baichuan2/benchmark_common/run_benchmark.sh
@@ -0,0 +1,249 @@
+#!/usr/bin/env bash
+
+# Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
+# 
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+# 
+#     http://www.apache.org/licenses/LICENSE-2.0
+# 
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Test training benchmark for a model.
+# Usage：bash benchmark/run_benchmark.sh ${model_name_or_path} ${per_device_train_batch_size} ${tensor_parallel_degree} ${pipeline_parallel_degree} ${virtual_pp_degree} ${sequence_parallel} ${sharding_parallel_degree} ${sharding} ${recompute} ${run_mode} ${device_num}
+function _set_params(){
+    model_item=${model_item:-"baichuan-inc-baichaun-2-13b_pretrain"}
+    run_mode=${run_mode:-"MP4-PP2"}
+    device_num=${device_num:-"N4C32"}
+    global_batch_size=${global_batch_size:-64}
+    fp_item="bf16"
+    MODEL_TYPE=${model_type:-"baichuan2_13b"}
+
+    ip_lists=($(echo $TRAINER_INSTANCES | tr ',' ' '))
+    master_ip=${ip_lists[0]}
+    nnodes=${nnodes:-1}
+
+    base_batch_size=${global_batch_size}
+    profiling=${PROFILING:-"false"}      # (必选) Profiling  开关，默认关闭，通过全局变量传递
+    model_repo="PaddleNLP"          # (必选) 模型套件的名字
+    speed_unit="tokens/s"         # (必选)速度指标单位
+    skip_steps=10                  # (必选)解析日志，跳过模型前几个性能不稳定的step
+    keyword="interval_tokens_per_second_per_device:"                 # (必选)解析日志，筛选出性能数据所在行的关键字
+    convergence_key="loss:"        # (可选)解析日志，筛选出收敛数据所在行的关键字 如：convergence_key="loss:"
+    model_mode=5                   # 获取ips数据及单位，仅跳过skip_steps后计算均值，单位保持token/s不变
+    
+    # 以下为通用执行命令，无特殊可不用修改
+    model_name=${model_item}_bs${global_batch_size}_${fp_item}_${run_mode}  # (必填) 且格式不要改动,与竞品名称对齐
+    device=${CUDA_VISIBLE_DEVICES//,/ }
+    arr=(${device})
+    num_gpu_devices=${#arr[*]}
+    run_log_path=${TRAIN_LOG_DIR:-$(pwd)}  # （必填） TRAIN_LOG_DIR  benchmark框架设置该参数为全局变量
+    profiling_log_path=${PROFILING_LOG_DIR:-$(pwd)}  # （必填） PROFILING_LOG_DIR benchmark框架设置该参数为全局变量
+    speed_log_path=${LOG_PATH_INDEX_DIR:-$(pwd)}
+    train_log_file=${run_log_path}/${model_repo}_${model_name}_${device_num}_log
+    mkdir -p $(dirname ${train_log_file})
+
+    profiling_log_file=${profiling_log_path}/${model_repo}_${model_name}_${device_num}_profiling
+    mkdir -p $(dirname ${profiling_log_file})
+
+    speed_log_file=${speed_log_path}/${model_repo}_${model_name}_${device_num}_speed
+    mkdir -p $(dirname ${speed_log_file})
+
+    OUTPUT_PATH=${run_log_path}/output
+}
+
+# 循环监控文件写入状态和进程状态
+monitor_log_file() {
+    local log_file="$1"  # 获取日志文件路径
+    local training_pid="$2"  # 获取训练进程的 PID
+    local no_update_duration=0  # 初始化无更新时长计数
+    local last_size=0
+    local kill_flag_file="/tmp/monitor_killed_$training_pid"
+
+    echo "$(date '+%Y-%m-%d %H:%M:%S') 开始监控进程 $training_pid 和日志文件 $log_file..."
+
+    while true; do
+        sleep 5  # 每隔 5 秒检查一次日志文件
+
+        # 判断日志文件是否存在
+        if [ ! -f "$log_file" ]; then
+            echo "日志文件 $log_file 不存在，检查进程状态..."
+            # 如果日志文件不存在，直接判断进程是否结束
+            if ! ps -p $training_pid > /dev/null; then
+                echo "$(date '+%Y-%m-%d %H:%M:%S') 进程 $training_pid 已经结束。"
+                break
+            fi
+            continue  # 如果文件不存在，跳过后续逻辑，继续循环
+        fi
+
+        # 获取当前日志文件的大小
+        new_size=$(stat -c %s "$log_file")
+
+        if [ "$last_size" -eq "$new_size" ]; then
+            # 文件大小未变化，增加无更新时长计数
+            no_update_duration=$((no_update_duration + 5))
+            echo "$(date '+%Y-%m-%d %H:%M:%S') 文件未写入..."
+            if [ "$no_update_duration" -ge 180 ]; then
+                echo "$(date '+%Y-%m-%d %H:%M:%S') 文件在过去的 3 分钟内没有继续写入，准备杀掉进程 $training_pid."
+                # 创建标志文件
+                touch "$kill_flag_file"
+                ls -l "$kill_flag_file"
+                kill -9 $training_pid  # 杀掉进程
+                echo "$(date '+%Y-%m-%d %H:%M:%S') 进程 $training_pid 已经被杀掉。"
+                break
+            fi
+        else
+            # 文件大小有变化，重置无更新时长计数
+            echo "$(date '+%Y-%m-%d %H:%M:%S') 文件仍在写入..."
+            no_update_duration=0
+            last_size=$new_size
+        fi
+
+        # 如果训练进程已经结束，退出监控
+        if ! ps -p $training_pid > /dev/null; then
+            echo "$(date '+%Y-%m-%d %H:%M:%S') 进程 $training_pid 已经结束。"
+            break
+        fi
+    done
+}
+
+function _train(){
+    batch_size=${per_device_train_batch_size}  # 如果模型跑多卡单进程时,请在_train函数中计算出多卡需要的bs
+
+    if [ -d $OUTPUT_PATH ]; then
+        rm -rf $OUTPUT_PATH
+    fi
+    mkdir $OUTPUT_PATH
+
+    echo "current CUDA_VISIBLE_DEVICES=${CUDA_VISIBLE_DEVICES}, model_name=${model_name}, device_num=${device_num}, is profiling=${profiling}"
+
+    if [ ${profiling} == "true" ];then
+        add_options="--profiler_options=\"batch_range=[10,20];state=GPU;tracer_option=Default;profile_path=model.profile\""
+        log_file=${profiling_log_file}
+    else
+        add_options=""
+        log_file=${train_log_file}
+    fi
+    
+    # Disable for hanging bug
+    # if [ "${tensor_parallel_degree}" != "1" ]; then
+    #     export CUDA_DEVICE_MAX_CONNECTIONS=1
+    # fi
+
+    # if [ ${run_mode} == "autotuner" ]; then
+    #     unset PADDLE_ELASTIC_JOB_ID
+    #     unset PADDLE_TRAINER_ENDPOINTS
+    #     unset DISTRIBUTED_TRAINER_ENDPOINTS
+    #     unset FLAGS_START_PORT
+    #     unset PADDLE_ELASTIC_TIMEOUT
+    #     unset PADDLE_TRAINERS_NUM
+    #     unset PADDLE_TRAINER_ID
+    #     autoconfig_args="--auto_tuner_json ./auto_config_${MODEL_TYPE}/${MODEL_TYPE}_pretrain_autoconfig.json"
+    # else
+    #     autoconfig_args=""
+    # fi
+    
+    if [ ${PADDLE_TRAINER_ID} ]; then
+        PADDLE_RANK_OPTION=" --rank ${PADDLE_TRAINER_ID}"
+    else
+        PADDLE_RANK_OPTION=""
+    fi
+
+    # if [ "$autoconfig_args" != "" ]; then
+    #     distributed_args="--master etcd://$master_ip:2379 --nnodes $nnodes:$nnodes"
+    # else
+    #     distributed_args="--master $master_ip:36677 --nnodes $nnodes ${PADDLE_RANK_OPTION} --run_mode=collective"
+    # fi
+
+    echo "==========System Env============="
+    env
+    echo "================================="
+
+    # 以下为通用执行命令，无特殊可不用修改
+    case ${device_num} in
+    N1C8) echo "Run with: device_num=${device_num}, run_mode=${run_mode}"
+        train_cmd="python -u -m paddle.distributed.launch --gpus=0,1,2,3,4,5,6,7 \
+            --nnodes 1 --nproc_per_node 8 \
+            --log_dir mylog run_pretrain_auto.py \
+            ./pretrain_config_${MODEL_TYPE}/pretrain-${MODEL_TYPE}.json"
+        ;;
+    N4C32) echo "Run with: device_num=${device_num} run_mode=${run_mode}"
+        train_cmd="python -u -m paddle.distributed.launch --gpus=0,1,2,3,4,5,6,7 \
+            --log_dir mylog run_pretrain_auto.py \
+            ./pretrain_config_${MODEL_TYPE}/pretrain-${MODEL_TYPE}.json"
+        ;;
+    *) echo "Run with: device_num=${device_num}, run_mode=${run_mode}"
+        train_cmd="python -u -m paddle.distributed.launch --gpus=0,1,2,3,4,5,6,7 \
+            --log_dir mylog run_pretrain_auto.py \
+            ./pretrain_config_${MODEL_TYPE}/pretrain-${MODEL_TYPE}.json"
+        ;;
+    esac
+    cd ../llm/auto_parallel/llama
+    # rm -rf ./auto_config_${MODEL_TYPE}/*GBS*
+    # rm -rf ./auto_config_${MODEL_TYPE}/*auto_tuner.log
+    # rm -rf ./auto_config_${MODEL_TYPE}/*csv
+    # rm -rf ./auto_config_${MODEL_TYPE}/best_*
+    rm -rf mylog && rm -rf checkpoints
+    
+    echo "train_cmd: ${train_cmd}  log_file: ${log_file}"
+    timeout 40m ${train_cmd} > ${log_file} 2>&1 &
+    training_pid=$!  # 获取后台进程的 PID
+
+    # 监控进程和日志的更新状态
+    monitor_log_file "$log_file" "$training_pid" & 
+    monitor_log_file_pid=$!  # 获取日志监控进程的 PID
+
+    # 等待训练进程完成
+    wait $training_pid
+    exit_code=$?
+
+    # 获取训练进程的退出码
+    echo "训练进程 $training_pid 的退出码是 $exit_code"
+
+    # 清理后台日志监控进程
+    kill $monitor_log_file_pid
+
+
+    if [ ${exit_code} -ne 0 ];then
+        echo -e "${model_name}, FAIL"
+        # 如果程序是主动报错退出，不是monitor_log_file函数kill掉的情况下，需要等待其它机器被kill
+        # 标志文件位置
+        kill_flag_file="/tmp/monitor_killed_$training_pid"
+        if [ -f "$kill_flag_file" ]; then
+            echo "$(date '+%Y-%m-%d %H:%M:%S') 训练进程 $training_pid 是被 monitor_log_file 函数杀掉的。"
+            rm -f "$kill_flag_file"  # 清理标志文件
+        else
+            echo "$(date '+%Y-%m-%d %H:%M:%S') 训练进程 $training_pid 是主动报错退出的。"
+            sleep 120
+        fi
+    else
+        echo -e "${model_name}, SUCCESS"
+    fi
+
+    #kill -9 `ps -ef|grep 'python'|awk '{print $2}'`
+    if [ ${device_num} != "N1C1" ]; then
+        case_path=$PWD && cd - && mkdir -p mylog      # PaddleNLP/tests/mylog
+        cp -r ${case_path}/mylog/workerlog.* ./mylog/
+    fi
+}
+
+export FLAGS_selected_gpus="0,1,2,3,4,5,6,7"
+export NCCL_IB_DISABLE=0
+export PYTHONPATH=$(dirname "$PWD"):$PYTHONPATH
+# https://github.com/PaddlePaddle/Paddle/pull/69410 合入影响
+# 如不设置参数为1,则默认选择不带tensor fusion的sharding stage1版本
+export FLAGS_enable_sharding_stage1_tensor_fusion=1
+
+# 只有13b的任务需要打开CUDA_DEVICE_MAX_CONNECTIONS,7b与13b关闭
+export CUDA_DEVICE_MAX_CONNECTIONS=1
+export PARALLEL_CROSS_ENTROPY=true
+
+source ${BENCHMARK_ROOT}/scripts/run_model.sh   # 在该脚本中会对符合benchmark规范的log使用analysis.py 脚本进行性能数据解析;如果不联调只想要产出训练log可以注掉本行,提交时需打开
+_set_params $@
+#_train       # 如果只产出训练log,不解析,可取消注释
+_run     # 该函数在run_model.sh中,执行时会调用_train; 如果不联调只产出训练log可以注掉本行,提交时需打开
diff --git a/tests/test_tipc/static/auto_parallel/baichuan2/pretrain_config_baichuan2_13b/pretrain-baichuan2_13b.json b/tests/test_tipc/static/auto_parallel/baichuan2/pretrain_config_baichuan2_13b/pretrain-baichuan2_13b.json
new file mode 100644
index 000000000000..5ef40643865b
--- /dev/null
+++ b/tests/test_tipc/static/auto_parallel/baichuan2/pretrain_config_baichuan2_13b/pretrain-baichuan2_13b.json
@@ -0,0 +1,55 @@
+{
+    "model_name_or_path": "baichuan-inc/Baichuan2-13B-Base",
+    "tokenizer_name_or_path": "baichuan-inc/Baichuan2-13B-Base",
+    "input_dir": "./data",
+    "output_dir": "./checkpoints/baichuan2_13b_ckpts",
+    "split": "949,50,1",
+    "to_static": true,
+    "pipeline_parallel_degree": 2,
+    "tensor_parallel_degree": 4,
+    "virtual_pp_degree": 2,
+    "pipeline_schedule_mode": "1F1B",
+    "weight_decay": 0.01,
+    "warmup_ratio": 0.01,
+    "max_grad_norm": 0.0,
+    "learning_rate": 0.00003,
+    "min_learning_rate": 0.000003,
+    "max_steps": 100,
+    "logging_steps": 1,
+    "eval_steps": 10000,
+    "save_steps": 1000,
+    "continue_training": 0,
+    "do_train": true,
+    "do_eval": false,
+    "do_predict": false,
+    "disable_tqdm": true,
+    "save_total_limit": 2,
+    "device": "gpu",
+    "dataloader_num_workers": 4,
+    "distributed_dataloader": 0,
+    "enable_auto_parallel": 1,
+    "per_device_train_batch_size": 1,
+    "gradient_accumulation_steps": 32,
+    "per_device_eval_batch_size": 1,
+    "recompute": false,
+    "recompute_use_reentrant": true,
+    "recompute_granularity": "full",
+    "pp_recompute_interval": 0,
+    "bf16": true,
+    "fp16_opt_level": "O2",
+    "amp_master_grad": true,
+    "fuse_attention_ffn": true,
+    "fuse_attention_qkv": true,
+    "use_flash_attention": true,
+    "fused_linear": 1,
+    "fused_linear_param_grad_add": 1,
+    "use_fused_rope": true,
+    "use_fused_rms_norm": false,
+    "max_seq_length": 4096,
+    "sequence_parallel": false,
+    "sharding": "stage1",
+    "sharding_parallel_config": "enable_stage1_tensor_fusion enable_stage1_overlap",
+    "tensor_parallel_config": "enable_mp_async_allreduce",
+    "data_parallel_config": "enable_allreduce_avg_in_gradinent_scale gradient_sync_after_accumulate",
+    "pipeline_parallel_config": "enable_send_recv_overlap enable_split_backward"
+}
\ No newline at end of file
diff --git a/tests/test_tipc/static/auto_parallel/gpt3/N4C32/meta-llama-Llama-2-13b_pretrain_dy2st_bs32_bf16_DP1_MP1_PP4_1F1B_Sharding4_Stage1.sh b/tests/test_tipc/static/auto_parallel/gpt3/N4C32/meta-llama-Llama-2-13b_pretrain_dy2st_bs32_bf16_DP1_MP1_PP4_1F1B_Sharding4_Stage1.sh
new file mode 100644
index 000000000000..3fe89367fb88
--- /dev/null
+++ b/tests/test_tipc/static/auto_parallel/gpt3/N4C32/meta-llama-Llama-2-13b_pretrain_dy2st_bs32_bf16_DP1_MP1_PP4_1F1B_Sharding4_Stage1.sh
@@ -0,0 +1,25 @@
+# Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
+# 
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+# 
+#     http://www.apache.org/licenses/LICENSE-2.0
+# 
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+param="model_item=gpt3-13b_pretrain_dy2st "
+param+="run_mode=DP1_MP2_PP4_1F1B_Sharding4_Stage1 "
+param+="device_num=N4C32 "
+param+="global_batch_size=32 "
+param+="nnodes=4 "
+param+="model_type=gpt3_13b "
+
+cd ./tests
+bash ./test_tipc/static/auto_parallel/gpt3/benchmark_common/prepare.sh
+
+bash -c "${param} bash ./test_tipc/static/auto_parallel/gpt3/benchmark_common/run_benchmark.sh"
diff --git a/tests/test_tipc/static/auto_parallel/gpt3/benchmark_common/prepare.sh b/tests/test_tipc/static/auto_parallel/gpt3/benchmark_common/prepare.sh
new file mode 100644
index 000000000000..a7be57632792
--- /dev/null
+++ b/tests/test_tipc/static/auto_parallel/gpt3/benchmark_common/prepare.sh
@@ -0,0 +1,37 @@
+# Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
+# 
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+# 
+#     http://www.apache.org/licenses/LICENSE-2.0
+# 
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+python -m pip install -r ../requirements.txt
+python -m pip install -r ../requirements-dev.txt
+
+# install fused_ln custom ops
+cd ../slm/model_zoo/gpt-3/external_ops/
+python setup.py install
+cd -
+
+# install fast_dataindex
+cd ../llm/auto_parallel/gpt-3
+python -m pip install fast_dataindex
+
+# download data
+wget https://bj.bcebos.com/paddlenlp/models/transformers/gpt/data/gpt_en_dataset_300m_ids.npy
+wget https://bj.bcebos.com/paddlenlp/models/transformers/gpt/data/gpt_en_dataset_300m_idx.npz
+
+mkdir data
+mv gpt_en_dataset_300m_ids.npy ./data
+mv gpt_en_dataset_300m_idx.npz ./data
+
+# mv pretrain_config
+rm -rf pretrain_config_*
+cp -r ../../../tests/test_tipc/static/auto_parallel/gpt3/pretrain_config_* ./
diff --git a/tests/test_tipc/static/auto_parallel/gpt3/benchmark_common/run_benchmark.sh b/tests/test_tipc/static/auto_parallel/gpt3/benchmark_common/run_benchmark.sh
new file mode 100644
index 000000000000..f11a624ad854
--- /dev/null
+++ b/tests/test_tipc/static/auto_parallel/gpt3/benchmark_common/run_benchmark.sh
@@ -0,0 +1,253 @@
+#!/usr/bin/env bash
+
+# Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
+# 
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+# 
+#     http://www.apache.org/licenses/LICENSE-2.0
+# 
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Test training benchmark for a model.
+# Usage：bash benchmark/run_benchmark.sh ${model_name_or_path} ${per_device_train_batch_size} ${tensor_parallel_degree} ${pipeline_parallel_degree} ${virtual_pp_degree} ${sequence_parallel} ${sharding_parallel_degree} ${sharding} ${recompute} ${run_mode} ${device_num}
+function _set_params(){
+    model_item=${model_item:-"gpt3-13b_pretrain"}
+    run_mode=${run_mode:-"MP2-PP4"}
+    device_num=${device_num:-"N4C32"}
+    global_batch_size=${global_batch_size:-64}
+    fp_item="bf16"
+    MODEL_TYPE=${model_type:-"gpt3_13b"}
+
+    ip_lists=($(echo $TRAINER_INSTANCES | tr ',' ' '))
+    master_ip=${ip_lists[0]}
+    nnodes=${nnodes:-1}
+
+    base_batch_size=${global_batch_size}
+    profiling=${PROFILING:-"false"}      # (必选) Profiling  开关，默认关闭，通过全局变量传递
+    model_repo="PaddleNLP"          # (必选) 模型套件的名字
+    speed_unit="tokens/s"         # (必选)速度指标单位
+    skip_steps=10                  # (必选)解析日志，跳过模型前几个性能不稳定的step
+    keyword="interval_tokens_per_second_per_device:"                 # (必选)解析日志，筛选出性能数据所在行的关键字
+    convergence_key="loss:"        # (可选)解析日志，筛选出收敛数据所在行的关键字 如：convergence_key="loss:"
+    model_mode=5                   # 获取ips数据及单位，仅跳过skip_steps后计算均值，单位保持token/s不变
+    
+    # 以下为通用执行命令，无特殊可不用修改
+    model_name=${model_item}_bs${global_batch_size}_${fp_item}_${run_mode}  # (必填) 且格式不要改动,与竞品名称对齐
+    device=${CUDA_VISIBLE_DEVICES//,/ }
+    arr=(${device})
+    num_gpu_devices=${#arr[*]}
+    run_log_path=${TRAIN_LOG_DIR:-$(pwd)}  # （必填） TRAIN_LOG_DIR  benchmark框架设置该参数为全局变量
+    profiling_log_path=${PROFILING_LOG_DIR:-$(pwd)}  # （必填） PROFILING_LOG_DIR benchmark框架设置该参数为全局变量
+    speed_log_path=${LOG_PATH_INDEX_DIR:-$(pwd)}
+    train_log_file=${run_log_path}/${model_repo}_${model_name}_${device_num}_log
+    mkdir -p $(dirname ${train_log_file})
+
+    profiling_log_file=${profiling_log_path}/${model_repo}_${model_name}_${device_num}_profiling
+    mkdir -p $(dirname ${profiling_log_file})
+
+    speed_log_file=${speed_log_path}/${model_repo}_${model_name}_${device_num}_speed
+    mkdir -p $(dirname ${speed_log_file})
+
+    OUTPUT_PATH=${run_log_path}/output
+}
+
+# 循环监控文件写入状态和进程状态
+monitor_log_file() {
+    local log_file="$1"  # 获取日志文件路径
+    local training_pid="$2"  # 获取训练进程的 PID
+    local no_update_duration=0  # 初始化无更新时长计数
+    local last_size=0
+    local kill_flag_file="/tmp/monitor_killed_$training_pid"
+
+    echo "$(date '+%Y-%m-%d %H:%M:%S') 开始监控进程 $training_pid 和日志文件 $log_file..."
+
+    while true; do
+        sleep 5  # 每隔 5 秒检查一次日志文件
+
+        # 判断日志文件是否存在
+        if [ ! -f "$log_file" ]; then
+            echo "日志文件 $log_file 不存在，检查进程状态..."
+            # 如果日志文件不存在，直接判断进程是否结束
+            if ! ps -p $training_pid > /dev/null; then
+                echo "$(date '+%Y-%m-%d %H:%M:%S') 进程 $training_pid 已经结束。"
+                break
+            fi
+            continue  # 如果文件不存在，跳过后续逻辑，继续循环
+        fi
+
+        # 获取当前日志文件的大小
+        new_size=$(stat -c %s "$log_file")
+
+        if [ "$last_size" -eq "$new_size" ]; then
+            # 文件大小未变化，增加无更新时长计数
+            no_update_duration=$((no_update_duration + 5))
+            echo "$(date '+%Y-%m-%d %H:%M:%S') 文件未写入..."
+            if [ "$no_update_duration" -ge 180 ]; then
+                echo "$(date '+%Y-%m-%d %H:%M:%S') 文件在过去的 3 分钟内没有继续写入，准备杀掉进程 $training_pid."
+                # 创建标志文件
+                touch "$kill_flag_file"
+                ls -l "$kill_flag_file"
+                kill -9 $training_pid  # 杀掉进程
+                echo "$(date '+%Y-%m-%d %H:%M:%S') 进程 $training_pid 已经被杀掉。"
+                break
+            fi
+        else
+            # 文件大小有变化，重置无更新时长计数
+            echo "$(date '+%Y-%m-%d %H:%M:%S') 文件仍在写入..."
+            no_update_duration=0
+            last_size=$new_size
+        fi
+
+        # 如果训练进程已经结束，退出监控
+        if ! ps -p $training_pid > /dev/null; then
+            echo "$(date '+%Y-%m-%d %H:%M:%S') 进程 $training_pid 已经结束。"
+            break
+        fi
+    done
+}
+
+function _train(){
+    batch_size=${per_device_train_batch_size}  # 如果模型跑多卡单进程时,请在_train函数中计算出多卡需要的bs
+
+    if [ -d $OUTPUT_PATH ]; then
+        rm -rf $OUTPUT_PATH
+    fi
+    mkdir $OUTPUT_PATH
+
+    echo "current CUDA_VISIBLE_DEVICES=${CUDA_VISIBLE_DEVICES}, model_name=${model_name}, device_num=${device_num}, is profiling=${profiling}"
+
+    if [ ${profiling} == "true" ];then
+        add_options="--profiler_options=\"batch_range=[10,20];state=GPU;tracer_option=Default;profile_path=model.profile\""
+        log_file=${profiling_log_file}
+    else
+        add_options=""
+        log_file=${train_log_file}
+    fi
+    
+    # 70b和7b需要关闭这个开关
+    if [[ "${MODEL_TYPE}" =~ "70b" || "${MODEL_TYPE}" =~ "7b" ]]; then
+        unset CUDA_DEVICE_MAX_CONNECTIONS
+    fi
+    # Disable for hanging bug
+    # if [ "${tensor_parallel_degree}" != "1" ]; then
+    #     export CUDA_DEVICE_MAX_CONNECTIONS=1
+    # fi
+
+    # if [ ${run_mode} == "autotuner" ]; then
+    #     unset PADDLE_ELASTIC_JOB_ID
+    #     unset PADDLE_TRAINER_ENDPOINTS
+    #     unset DISTRIBUTED_TRAINER_ENDPOINTS
+    #     unset FLAGS_START_PORT
+    #     unset PADDLE_ELASTIC_TIMEOUT
+    #     unset PADDLE_TRAINERS_NUM
+    #     unset PADDLE_TRAINER_ID
+    #     autoconfig_args="--auto_tuner_json ./auto_config_${MODEL_TYPE}/${MODEL_TYPE}_pretrain_autoconfig.json"
+    # else
+    #     autoconfig_args=""
+    # fi
+    
+    if [ ${PADDLE_TRAINER_ID} ]; then
+        PADDLE_RANK_OPTION=" --rank ${PADDLE_TRAINER_ID}"
+    else
+        PADDLE_RANK_OPTION=""
+    fi
+
+    # if [ "$autoconfig_args" != "" ]; then
+    #     distributed_args="--master etcd://$master_ip:2379 --nnodes $nnodes:$nnodes"
+    # else
+    #     distributed_args="--master $master_ip:36677 --nnodes $nnodes ${PADDLE_RANK_OPTION} --run_mode=collective"
+    # fi
+
+    echo "==========System Env============="
+    env
+    echo "================================="
+
+    # 以下为通用执行命令，无特殊可不用修改
+    case ${device_num} in
+    N1C8) echo "Run with: device_num=${device_num}, run_mode=${run_mode}"
+        train_cmd="python -u -m paddle.distributed.launch --gpus=0,1,2,3,4,5,6,7 \
+            --nnodes 1 --nproc_per_node 8 \
+            --log_dir mylog run_pretrain_auto.py \
+            ./pretrain_config_${MODEL_TYPE}/pretrain-${MODEL_TYPE}.json"
+        ;;
+    N4C32) echo "Run with: device_num=${device_num} run_mode=${run_mode}"
+        train_cmd="python -u -m paddle.distributed.launch --gpus=0,1,2,3,4,5,6,7 \
+            --log_dir mylog run_pretrain_auto.py \
+            ./pretrain_config_${MODEL_TYPE}/pretrain-${MODEL_TYPE}.json"
+        ;;
+    *) echo "Run with: device_num=${device_num}, run_mode=${run_mode}"
+        train_cmd="python -u -m paddle.distributed.launch --gpus=0,1,2,3,4,5,6,7 \
+            --log_dir mylog run_pretrain_auto.py \
+            ./pretrain_config_${MODEL_TYPE}/pretrain-${MODEL_TYPE}.json"
+        ;;
+    esac
+    cd ../llm/auto_parallel/gpt-3
+    # rm -rf ./auto_config_${MODEL_TYPE}/*GBS*
+    # rm -rf ./auto_config_${MODEL_TYPE}/*auto_tuner.log
+    # rm -rf ./auto_config_${MODEL_TYPE}/*csv
+    # rm -rf ./auto_config_${MODEL_TYPE}/best_*
+    rm -rf mylog && rm -rf checkpoints
+    
+    echo "train_cmd: ${train_cmd}  log_file: ${log_file}"
+    timeout 40m ${train_cmd} > ${log_file} 2>&1 &
+    training_pid=$!  # 获取后台进程的 PID
+
+    # 监控进程和日志的更新状态
+    monitor_log_file "$log_file" "$training_pid" & 
+    monitor_log_file_pid=$!  # 获取日志监控进程的 PID
+
+    # 等待训练进程完成
+    wait $training_pid
+    exit_code=$?
+
+    # 获取训练进程的退出码
+    echo "训练进程 $training_pid 的退出码是 $exit_code"
+
+    # 清理后台日志监控进程
+    kill $monitor_log_file_pid
+
+
+    if [ ${exit_code} -ne 0 ];then
+        echo -e "${model_name}, FAIL"
+        # 如果程序是主动报错退出，不是monitor_log_file函数kill掉的情况下，需要等待其它机器被kill
+        # 标志文件位置
+        kill_flag_file="/tmp/monitor_killed_$training_pid"
+        if [ -f "$kill_flag_file" ]; then
+            echo "$(date '+%Y-%m-%d %H:%M:%S') 训练进程 $training_pid 是被 monitor_log_file 函数杀掉的。"
+            rm -f "$kill_flag_file"  # 清理标志文件
+        else
+            echo "$(date '+%Y-%m-%d %H:%M:%S') 训练进程 $training_pid 是主动报错退出的。"
+            sleep 120
+        fi
+    else
+        echo -e "${model_name}, SUCCESS"
+    fi
+
+    #kill -9 `ps -ef|grep 'python'|awk '{print $2}'`
+    if [ ${device_num} != "N1C1" ]; then
+        case_path=$PWD && cd - && mkdir -p mylog      # PaddleNLP/tests/mylog
+        cp -r ${case_path}/mylog/workerlog.* ./mylog/
+    fi
+}
+
+export FLAGS_selected_gpus="0,1,2,3,4,5,6,7"
+export NCCL_IB_DISABLE=0
+export PYTHONPATH=$(dirname "$PWD"):$PYTHONPATH
+# https://github.com/PaddlePaddle/Paddle/pull/69410 合入影响
+# 如不设置参数为1,则默认选择不带tensor fusion的sharding stage1版本
+export FLAGS_enable_sharding_stage1_tensor_fusion=1
+
+# 只有13b的任务需要打开CUDA_DEVICE_MAX_CONNECTIONS,7b与13b关闭
+export CUDA_DEVICE_MAX_CONNECTIONS=1
+export PARALLEL_CROSS_ENTROPY=true
+
+source ${BENCHMARK_ROOT}/scripts/run_model.sh   # 在该脚本中会对符合benchmark规范的log使用analysis.py 脚本进行性能数据解析;如果不联调只想要产出训练log可以注掉本行,提交时需打开
+_set_params $@
+#_train       # 如果只产出训练log,不解析,可取消注释
+_run     # 该函数在run_model.sh中,执行时会调用_train; 如果不联调只产出训练log可以注掉本行,提交时需打开
diff --git a/tests/test_tipc/static/auto_parallel/gpt3/pretrain_config_gpt3_13b/pretrain-gpt3_13b.json b/tests/test_tipc/static/auto_parallel/gpt3/pretrain_config_gpt3_13b/pretrain-gpt3_13b.json
new file mode 100644
index 000000000000..129f55a349e2
--- /dev/null
+++ b/tests/test_tipc/static/auto_parallel/gpt3/pretrain_config_gpt3_13b/pretrain-gpt3_13b.json
@@ -0,0 +1,58 @@
+{
+    "model_name_or_path": "gpt3-13B-en",
+    "tokenizer_name_or_path": "gpt3-13B-en",
+    "to_static": true,
+    "enable_auto_parallel": 1,
+    "input_dir": "../llama_data",
+    "output_dir": "./checkpoints/gpt_pretrain_ckpts",
+    "split": "949,50,1",
+    "max_seq_length": 4096,
+    "per_device_train_batch_size": 1,
+    "per_device_eval_batch_size": 1,
+    "scale_loss": 1024,
+    "learning_rate": 0.00001,
+    "min_learning_rate": 0.000001,
+    "max_steps": 100,
+    "save_steps": 50000,
+    "weight_decay": 0.01,
+    "warmup_ratio": 0.01,
+    "logging_steps": 1,
+    "continue_training": 0,
+    "dataloader_num_workers": 4,
+    "eval_steps": 100000,
+    "report_to": "visualdl",
+    "disable_tqdm": true,
+    "do_train": true,
+    "do_eval": true,
+    "device": "gpu",
+    "model_type": "gpt",
+    "sharding": "stage1",
+    "tensor_parallel_degree": 2,
+    "pipeline_parallel_degree": 4,
+    "virtual_pp_degree": 2,
+    "pipeline_schedule_mode": "1F1B",
+    "virtual_pipeline_seg_method": "GPTDecoderLayerAuto",
+    "sequence_parallel": 0,
+    "use_flash_attention": 1,
+    "fused_linear": 1,
+    "fuse_attention_ffn": 1,
+    "fuse_attention_qkv": 1,
+    "fused_linear_param_grad_add": 1,
+    "use_fused_rope": true,
+    "use_fused_rms_norm": false,
+    "recompute": 0,
+    "recompute_use_reentrant": true,
+    "recompute_granularity": "full",
+    "pp_recompute_interval": 1,
+    "gradient_accumulation_steps": 32,
+    "max_grad_norm": 0.1,
+    "bf16": 1,
+    "fp16_opt_level": "O2",
+    "amp_master_grad": true,
+    "attention_probs_dropout_prob": 0.1,
+    "hidden_dropout_prob": 0.1,
+    "sharding_parallel_config": "enable_stage1_tensor_fusion enable_stage1_overlap",
+    "tensor_parallel_config": "enable_mp_async_allreduce",
+    "data_parallel_config": "enable_allreduce_avg_in_gradinent_scale gradient_sync_after_accumulate",
+    "pipeline_parallel_config": "enable_send_recv_overlap enable_split_backward"
+}
\ No newline at end of file
diff --git a/tests/test_tipc/static/auto_parallel/qwen/N4C32/qwen-14b_pretrain_dy2st_bs32_bf16_DP1_MP2_PP4_1F1B_Sharding4_Stage1.sh b/tests/test_tipc/static/auto_parallel/qwen/N4C32/qwen-14b_pretrain_dy2st_bs32_bf16_DP1_MP2_PP4_1F1B_Sharding4_Stage1.sh
new file mode 100644
index 000000000000..be88c3dcddb9
--- /dev/null
+++ b/tests/test_tipc/static/auto_parallel/qwen/N4C32/qwen-14b_pretrain_dy2st_bs32_bf16_DP1_MP2_PP4_1F1B_Sharding4_Stage1.sh
@@ -0,0 +1,25 @@
+# Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
+# 
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+# 
+#     http://www.apache.org/licenses/LICENSE-2.0
+# 
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+param="model_item=qwen-2-14b_pretrain_dy2st "
+param+="run_mode=DP1_MP2_PP4_1F1B_Sharding4_Stage1 "
+param+="device_num=N4C32 "
+param+="global_batch_size=32 "
+param+="nnodes=4 "
+param+="model_type=qwen_14b "
+
+cd ./tests
+bash ./test_tipc/static/auto_parallel/qwen/benchmark_common/prepare.sh
+
+bash -c "${param} bash ./test_tipc/static/auto_parallel/qwen/benchmark_common/run_benchmark.sh"
diff --git a/tests/test_tipc/static/auto_parallel/qwen/benchmark_common/prepare.sh b/tests/test_tipc/static/auto_parallel/qwen/benchmark_common/prepare.sh
new file mode 100644
index 000000000000..7239135a96c5
--- /dev/null
+++ b/tests/test_tipc/static/auto_parallel/qwen/benchmark_common/prepare.sh
@@ -0,0 +1,36 @@
+# Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
+# 
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+# 
+#     http://www.apache.org/licenses/LICENSE-2.0
+# 
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+python -m pip install -r ../requirements.txt
+python -m pip install -r ../requirements-dev.txt
+
+# install fused_ln custom ops
+cd ../slm/model_zoo/gpt-3/external_ops/
+python setup.py install
+cd -
+
+# install fast_dataindex
+cd ../llm/auto_parallel/qwen
+python -m pip install fast_dataindex
+
+# download data
+wget https://bj.bcebos.com/paddlenlp/models/transformers/llama/data/llama_openwebtext_100k_ids.npy
+wget https://bj.bcebos.com/paddlenlp/models/transformers/llama/data/llama_openwebtext_100k_idx.npz
+mkdir data
+mv llama_openwebtext_100k_ids.npy ./data
+mv llama_openwebtext_100k_idx.npz ./data
+
+# mv pretrain_config
+rm -rf pretrain_config_*
+cp -r ../../../tests/test_tipc/static/auto_parallel/qwen/pretrain_config_* ./
diff --git a/tests/test_tipc/static/auto_parallel/qwen/benchmark_common/run_benchmark.sh b/tests/test_tipc/static/auto_parallel/qwen/benchmark_common/run_benchmark.sh
new file mode 100644
index 000000000000..41471ef5b21f
--- /dev/null
+++ b/tests/test_tipc/static/auto_parallel/qwen/benchmark_common/run_benchmark.sh
@@ -0,0 +1,253 @@
+#!/usr/bin/env bash
+
+# Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
+# 
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+# 
+#     http://www.apache.org/licenses/LICENSE-2.0
+# 
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Test training benchmark for a model.
+# Usage：bash benchmark/run_benchmark.sh ${model_name_or_path} ${per_device_train_batch_size} ${tensor_parallel_degree} ${pipeline_parallel_degree} ${virtual_pp_degree} ${sequence_parallel} ${sharding_parallel_degree} ${sharding} ${recompute} ${run_mode} ${device_num}
+function _set_params(){
+    model_item=${model_item:-"qwen-14b_pretrain"}
+    run_mode=${run_mode:-"MP2-PP1"}
+    device_num=${device_num:-"N1C8"}
+    global_batch_size=${global_batch_size:-64}
+    fp_item="bf16"
+    MODEL_TYPE=${model_type:-"qwen_14b"}
+
+    ip_lists=($(echo $TRAINER_INSTANCES | tr ',' ' '))
+    master_ip=${ip_lists[0]}
+    nnodes=${nnodes:-1}
+
+    base_batch_size=${global_batch_size}
+    profiling=${PROFILING:-"false"}      # (必选) Profiling  开关，默认关闭，通过全局变量传递
+    model_repo="PaddleNLP"          # (必选) 模型套件的名字
+    speed_unit="tokens/s"         # (必选)速度指标单位
+    skip_steps=10                  # (必选)解析日志，跳过模型前几个性能不稳定的step
+    keyword="interval_tokens_per_second_per_device:"                 # (必选)解析日志，筛选出性能数据所在行的关键字
+    convergence_key="loss:"        # (可选)解析日志，筛选出收敛数据所在行的关键字 如：convergence_key="loss:"
+    model_mode=5                   # 获取ips数据及单位，仅跳过skip_steps后计算均值，单位保持token/s不变
+    
+    # 以下为通用执行命令，无特殊可不用修改
+    model_name=${model_item}_bs${global_batch_size}_${fp_item}_${run_mode}  # (必填) 且格式不要改动,与竞品名称对齐
+    device=${CUDA_VISIBLE_DEVICES//,/ }
+    arr=(${device})
+    num_gpu_devices=${#arr[*]}
+    run_log_path=${TRAIN_LOG_DIR:-$(pwd)}  # （必填） TRAIN_LOG_DIR  benchmark框架设置该参数为全局变量
+    profiling_log_path=${PROFILING_LOG_DIR:-$(pwd)}  # （必填） PROFILING_LOG_DIR benchmark框架设置该参数为全局变量
+    speed_log_path=${LOG_PATH_INDEX_DIR:-$(pwd)}
+    train_log_file=${run_log_path}/${model_repo}_${model_name}_${device_num}_log
+    mkdir -p $(dirname ${train_log_file})
+
+    profiling_log_file=${profiling_log_path}/${model_repo}_${model_name}_${device_num}_profiling
+    mkdir -p $(dirname ${profiling_log_file})
+
+    speed_log_file=${speed_log_path}/${model_repo}_${model_name}_${device_num}_speed
+    mkdir -p $(dirname ${speed_log_file})
+
+    OUTPUT_PATH=${run_log_path}/output
+}
+
+# 循环监控文件写入状态和进程状态
+monitor_log_file() {
+    local log_file="$1"  # 获取日志文件路径
+    local training_pid="$2"  # 获取训练进程的 PID
+    local no_update_duration=0  # 初始化无更新时长计数
+    local last_size=0
+    local kill_flag_file="/tmp/monitor_killed_$training_pid"
+
+    echo "$(date '+%Y-%m-%d %H:%M:%S') 开始监控进程 $training_pid 和日志文件 $log_file..."
+
+    while true; do
+        sleep 5  # 每隔 5 秒检查一次日志文件
+
+        # 判断日志文件是否存在
+        if [ ! -f "$log_file" ]; then
+            echo "日志文件 $log_file 不存在，检查进程状态..."
+            # 如果日志文件不存在，直接判断进程是否结束
+            if ! ps -p $training_pid > /dev/null; then
+                echo "$(date '+%Y-%m-%d %H:%M:%S') 进程 $training_pid 已经结束。"
+                break
+            fi
+            continue  # 如果文件不存在，跳过后续逻辑，继续循环
+        fi
+
+        # 获取当前日志文件的大小
+        new_size=$(stat -c %s "$log_file")
+
+        if [ "$last_size" -eq "$new_size" ]; then
+            # 文件大小未变化，增加无更新时长计数
+            no_update_duration=$((no_update_duration + 5))
+            echo "$(date '+%Y-%m-%d %H:%M:%S') 文件未写入..."
+            if [ "$no_update_duration" -ge 180 ]; then
+                echo "$(date '+%Y-%m-%d %H:%M:%S') 文件在过去的 3 分钟内没有继续写入，准备杀掉进程 $training_pid."
+                # 创建标志文件
+                touch "$kill_flag_file"
+                ls -l "$kill_flag_file"
+                kill -9 $training_pid  # 杀掉进程
+                echo "$(date '+%Y-%m-%d %H:%M:%S') 进程 $training_pid 已经被杀掉。"
+                break
+            fi
+        else
+            # 文件大小有变化，重置无更新时长计数
+            echo "$(date '+%Y-%m-%d %H:%M:%S') 文件仍在写入..."
+            no_update_duration=0
+            last_size=$new_size
+        fi
+
+        # 如果训练进程已经结束，退出监控
+        if ! ps -p $training_pid > /dev/null; then
+            echo "$(date '+%Y-%m-%d %H:%M:%S') 进程 $training_pid 已经结束。"
+            break
+        fi
+    done
+}
+
+function _train(){
+    batch_size=${per_device_train_batch_size}  # 如果模型跑多卡单进程时,请在_train函数中计算出多卡需要的bs
+
+    if [ -d $OUTPUT_PATH ]; then
+        rm -rf $OUTPUT_PATH
+    fi
+    mkdir $OUTPUT_PATH
+
+    echo "current CUDA_VISIBLE_DEVICES=${CUDA_VISIBLE_DEVICES}, model_name=${model_name}, device_num=${device_num}, is profiling=${profiling}"
+
+    if [ ${profiling} == "true" ];then
+        add_options="--profiler_options=\"batch_range=[10,20];state=GPU;tracer_option=Default;profile_path=model.profile\""
+        log_file=${profiling_log_file}
+    else
+        add_options=""
+        log_file=${train_log_file}
+    fi
+    
+    # 70b和7b需要关闭这个开关
+    if [[ "${MODEL_TYPE}" =~ "70b" || "${MODEL_TYPE}" =~ "7b" ]]; then
+        unset CUDA_DEVICE_MAX_CONNECTIONS
+    fi
+    # Disable for hanging bug
+    # if [ "${tensor_parallel_degree}" != "1" ]; then
+    #     export CUDA_DEVICE_MAX_CONNECTIONS=1
+    # fi
+
+    # if [ ${run_mode} == "autotuner" ]; then
+    #     unset PADDLE_ELASTIC_JOB_ID
+    #     unset PADDLE_TRAINER_ENDPOINTS
+    #     unset DISTRIBUTED_TRAINER_ENDPOINTS
+    #     unset FLAGS_START_PORT
+    #     unset PADDLE_ELASTIC_TIMEOUT
+    #     unset PADDLE_TRAINERS_NUM
+    #     unset PADDLE_TRAINER_ID
+    #     autoconfig_args="--auto_tuner_json ./auto_config_${MODEL_TYPE}/${MODEL_TYPE}_pretrain_autoconfig.json"
+    # else
+    #     autoconfig_args=""
+    # fi
+    
+    if [ ${PADDLE_TRAINER_ID} ]; then
+        PADDLE_RANK_OPTION=" --rank ${PADDLE_TRAINER_ID}"
+    else
+        PADDLE_RANK_OPTION=""
+    fi
+
+    # if [ "$autoconfig_args" != "" ]; then
+    #     distributed_args="--master etcd://$master_ip:2379 --nnodes $nnodes:$nnodes"
+    # else
+    #     distributed_args="--master $master_ip:36677 --nnodes $nnodes ${PADDLE_RANK_OPTION} --run_mode=collective"
+    # fi
+
+    echo "==========System Env============="
+    env
+    echo "================================="
+
+    # 以下为通用执行命令，无特殊可不用修改
+    case ${device_num} in
+    N1C8) echo "Run with: device_num=${device_num}, run_mode=${run_mode}"
+        train_cmd="python -u -m paddle.distributed.launch --gpus=0,1,2,3,4,5,6,7 \
+            --nnodes 1 --nproc_per_node 8 \
+            --log_dir mylog run_pretrain_auto.py \
+            ./pretrain_config_${MODEL_TYPE}/pretrain-${MODEL_TYPE}.json"
+        ;;
+    N4C32) echo "Run with: device_num=${device_num} run_mode=${run_mode}"
+        train_cmd="python -u -m paddle.distributed.launch --gpus=0,1,2,3,4,5,6,7 \
+            --log_dir mylog run_pretrain_auto.py \
+            ./pretrain_config_${MODEL_TYPE}/pretrain-${MODEL_TYPE}.json"
+        ;;
+    *) echo "Run with: device_num=${device_num}, run_mode=${run_mode}"
+        train_cmd="python -u -m paddle.distributed.launch --gpus=0,1,2,3,4,5,6,7 \
+            --log_dir mylog run_pretrain_auto.py \
+            ./pretrain_config_${MODEL_TYPE}/pretrain-${MODEL_TYPE}.json"
+        ;;
+    esac
+    cd ../llm/auto_parallel/qwen
+    # rm -rf ./auto_config_${MODEL_TYPE}/*GBS*
+    # rm -rf ./auto_config_${MODEL_TYPE}/*auto_tuner.log
+    # rm -rf ./auto_config_${MODEL_TYPE}/*csv
+    # rm -rf ./auto_config_${MODEL_TYPE}/best_*
+    rm -rf mylog && rm -rf checkpoints
+    
+    echo "train_cmd: ${train_cmd}  log_file: ${log_file}"
+    timeout 40m ${train_cmd} > ${log_file} 2>&1 &
+    training_pid=$!  # 获取后台进程的 PID
+
+    # 监控进程和日志的更新状态
+    monitor_log_file "$log_file" "$training_pid" & 
+    monitor_log_file_pid=$!  # 获取日志监控进程的 PID
+
+    # 等待训练进程完成
+    wait $training_pid
+    exit_code=$?
+
+    # 获取训练进程的退出码
+    echo "训练进程 $training_pid 的退出码是 $exit_code"
+
+    # 清理后台日志监控进程
+    kill $monitor_log_file_pid
+
+
+    if [ ${exit_code} -ne 0 ];then
+        echo -e "${model_name}, FAIL"
+        # 如果程序是主动报错退出，不是monitor_log_file函数kill掉的情况下，需要等待其它机器被kill
+        # 标志文件位置
+        kill_flag_file="/tmp/monitor_killed_$training_pid"
+        if [ -f "$kill_flag_file" ]; then
+            echo "$(date '+%Y-%m-%d %H:%M:%S') 训练进程 $training_pid 是被 monitor_log_file 函数杀掉的。"
+            rm -f "$kill_flag_file"  # 清理标志文件
+        else
+            echo "$(date '+%Y-%m-%d %H:%M:%S') 训练进程 $training_pid 是主动报错退出的。"
+            sleep 120
+        fi
+    else
+        echo -e "${model_name}, SUCCESS"
+    fi
+
+    #kill -9 `ps -ef|grep 'python'|awk '{print $2}'`
+    if [ ${device_num} != "N1C1" ]; then
+        case_path=$PWD && cd - && mkdir -p mylog      # PaddleNLP/tests/mylog
+        cp -r ${case_path}/mylog/workerlog.* ./mylog/
+    fi
+}
+
+export FLAGS_selected_gpus="0,1,2,3,4,5,6,7"
+export NCCL_IB_DISABLE=0
+export PYTHONPATH=$(dirname "$PWD"):$PYTHONPATH
+# https://github.com/PaddlePaddle/Paddle/pull/69410 合入影响
+# 如不设置参数为1,则默认选择不带tensor fusion的sharding stage1版本
+export FLAGS_enable_sharding_stage1_tensor_fusion=1
+
+# 只有13b的任务需要打开CUDA_DEVICE_MAX_CONNECTIONS,7b与13b关闭
+export CUDA_DEVICE_MAX_CONNECTIONS=1
+export PARALLEL_CROSS_ENTROPY=true
+
+source ${BENCHMARK_ROOT}/scripts/run_model.sh   # 在该脚本中会对符合benchmark规范的log使用analysis.py 脚本进行性能数据解析;如果不联调只想要产出训练log可以注掉本行,提交时需打开
+_set_params $@
+#_train       # 如果只产出训练log,不解析,可取消注释
+_run     # 该函数在run_model.sh中,执行时会调用_train; 如果不联调只产出训练log可以注掉本行,提交时需打开
diff --git a/tests/test_tipc/static/auto_parallel/qwen/pretrain_config_qwen2_14b/pretrain-qwen2_14b.json b/tests/test_tipc/static/auto_parallel/qwen/pretrain_config_qwen2_14b/pretrain-qwen2_14b.json
new file mode 100644
index 000000000000..9e8bacb911af
--- /dev/null
+++ b/tests/test_tipc/static/auto_parallel/qwen/pretrain_config_qwen2_14b/pretrain-qwen2_14b.json
@@ -0,0 +1,55 @@
+{
+    "model_name_or_path": "qwen/qwen-14b",
+    "tokenizer_name_or_path": "qwen/qwen-14b",
+    "input_dir": "./data",
+    "output_dir": "./checkpoints/qwen_pretrain_ckpts",
+    "per_device_train_batch_size": 1,
+    "gradient_accumulation_steps": 32,
+    "per_device_eval_batch_size": 16,
+    "sharding": "stage1",
+    "tensor_parallel_degree": 2,
+    "pipeline_parallel_degree": 4,
+    "virtual_pp_degree": 5,
+    "pipeline_schedule_mode": "1F1B",
+    "virtual_pipeline_seg_method": "QWenBlockAuto",
+    "use_flash_attention": true,
+    "use_fused_rms_norm": false,
+    "use_fused_rope": true,
+    "fused_linear": 1,
+    "fuse_attention_ffn": 1,
+    "fuse_attention_qkv": 1,
+    "fused_linear_param_grad_add": 1,
+    "max_seq_length": 4096,
+    "learning_rate": 0.00003,
+    "min_learning_rate": 0.000003,
+    "scale_loss": 1024,
+    "warmup_steps": 30,
+    "logging_steps": 1,
+    "max_steps": 100,
+    "save_steps": 1000,
+    "eval_steps": 10000,
+    "weight_decay": 0.01,
+    "bf16": true,
+    "fp16_opt_level": "O2",
+    "amp_master_grad": true,
+    "warmup_ratio": 0.01,
+    "max_grad_norm": 0.0,
+    "dataloader_num_workers": 4,
+    "continue_training": 0,
+    "do_train": true,
+    "do_eval": false,
+    "do_predict": false,
+    "disable_tqdm": true,
+    "recompute": false,
+    "recompute_granularity": "core_attn",
+    "recompute_use_reentrant": true,
+    "distributed_dataloader": 0,
+    "save_total_limit": 2,
+    "enable_auto_parallel": 1,
+    "to_static": 1,
+    "auto_parallel_resume_form_hybrid_parallel": true,
+    "data_parallel_config": "enable_allreduce_avg_in_gradinent_scale gradient_sync_after_accumulate",
+    "sharding_parallel_config": "enable_stage1_overlap",
+    "tensor_parallel_config": "enable_mp_async_allreduce",
+    "pipeline_parallel_config": "enable_send_recv_overlap enable_split_backward"
+}
\ No newline at end of file