增加线上提交作业方式

README.md

@@ -61,8 +61,33 @@ python依赖安装（如已经拉取镜像可忽略下述依赖安装步骤）
     #不同环境的配置及数据的存放路径会有不同，请根据实际情况进行调整run_benchmark_8gpu.sh脚本中的如下内容：
     BASE_DATA_DIR=${BASE_DATA_DIR:-"/public/DL_DATA/mlperf/bert"} //调整为具体的数据的路径
 
+### 测试规模
+
+线上资源进行双节点八卡规模性能&&精度测试
+
+### 环境配置
+需要采用提供的环境，已经打包，可通过网盘下载，下载链接如下：
+链接：https://pan.baidu.com/s/1YI4SPDEzb4F5_oJaEuhWEw?pwd=a9xx
+提取码：a9xx
+
+### 预训练模型
+/workspace/bert_data文件夹存放预训练模型如下：
+
+    ├── /workpalce/bert_data/phase1
+    └── └──model.ckpt-28252.tf_pickled #预训练模型
+
+### 训练
+训练命令：
+
+    step1. 将获取的环境tar包解压至适宜位置，并按照位置信息&&平台信息对env.sh进行修改
+    step2. 依照平台信息对run_sbatch.sh作业提交脚本进行修改
+    step3. 执行命令 sbatch run_sbatch.sh
+    
+    #不同环境的配置及数据的存放路径会有不同，请根据实际情况进行调整run_benchmark.sh脚本中的如下内容：
+    BASE_DATA_DIR=${BASE_DATA_DIR:-"/public/DL_DATA/mlperf/bert"} //调整为具体的数据的路径
+
 ## 测试结果
-采用上述输入数据，加速卡采用Z100L*8，可最终达到官方收敛要求；
+采用上述输入数据，加速卡采用Z100L*8 && 线上Z100*8，可最终达到官方收敛要求；
 
 ## 源码仓库及问题反馈
 * https://developer.hpccube.com/codes/modelzoo/mlperf_bert-large

env.sh

+#!/bin/bash
+
+module purge
+module load compiler/devtoolset/7.3.1
+module load mpi/hpcx/2.4.1/gcc-7.3.1
+module load compiler/rocm/dtk-21.04
+
+#for mlperf_bert
+export PYTHONPATH=/path_to_env
+export PATH=/path_to_env/bin:$PATH
+export LD_LIBRARY_PATH=/path_to_env/lib/python3.6:$LD_LIBRARY_PATH
+export LD_LIBRARY_PATH=/path_to_env/lib:$LD_LIBRARY_PATH
+export INCLUDE=/path_to_env/include/python3.6m:$INCLUDE
+export ROCM_DIR=/public/software/compiler/rocm/dtk-21.04
+
+export PATH=${ROCM_DIR}/bin:${ROCM_DIR}/hcc/bin:${ROCM_DIR}/hip/bin:${ROCM_DIR}/miopen/bin:${ROCM_DIR}/opencl/bin/x86_64:$PATH
+export LD_LIBRARY_PATH=${ROCM_DIR}/lib:${ROCM_DIR}/lib64:${ROCM_DIR}/miopen/lib:${ROCM_DIR}/opencl/lib/x86_64:/public/software/compiler/rocm/dtk-21.04/roctracer/lib:$LD_LIBRARY_PATH
+export C_INCLUDE_PATH=${ROCM_DIR}/include:${ROCM_DIR}/opencl/include:${ROCM_DIR}/hiprand/include:${ROCM_DIR}/miopen/include:$C_INCLUDE_PATH
+export CPLUS_INCLUDE_PATH=${ROCM_DIR}/include:${ROCM_DIR}/miopen/include:$CPLUS_INCLUDE_PATH
+export ROCM_PATH=${ROCM_DIR}
+export HIP_PATH=${ROCM_DIR}/hip/
+
+
+export C_INCLUDE_PATH=/public/software/apps/DeepLearning/PyTorch/gflags-2.1.2-build/include:/public/software/apps/DeepLearning/PyTorch/glog-build/include:$C_INCLUDE_PATH
+export CPLUS_INCLUDE_PATH=/public/software/apps/DeepLearning/PyTorch/gflags-2.1.2-build/include:/public/software/apps/DeepLearning/PyTorch/glog-build/include:$CPLUS_INCLUDE_PATH
+export LD_LIBRARY_PATH=/public/software/apps/DeepLearning/PyTorch/glog-build/lib/:$LD_LIBRARY_PATH
+export LD_LIBRARY_PATH=/public/software/apps/DeepLearning/PyTorch/lmdb-0.9.24-build/lib/:$LD_LIBRARY_PATH
+export LD_LIBRARY_PATH=/public/software/apps/DeepLearning/PyTorch/opencv-2.4.13.6-build/lib/:$LD_LIBRARY_PATH
+export LD_LIBRARY_PATH=/public/software/apps/DeepLearning/PyTorch/openblas-0.3.7-build/lib/:$LD_LIBRARY_PATH
+export LD_LIBRARY_PATH=/public/software/apps/DeepLearning/PyTorch/gflags-2.1.2-build/lib/:$LD_LIBRARY_PATH
+export LD_LIBRARY_PATH=/public/software/apps/DeepLearning/PyTorch/lib/:$LD_LIBRARY_PATH
+export LD_LIBRARY_PATH=/public/software/apps/DeepLearning/PyTorch/openmp-build/lib:$LD_LIBRARY_PATH

external_ops/env.sh

-#!/bin/bash
-################################################################################
-# Copyright 2016-2021 by SW Group, Chengdu Hygon IC Design Co., Ltd.
-# All right reserved. See COPYRIGHT for detailed Information.
-#
-# @file     set_env.sh
-# @brief    set env variables for running test.
-#
-# @author   wangmingliang <wangmingliang@hygon.cn>
-# @date     2022/03/23
-# @history  1.0
-################################################################################
-export HYGON_ROCM_INSTALL=/opt/dtk-21.04
-
-# library path
-# fix LD_LIBRARY_PATH begin or end with colon
-export LD_LIBRARY_PATH=$(echo ${LD_LIBRARY_PATH} | sed 's/:$//; s/^://;')
-export LD_LIBRARY_PATH=${HYGON_ROCM_INSTALL}/lib:${HYGON_ROCM_INSTALL}/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
-export LD_LIBRARY_PATH=${HYGON_ROCM_INSTALL}/hip/lib${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
-export LD_LIBRARY_PATH=${HYGON_ROCM_INSTALL}/hipblas/lib${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
-export LD_LIBRARY_PATH=${HYGON_ROCM_INSTALL}/hipcub/lib${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
-export LD_LIBRARY_PATH=${HYGON_ROCM_INSTALL}/hipfft/lib${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
-export LD_LIBRARY_PATH=${HYGON_ROCM_INSTALL}/hiprand/lib${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
-export LD_LIBRARY_PATH=${HYGON_ROCM_INSTALL}/hipsolver/lib${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
-export LD_LIBRARY_PATH=${HYGON_ROCM_INSTALL}/hipsparse/lib${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
-export LD_LIBRARY_PATH=${HYGON_ROCM_INSTALL}/hsa/lib${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
-export LD_LIBRARY_PATH=${HYGON_ROCM_INSTALL}/llvm/lib${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
-export LD_LIBRARY_PATH=${HYGON_ROCM_INSTALL}/miopen/lib${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
-export LD_LIBRARY_PATH=${HYGON_ROCM_INSTALL}/miopengemm/lib${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
-export LD_LIBRARY_PATH=${HYGON_ROCM_INSTALL}/oam/lib${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
-export LD_LIBRARY_PATH=${HYGON_ROCM_INSTALL}/opencl/lib${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
-export LD_LIBRARY_PATH=${HYGON_ROCM_INSTALL}/rccl/lib${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
-export LD_LIBRARY_PATH=${HYGON_ROCM_INSTALL}/rocalution/lib${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
-export LD_LIBRARY_PATH=${HYGON_ROCM_INSTALL}/rocblas/lib${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
-export LD_LIBRARY_PATH=${HYGON_ROCM_INSTALL}/rocclr/lib${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
-export LD_LIBRARY_PATH=${HYGON_ROCM_INSTALL}/rocfft/lib${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
-export LD_LIBRARY_PATH=${HYGON_ROCM_INSTALL}/rocm_smi/lib${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
-export LD_LIBRARY_PATH=${HYGON_ROCM_INSTALL}/rocprim/lib${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
-export LD_LIBRARY_PATH=${HYGON_ROCM_INSTALL}/rocprofiler/lib${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
-export LD_LIBRARY_PATH=${HYGON_ROCM_INSTALL}/rocrand/lib${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
-export LD_LIBRARY_PATH=${HYGON_ROCM_INSTALL}/rocsolver/lib${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
-export LD_LIBRARY_PATH=${HYGON_ROCM_INSTALL}/rocsparse/lib${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
-export LD_LIBRARY_PATH=${HYGON_ROCM_INSTALL}/rocthrust/lib${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
-export LD_LIBRARY_PATH=${HYGON_ROCM_INSTALL}/roctracer/lib${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
-export LD_LIBRARY_PATH=${HYGON_ROCM_INSTALL}/rocblas/lib${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
-
-export C_INCLUDE_PATH=${HYGON_ROCM_INSTALL}/rocrand/include:$C_INCLUDE_PATH
-export CPLUS_INCLUDE_PATH=${HYGON_ROCM_INSTALL}/rocrand/include:$CPLUS_INCLUDE_PATH
-export C_INCLUDE_PATH=${HYGON_ROCM_INSTALL}/hiprand/include:$C_INCLUDE_PATH
-export CPLUS_INCLUDE_PATH=${HYGON_ROCM_INSTALL}/hiprand/include:$CPLUS_INCLUDE_PATH
-# executable path
-# fix PATH begin or end with colon
-export PATH=$(echo ${PATH} | sed 's/:$//; s/^://;')
-export PATH=${HYGON_ROCM_INSTALL}/bin${PATH:+:${PATH}}
-export PATH=${HYGON_ROCM_INSTALL}/hip/bin${PATH:+:${PATH}}
-export PATH=${HYGON_ROCM_INSTALL}/llvm/bin${PATH:+:${PATH}}
-export PATH=${HYGON_ROCM_INSTALL}/llvm/lib/clang/13.0.0/bin${PATH:+:${PATH}}
-export PATH=${HYGON_ROCM_INSTALL}/miopen/bin${PATH:+:${PATH}}
-export PATH=${HYGON_ROCM_INSTALL}/opencl/bin${PATH:+:${PATH}}
-export PATH=${HYGON_ROCM_INSTALL}/rocprofiler/bin${PATH:+:${PATH}}
-export PATH=/public/home/zhangqha/bladisc/hmmer/bin${PATH:+:${PATH}}
-export PATH=/public/home/zhangqha/bladisc/hh-suite-master/build/bin${PATH:+:${PATH}}
-export PATH=/public/home/zhangqha/bladisc/kalign/build/bin${PATH:+:${PATH}}
-# component path
-export ROCM_PATH=${HYGON_ROCM_INSTALL}
-export HSA_PATH=${HYGON_ROCM_INSTALL}/hsa
-export HIP_PATH=${HYGON_ROCM_INSTALL}/hip
-export HIP_ROCCLR_HOME=${HYGON_ROCM_INSTALL}/hip
-export HIP_LIB_PATH=${HYGON_ROCM_INSTALL}/hip/lib
-export DEVICE_LIB_PATH=${HYGON_ROCM_INSTALL}/amdgcn/bitcode
-export HIP_CLANG_PATH=${HYGON_ROCM_INSTALL}/llvm/bin
-export HIP_RUNTIME="rocclr"
-export HIP_COMPILER="clang"
-
-#export CPLUS_INCLUDE_PATH=/public/software/apps/DeepLearning/PyTorch_Lib/gflags-2.1.2-build/include:/public/home/zhangqha/openssl_install/include:/public/home/zhangqha/openssl_install/include/openssl:/public/software/apps/DeepLearning/PyTorch_Lib/glog-build/include:${HIP_PATH}/include:$CPLUS_INCLUDE_PATH
-#export INCLUDE=/public/software/apps/DeepLearning/PyTorch_Lib/gflags-2.1.2-build/include:/public/home/zhangqha/openssl_install/include:/public/home/zhangqha/openssl_install/include/openssl:/public/software/apps/DeepLearning/PyTorch_Lib/glog-build/include:${HIP_PATH}/include:$INCLUDE
-#export C_INCLUDE_PATH=/public/software/apps/DeepLearning/PyTorch_Lib/gflags-2.1.2-build/include:/public/home/zhangqha/openssl_install/include:/public/home/zhangqha/openssl_install/include/openssl:/public/software/apps/DeepLearning/PyTorch_Lib/glog-build/include:${HIP_PATH}/include:$C_INCLUDE_PATH

run_benchmark.sh

+#!/bin/bash
+
+# Copyright (c) 2021 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.
+
+set -ex
+export FLAGS_rocm_dir=/public/software/compiler/rocm/dtk-21.04/
+export FLAGS_max_inplace_grad_add=2
+export HSA_FORCE_FINE_GRAIN_PCIE=1
+export NCCL_P2P_LEVEL=5
+
+export USE_NV_INPUT=1
+
+USE_UNCOMPRESSED_DATASET=1
+
+BASE_DATA_DIR=${BASE_DATA_DIR:-"/public/software/apps/DeepLearning/Data/mlperf/bert"}
+export USE_NV_INPUT
+UNCOMPRESSED_DATA_DIR=$BASE_DATA_DIR/hdf5/training-4320/hdf5_4320_shards_uncompressed
+VARLENGTH_DATA_DIR=$BASE_DATA_DIR/hdf5/training-4320/hdf5_4320_shards_varlength
+
+export DATA_DIR=$UNCOMPRESSED_DATA_DIR
+export EVAL_DIR=$BASE_DATA_DIR/hdf5/eval
+if [[ "$USE_NV_INPUT" == "1" && "$USE_UNCOMPRESSED_DATASET" == "0" ]]; then
+  export DATA_DIR="$VARLENGTH_DATA_DIR"
+  export EVAL_DIR=$BASE_DATA_DIR/hdf5/eval
+else
+  export USE_UNCOMPRESSED_DATASET=1
+fi
+export USE_UNCOMPRESSED_DATASET
+export TF_CKPT_PATH=$BASE_DATA_DIR/phase1/model.ckpt-28252.tf_pickled
+export BERT_CONFIG_PATH=$BASE_DATA_DIR/phase1/bert_config.json
+
+export PYTHON=python3
+export PADDLE_TRAINER_ID=${OMPI_COMM_WORLD_RANK}
+export PADDLE_TRAINERS_NUM=${PADDLE_TRAINERS_NUM:-"1"}
+export PADDLE_TRAINER_ENDPOINTS=${PADDLE_TRAINER_ENDPOINTS:-""}
+
+OMPI_COMM_WORLD_RANK=${OMPI_COMM_WORLD_RANK:-"0"}
+lrank=$OMPI_COMM_WORLD_LOCAL_RANK
+
+function get_device_id() {
+$PYTHON <<EOF
+import paddle
+import os
+gpus = os.environ.get("CUDA_VISIBLE_DEVICES", None)
+if gpus is None:
+    print($OMPI_COMM_WORLD_RANK)
+else:
+    gpus = gpus.split(",")
+    print(gpus[$OMPI_COMM_WORLD_RANK])
+EOF
+}
+if [[ $PADDLE_TRAINER_ID -lt $PADDLE_TRAINERS_NUM ]]; then
+  export CUDA_VISIBLE_DEVICES=0,1,2,3 #$(expr $OMPI_COMM_WORLD_RANK % 4) #`get_device_id`
+  export IS_TRAINER=1
+  export IS_READER=0
+else
+  export CUDA_VISIBLE_DEVICES=""
+  export IS_TRAINER=0
+  export IS_READER=1
+fi
+
+echo "Trainer :" $CUDA_VISIBLE_DEVICES $PADDLE_TRAINER_ENDPOINTS $PADDLE_TRAINERS_NUM
+
+export FLAGS_sync_nccl_allreduce=0
+export FLAGS_fraction_of_gpu_memory_to_use=0.99
+export FLAGS_call_stack_level=2
+export FLAGS_use_fast_math=0
+export FLAGS_enable_nvtx=1
+
+
+batch_size=4
+eval_batch_size=63
+use_amp=True
+use_pure_fp16=True
+
+max_steps=7100
+log_freq=50
+eval_iter_start_samples=150000
+eval_iter_samples=150000
+max_seq_length=512
+
+dense_seq_output=True
+unpad=False
+unpad_fmha=False
+fused_bias_mha=True
+fused_bias_fc=True
+## can be False or True 
+weight_transpose=True
+
+fused_dropout_add_ln=False
+exchange_padding=True
+cpu_exchange_padding=True
+
+distributed_lamb=True
+
+unpad_embed=False
+unpad_fmha_mke_opt=True
+
+sort_eval_data=False
+
+LOG_DIR="log_${PADDLE_TRAINERS_NUM}"
+mkdir -p ${LOG_DIR}
+LOG_FILE=${LOG_DIR}/worker.${PADDLE_TRAINER_ID}
+
+export FLAGS_max_inplace_grad_add=2
+
+if [[ "$exchange_padding" == "true" || "$exchange_padding" == "True" ]]; then
+  if [[ "$cpu_exchange_padding" == "true" || "$cpu_exchange_padding" == "True" ]]; then
+    export DATA_DIR="$UNCOMPRESSED_DATA_DIR"
+  fi
+fi
+
+
+BERT_CMD="run_pretrain.py \
+   --max_predictions_per_seq 76 \
+   --train_batch_size $batch_size   \
+   --eval_batch_size $eval_batch_size \
+   --sort_eval_data $sort_eval_data \
+   --learning_rate 0.000425 \
+   --weight_decay 1e-2 \
+   --lamb_epsilon 1e-6 \
+   --start_warmup_step 0 \
+   --warmup_proportion 0.0 \
+   --warmup_steps 0 \
+   --input_dir $DATA_DIR \
+   --log_freq $log_freq \
+   --max_steps $max_steps \
+   --tf_ckpt_path $TF_CKPT_PATH \
+   --bert_config_path $BERT_CONFIG_PATH \
+   --unpad $unpad \
+   --unpad_fmha $unpad_fmha \
+   --unpad_fmha_mke_opt $unpad_fmha_mke_opt \
+   --unpad_embed $unpad_embed \
+   --fused_bias_mha $fused_bias_mha \
+   --fused_bias_fc $fused_bias_fc \
+   --fused_dropout_add_ln $fused_dropout_add_ln \
+   --weight_transpose $weight_transpose \
+   --max_seq_length $max_seq_length \
+   --eval_dir $EVAL_DIR \
+   --distributed_lamb $distributed_lamb \
+   --exchange_padding $exchange_padding \
+   --cpu_exchange_padding $cpu_exchange_padding \
+   --seed $SEED \
+   --use_uncompressed_dataset $USE_UNCOMPRESSED_DATASET \
+   --dense_seq_output $dense_seq_output \
+   --gradient_accumulation_steps 14 \
+   --opt_lamb_beta_1 0.9 \
+   --opt_lamb_beta_2 0.999 \
+   --enable_addto True \
+   --use_pure_fp16 $use_pure_fp16 \
+   --use_amp $use_amp"
+
+
+APP="python3 -u $BERT_CMD"
+case $(expr $lrank % 4) in
+[0])
+  echo "work ${lrank} less than ${PADDLE_TRAINERS_NUM} on DCU $(expr $lrank % 4)"
+  export HIP_VISIBLE_DEVICES=0,1,2,3
+  export FLAGS_selected_gpus=0
+  export UCX_NET_DEVICES=mlx5_0:1
+  export UCX_IB_PCI_BW=mlx5_0:50Gbs
+  numactl --cpunodebind=0 --membind=0 ${APP} >& $LOG_FILE
+  ;;
+[1])
+  echo "work ${lrank} less than ${PADDLE_TRAINERS_NUM} on DCU $(expr $lrank % 4)"
+  export HIP_VISIBLE_DEVICES=0,1,2,3
+  export FLAGS_selected_gpus=1
+  export UCX_NET_DEVICES=mlx5_1:1
+  export UCX_IB_PCI_BW=mlx5_1:50Gbs
+  numactl --cpunodebind=1 --membind=1 ${APP} >& $LOG_FILE
+  ;;
+[2])
+  echo "work ${lrank} less than ${PADDLE_TRAINERS_NUM} on DCU $(expr $lrank % 4)"
+  export HIP_VISIBLE_DEVICES=0,1,2,3
+  export FLAGS_selected_gpus=2
+  export UCX_NET_DEVICES=mlx5_2:1
+  export UCX_IB_PCI_BW=mlx5_2:50Gbs
+  numactl --cpunodebind=2 --membind=2 ${APP} >& $LOG_FILE
+  ;;
+[3])
+  echo "work ${lrank} less than ${PADDLE_TRAINERS_NUM} on DCU $(expr $lrank % 4)"
+  export HIP_VISIBLE_DEVICES=0,1,2,3
+  export FLAGS_selected_gpus=3
+  export UCX_NET_DEVICES=mlx5_3:1
+  export UCX_IB_PCI_BW=mlx5_3:50Gbs
+  numactl --cpunodebind=3 --membind=3 ${APP}  >& $LOG_FILE
+  ;;
+esac

run_sbatch.sh

+#!/bin/bash
+#SBATCH -p xxx 
+#SBATCH -N 2
+#SBATCH --ntasks-per-node=8
+#SBATCH --cpus-per-task=4
+#SBATCH --gres=dcu:4
+#SBATCH -J mlperf
+
+source ./env.sh
+cp ./rundir_8gpu/init_env.py .
+export HSA_FORCE_FINE_GRAIN_PCIE=1
+export PYTHON=python3
+export PADDLE_TRAINER_ENDPOINTS=`$PYTHON -c "import list;print(list.get_list())"`
+echo $PADDLE_TRAINER_ENDPOINTS
+
+#set -e
+hostfile=./hostfile
+scontrol show hostnames $SLURM_JOB_NODELIST > ${hostfile}
+num_node=$(cat $hostfile|sort|uniq |wc -l)
+num_DCU=$(($num_node*4))
+export LD_LIBRARY_PATH=/public/software/compiler/rocm/dtk-21.04/rccl/lib:$LD_LIBRARY_PATH
+export HSA_FORCE_FINE_GRAIN_PCIE=1
+export NCCL_P2P_LEVEL=5
+export use_hierarchical_allreduce=True
+export NCCL_IB_HCA=mlx5_0:1
+rm hosts
+p=0
+for i in `cat hostfile`;do
+    for j in `seq 1 4`;do
+        num=$[ $p * 4 + $j - 1]
+        echo "rank ${num}=${i} slot=$[j-1]" >> hosts
+    done
+    p=$(expr $p + 1)
+done
+for i in `cat hostfile`;do
+    for j in `seq 1 4`;do
+        num=$[ $p * 4 + $j - 1]
+        echo "rank ${num}=${i} slot=$[j+3]" >> hosts
+    done
+    p=$(expr $p + 1)
+done
+
+
+export HIP_LAUNCH_BLOCKING=1
+export PADDLE_TRAINERS_NUM=$num_DCU
+
+export SEED=${SEED:-"$RANDOM"}
+echo "PADDLE_TRAINER_ENDPOINTS " $PADDLE_TRAINER_ENDPOINTS
+
+
+num_process=$(($PADDLE_TRAINERS_NUM*2))
+if [[ $num_process -gt 1 ]]; then
+  ORTERUN=`which orterun`
+  mpirun="mpirun --allow-run-as-root -np $num_process --rankfile hosts -mca btl_tcp_if_exclude ib0 --bind-to none -x PADDLE_TRAINERS_NUM -x PADDLE_TRAINER_ENDPOINTS -x LD_LIBRARY_PATH -x SEED -x PYTHON -x NPROC_PER_NODE -x use_hierarchical_allreduce ./run_benchmark.sh"
+else
+  mpirun=""
+fi
+
+for NPROC_PER_NODE in 4; do
+  echo "command is " $mpirun $CMD
+  export NPROC_PER_NODE=$NPROC_PER_NODE
+  $mpirun $CMD
+done