diff --git a/Classification/resnet50/0_dist_ssh_key/README.md b/Classification/resnet50/0_dist_ssh_key/README.md
new file mode 100644
index 0000000..856df74
--- /dev/null
+++ b/Classification/resnet50/0_dist_ssh_key/README.md
@@ -0,0 +1,83 @@
+# 使用 Ansible 将 SSH 公钥分发到多个目标主机
+
+## 1. 创建变量文件并加密
+
+创建一个包含密码的变量文件vars.yml：
+
+```yaml
+all:
+  hosts:
+    192.168.1.27:
+      ansible_user: myuser
+      ansible_password: mypassword
+    192.168.1.28:
+      ansible_user: myuser
+      ansible_password: mypassword
+```
+
+然后使用Ansible Vault加密这个文件：
+
+```bash
+ansible-vault encrypt vars.yml
+```
+
+注意：
+
+1. 执行 `ansible-vault` 的过程中需要设定一个密码，请记住或保存好这个密码
+2. `vars.yml`将被替换为加密后的文件
+
+## 2. 创建主机清单文件
+
+创建一个主机清单文件`inventory.ini`：
+
+```ini
+[all]
+node1 ansible_host=192.168.1.27 ansible_user=myuser
+node2 ansible_host=192.168.1.28 ansible_user=myuser
+```
+
+注：需要根据情况修改 `ansible_user` 的值
+
+## 3. 创建Playbook
+
+如果文件存在，这一步可以忽略。
+
+创建一个Playbook distribute_ssh_key.yml：
+
+```yaml
+---
+- name: Distribute SSH key
+  hosts: all
+  vars_files:
+    - vars.yml
+  tasks:
+    - name: Create .ssh directory if it doesn't exist
+      file:
+        path: /home/{{ ansible_user }}/.ssh
+        state: directory
+        mode: '0700'
+        owner: "{{ ansible_user }}"
+        group: "{{ ansible_user }}"
+
+    - name: Copy the SSH key to the authorized_keys file
+      authorized_key:
+        user: "{{ ansible_user }}"
+        state: present
+        key: "{{ lookup('file', '/path/to/id_rsa.pub') }}"
+```
+
+注：`vars_files` 配置为 `vars.yml`
+
+## 4. 运行Playbook
+
+使用以下命令运行Playbook，并解密变量文件：
+
+```bash
+ansible-playbook -i inventory.ini distribute_ssh_key.yml --ask-vault-pass
+```
+或者运行
+
+```bash
+./dist_ssh_key.sh
+```
+
diff --git a/Classification/resnet50/0_dist_ssh_key/dist_ssh_key.sh b/Classification/resnet50/0_dist_ssh_key/dist_ssh_key.sh
new file mode 100755
index 0000000..7cb0f1d
--- /dev/null
+++ b/Classification/resnet50/0_dist_ssh_key/dist_ssh_key.sh
@@ -0,0 +1 @@
+ansible-playbook -i inventory.ini distribute_ssh_key.yml --ask-vault-pass
diff --git a/Classification/resnet50/0_dist_ssh_key/distribute_ssh_key.yml b/Classification/resnet50/0_dist_ssh_key/distribute_ssh_key.yml
new file mode 100644
index 0000000..95147c2
--- /dev/null
+++ b/Classification/resnet50/0_dist_ssh_key/distribute_ssh_key.yml
@@ -0,0 +1,19 @@
+---
+- name: Distribute SSH key
+  hosts: all
+  vars_files:
+    - vars.yml
+  tasks:
+    - name: Create .ssh directory if it doesn't exist
+      file:
+        path: /home/{{ ansible_user }}/.ssh
+        state: directory
+        mode: '0700'
+        owner: "{{ ansible_user }}"
+        group: "{{ ansible_user }}"
+
+    - name: Copy the SSH key to the authorized_keys file
+      authorized_key:
+        user: "{{ ansible_user }}"
+        state: present
+        key: "{{ lookup('file', '/home/xiexuan/.ssh/id_rsa.pub') }}"
diff --git a/Classification/resnet50/0_dist_ssh_key/inventory.ini b/Classification/resnet50/0_dist_ssh_key/inventory.ini
new file mode 100644
index 0000000..894b65d
--- /dev/null
+++ b/Classification/resnet50/0_dist_ssh_key/inventory.ini
@@ -0,0 +1,3 @@
+[all]
+of27 ansible_host=192.168.1.27 ansible_user=xiexuan
+of28 ansible_host=192.168.1.28 ansible_user=xiexuan
diff --git a/Classification/resnet50/0_dist_ssh_key/vars.yml b/Classification/resnet50/0_dist_ssh_key/vars.yml
new file mode 100644
index 0000000..49c7dbf
--- /dev/null
+++ b/Classification/resnet50/0_dist_ssh_key/vars.yml
@@ -0,0 +1,8 @@
+all:
+  hosts:
+    192.168.1.27:
+      ansible_user: myuser
+      ansible_password: mypassword
+    192.168.1.28:
+      ansible_user: myuser
+      ansible_password: mypassword
diff --git a/Classification/resnet50/1_get_docker_image/README.md b/Classification/resnet50/1_get_docker_image/README.md
new file mode 100644
index 0000000..3f6282a
--- /dev/null
+++ b/Classification/resnet50/1_get_docker_image/README.md
@@ -0,0 +1,61 @@
+# 拉取或导入镜像
+
+## 拉取镜像
+
+适用于直接从 dockerhub 拉取镜像。
+
+用法: `./pull.sh [镜像标签]`
+
+参数说明:
+
+- 镜像标签 (可选)  : 要拉取的Docker镜像标签，例如 alpine:latest。如果未提供，则使用playbook中的默认值。
+
+示例:
+
+- 默认使用:
+
+```bash
+./pull.sh
+```
+
+- 指定镜像标签:
+
+ ```bash
+./pull.sh alpine:latest
+ ```
+
+## 导入镜像
+
+适用于本地共享目录有已经保存镜像的tar文件，使用 `docker load` 导入。
+
+用法: `./load.sh [镜像文件路径] [镜像标签] [强制导入]`
+
+参数说明:
+
+- 镜像文件路径 (可选)  : 要导入的Docker镜像tar文件路径，默认为 `/share_nfs/k85/oneflow.0.9.1.dev20240203-cuda11.8.tar`
+- 镜像标签 (可选)      : 导入后设置的Docker镜像标签，默认为 `oneflowinc/oneflow:0.9.1.dev20240203-cuda11.8`
+- 强制导入 (可选)      : 是否强制导入镜像（true 或 false），默认为 false
+
+示例:
+
+- 默认使用:
+
+  ```bash
+ ./load.sh
+  ```
+
+- 指定镜像文件路径和标签:
+
+```bash
+./load.sh /path/to/shared/abc.tar myrepo/myimage:latest
+```
+
+- 强制导入镜像:    
+
+```bash
+./load.sh /path/to/shared/abc.tar myrepo/myimage:latest true
+```
+
+
+
+
diff --git a/Classification/resnet50/1_get_docker_image/load.sh b/Classification/resnet50/1_get_docker_image/load.sh
new file mode 100755
index 0000000..5df8bcd
--- /dev/null
+++ b/Classification/resnet50/1_get_docker_image/load.sh
@@ -0,0 +1,26 @@
+#!/bin/bash
+
+if [ -n "$1" ]; then
+  docker_image_path=$1
+else
+  docker_image_path="/share_nfs/k85/oneflow.0.9.1.dev20240203-cuda11.8.tar"
+fi
+
+if [ -n "$2" ]; then
+  docker_image_tag=$2
+else
+  docker_image_tag="oneflowinc/oneflow:0.9.1.dev20240203-cuda11.8"
+fi
+
+if [ -n "$3" ]; then
+  force_load=$3
+else
+  force_load=false
+fi
+
+ansible-playbook \
+    -i ../inventory.ini \
+    load_and_tag_docker_image.yml \
+    -e "docker_image_path=$docker_image_path" \
+    -e "docker_image_tag=$docker_image_tag" \
+    -e "force_load=$force_load"
diff --git a/Classification/resnet50/1_get_docker_image/load_and_tag_docker_image.yml b/Classification/resnet50/1_get_docker_image/load_and_tag_docker_image.yml
new file mode 100644
index 0000000..5a2f92d
--- /dev/null
+++ b/Classification/resnet50/1_get_docker_image/load_and_tag_docker_image.yml
@@ -0,0 +1,28 @@
+---
+- name: Load and tag Docker image
+  hosts: all
+  vars:
+    docker_image_path: "/share_nfs/k85/oneflow.0.9.1.dev20240203-cuda11.8.tar"
+    docker_image_tag: "oneflowinc/oneflow:0.9.1.dev20240203-cuda11.8"
+    force_load: false
+
+  tasks:
+    - name: Check if Docker image with the specified tag already exists
+      command: "docker images -q {{ docker_image_tag }}"
+      register: image_id
+      changed_when: false
+      when: not force_load
+
+    - name: Load Docker image from tar file
+      command: "docker load -i {{ docker_image_path }}"
+      when: force_load or image_id.stdout == ""
+      register: load_output
+
+    - name: Get image ID from load output
+      set_fact:
+        loaded_image_id: "{{ load_output.stdout_lines[-1] | regex_search('sha256:[0-9a-f]+') }}"
+      when: force_load or image_id.stdout == ""
+
+    - name: Tag the loaded Docker image
+      command: "docker tag {{ loaded_image_id }} {{ docker_image_tag }}"
+      when: force_load or image_id.stdout == ""
diff --git a/Classification/resnet50/1_get_docker_image/pull.sh b/Classification/resnet50/1_get_docker_image/pull.sh
new file mode 100755
index 0000000..8787fea
--- /dev/null
+++ b/Classification/resnet50/1_get_docker_image/pull.sh
@@ -0,0 +1,7 @@
+#!/bin/bash
+
+if [ -n "$1" ]; then
+  ansible-playbook -i ../inventory.ini pull_docker_image.yml -e "docker_image=$1"
+else
+  ansible-playbook -i ../inventory.ini pull_docker_image.yml
+fi
diff --git a/Classification/resnet50/1_get_docker_image/pull_docker_image.yml b/Classification/resnet50/1_get_docker_image/pull_docker_image.yml
new file mode 100644
index 0000000..d086e6a
--- /dev/null
+++ b/Classification/resnet50/1_get_docker_image/pull_docker_image.yml
@@ -0,0 +1,17 @@
+---
+- name: Pull specified Docker image
+  hosts: all
+  vars:
+    docker_image: "oneflowinc/oneflow:0.9.1.dev20240203-cuda11.8"
+
+  tasks:
+    - name: Check if the Docker image is already present
+      command: "docker images -q {{ docker_image }}"
+      register: docker_image_id
+      changed_when: false
+
+    - name: Pull Docker image if not present
+      docker_image:
+        name: "{{ docker_image }}"
+        source: pull
+      when: docker_image_id.stdout == ""
diff --git a/Classification/resnet50/Readme.md b/Classification/resnet50/Readme.md
new file mode 100644
index 0000000..b19a080
--- /dev/null
+++ b/Classification/resnet50/Readme.md
@@ -0,0 +1,225 @@
+# 千卡 0.85
+
+[toc]
+
+## 文件目录结构
+```
+├── ansible_workspace # 主节点上的工作目录
+│   ├── inventory.ini # 用来配置节点信息
+│   ├── set_docker.sh # 在各节点上创建docker，并且配置好docker内环境
+│   ├── profile.sh # 根据节点数启动profile
+│   ├── train.sh # 根据节点数启动训练
+│   └── update_tools.sh # 将主节点的tools文件夹复制到各个子节点
+├── tools # 在各个节点使用的文件
+│   ├── args_train_ddp_graph_resnet50.sh # 接受模型训练参数并启动训练
+│   ├── models.tar.gz # 模型，为防止git网络问题，建议先下载放在共享目录下
+│   ├── extract.py # 提取log中train阶段的throughput的平均值
+│   ├── prepare_docker.sh # 用于配置docker内环境
+│   ├── profile.sh # 根据节点数在本机启动profile
+│   └── train.sh # 根据节点数在本机启动训练
+└── Readme.md
+```
+
+需求：有NVLink，以及 shared_nfs
+
+以下供参考
+
+## 第一步： 配置环境
+
+### 1.1 所有节点配置SSH Key，并设置authorized_keys
+
+（怎么自动化）
+
+需要一个共享的存储空间，如：`/shared_nfs/k85`，在一个文件夹下准备好 
+
+- authorized_keys : 在主节点运行 
+
+  ```bash
+  #!/bin/bash
+  
+  # 设置 SSH 目录路径
+  SSH_DIR="$HOME/.ssh"
+  
+  # 检查 SSH 目录是否存在，如果不存在则创建
+  if [ ! -d "$SSH_DIR" ]; then
+    mkdir -p "$SSH_DIR"
+    echo "Created directory: $SSH_DIR"
+  fi
+  
+  # 设置密钥文件路径
+  KEY_PATH="$SSH_DIR/id_rsa"
+  
+  # 生成 SSH 密钥对
+  ssh-keygen -t rsa -b 2048 -f "$KEY_PATH" -N "" -q
+  
+  # 创建 authorized_keys 文件
+  cat $SSH_DIR/id_rsa.pub > $SSH_DIR/authorized_keys
+  
+  # 将 authorized_keys 文件拷贝到共享目录
+  cp $SSH_DIR/authorized_keys shared_nfs/k85
+  ```
+
+- 在子节点运行
+
+  ```bash
+  #!/bin/bash
+  
+  # 设置 SSH 目录路径
+  SSH_DIR="$HOME/.ssh"
+  
+  # 检查 SSH 目录是否存在，如果不存在则创建
+  if [ ! -d "$SSH_DIR" ]; then
+    mkdir -p "$SSH_DIR"
+    echo "Created directory: $SSH_DIR"
+  fi
+  
+  # 设置密钥文件路径
+  KEY_PATH="$SSH_DIR/id_rsa"
+  
+  # 生成 SSH 密钥对
+  ssh-keygen -t rsa -b 2048 -f "$KEY_PATH" -N "" -q
+  
+  # 将 authorized_keys 文件拷贝到 .ssh 目录
+  cp shared_nfs/k85/authorized_keys $SSH_DIR
+  ```
+
+### 1.2 主节点安装 Ansible，并配置节点ip
+示例文件：./ansible_workspace/inventory.ini
+```ini
+[hosts]
+of27 ansible_host=192.168.1.27 ansible_ssh_common_args='-o StrictHostKeyChecking=no'
+of25 ansible_host=192.168.1.25 ansible_ssh_common_args='-o StrictHostKeyChecking=no'
+of26 ansible_host=192.168.1.26 ansible_ssh_common_args='-o StrictHostKeyChecking=no'
+of28 ansible_host=192.168.1.28 ansible_ssh_common_args='-o StrictHostKeyChecking=no'
+```
+
+### 1.3 共享目录中拷贝镜像、数据集、models脚本
+主要为设置docker内环境的脚本 和 启动docker内训练的脚本
+设置docker内环境脚本(./tools/prepare_docker.sh)如下：
+```Bash
+#!/bin/bash
+# 将tools视为共享目录
+pip config set global.index-url https://pypi.tuna.tsinghua.edu.cn/simple
+python3 -m pip install --upgrade pip
+python3 -m pip install --pre oneflow -f https://oneflow-staging.oss-cn-beijing.aliyuncs.com/branch/master/cu121
+
+
+cd /workspace 
+cp tools/models.tar.gz ./
+tar -xvf models.tar.gz
+pip install -r models/dev-requirements.txt
+pip install -r models/Vision/classification/image/resnet50/requirements.txt
+
+# 将需要使用到的脚本拷到对应文件夹下
+cp tools/args_train_ddp_graph_resnet50.sh models/Vision/classification/image/resnet50/examples/
+cp tools/train.sh models/Vision/classification/image/resnet50/
+cp tools/profile.sh models/Vision/classification/image/resnet50/
+```
+启动dokcer内训练的脚本(./tools/train.sh)如下：
+```Bash
+# 根据使用的节点数，来判断本机是否开始训练
+NUM_NODES=${1:-1}
+
+if [ "$NODE_RANK" -lt "$NUM_NODES" ]; then
+  bash examples/args_train_ddp_graph_resnet50.sh "$NUM_NODES" 8 "$NODE_RANK" 192.168.1.27 /data/dataset/ImageNet/ofrecord 192 50 true python3 graph gpu 100 false '' 1
+else
+  echo do nothing
+fi
+```
+启动dokcer内profile(./tools/profile.sh)如下：
+```Bash
+# 根据使用的节点数，来判断是否在本地开始profile
+NUM_NODES=${1:-1}
+
+if [ "$NODE_RANK" -lt "$NUM_NODES" ]; then
+  # 在启动训练时添加nsys启动路径，即可进行profile
+  bash examples/args_train_ddp_graph_resnet50.sh "$NUM_NODES" 8 "$NODE_RANK" 192.168.1.27 /data/dataset/ImageNet/ofrecord 192 50 true python3 graph gpu 100 false '/usr/local/cuda/bin/nsys' 1
+else
+  echo do nothing
+fi
+```
+args_train_ddp_graph_resnet50.sh文件参考自OneAutoTest仓库[args_train_ddp_graph.sh](https://github.com/Oneflow-Inc/OneAutoTest/blob/main/ResNet50/args_train_ddp_graph.sh)，其中包含使用nsys启动的选项
+### 1.4 使用ansible 在所有节点执行 docker load, docker tag命令
+根据上文中inventory.ini文件依次在节点上创建docker，并将NODE_RANK写入docker的环境变量内，脚本(./ansible_workspace/set_docker.sh)内容如下：
+```Bash
+set -ex
+if [ $# -ne 1 ]; then
+  echo "Usage: $0 filename"
+  exit 1
+fi
+host_file="$1"
+num_hosts=$(wc -l < "$host_file")
+docker_name="cd_test"
+
+mapfile -t lines < "$host_file"
+
+for (( i=1; i<${#lines[@]}; i++ )); do
+  line="${lines[$i]}"
+  host_name=$(echo "$line" | awk '{print $1}')
+  # 根据inventory.ini文件中节点顺序，将NODE_RANK写入docker的环境变量中
+  ansible $host_name -i $host_file -m shell -a "docker run -itd -e NODE_RANK=$((i-1)) -v /data/dataset/ImageNet:/data/dataset/ImageNet -v /data/home/chende/tools:/workspace/tools --network host --gpus all --shm-size=16g --ulimit memlock=-1 --ulimit core=0 --ulimit stack=67108864 --privileged --ipc host --cap-add=IPC_LOCK --name $docker_name nvcr.io/nvidia/pytorch:24.03-py3 bash"
+done
+# 在docker内运行环境设置的脚本
+ansible hosts -i "$host_file" -m shell -a "docker exec $docker_name bash -c 'bash /workspace/tools/prepare_docker.sh'"
+```
+使用方式：
+```Bash
+bash set_docker.sh inventory.ini
+```
+
+## 第二步：进行测试
+
+### 2.1 自动测试与日志搜集
+
+编写一个测试命令脚本文件（./ansible_workspace/train.sh）
+```Bash
+#!/bin/bash
+set -ex
+if [ $# -ne 1 ]; then
+  echo "Usage: $0 num_nodes"
+  exit 1
+fi
+NUM_NODES="$1"
+docker_name="cd_test_new"
+ansible hosts -i inventory.ini -m shell -a "docker exec $docker_name bash -c 'cd /workspace/models/Vision/classification/image/resnet50 && bash train.sh $NUM_NODES'"
+```
+
+- 需要一个参数: 节点数，
+- 运行该命令能够自动启动相应数量的节点运行。
+- 运行结束后收集日志到主节点。
+- 保存日志的目录可以以：`prefix_节点数_日期时间_surfix` 命名，前缀和后缀可以自定义
+
+### 2.2 自动日志解析
+
+可以使用2.1节提供的命令运行多次，比如：
+
+```bash
+train.sh 1
+train.sh 2
+train.sh 4
+train.sh 8
+train.sh 16
+```
+
+完成后应该保存了多个日志目录，需要编写一个日志处理脚本，从这些日志目录中提取性能数据并制成 markdown 格式的表格
+
+注：不需要完整训练，训练稳定后获取到数据就可以了。
+
+### 2.3 自动 nsys 性能测试
+
+需要编写一个能够运行 nsys 的性能测试脚本文件（./ansible_workspace/profile.sh），和2.1的脚本类似，只是启动时需要调用nsys，我们需要搜集这些信息分析，然后进行优化。这个脚本文件。
+```Bash
+#!/bin/bash
+set -ex
+if [ $# -ne 1 ]; then
+  echo "Usage: $0 num_nodes"
+  exit 1
+fi
+NUM_NODES="$1"
+docker_name="cd_test_new"
+ansible hosts -i inventory.ini -m shell -a "docker exec $docker_name bash -c 'cd /workspace/models/Vision/classification/image/resnet50 && bash profile.sh $NUM_NODES'"
+```
+- 需要一个参数: 节点数，
+- 运行该命令能够自动启动相应数量的节点运行。
+- 运行结束后收集日志和nsys相关文件到主节点。
+- 保存日志的目录可以以：`prefix_节点数_日期时间_surfix` 命名，前缀和后缀可以自定义
\ No newline at end of file
diff --git a/Classification/resnet50/ansible_workspace/profile.sh b/Classification/resnet50/ansible_workspace/profile.sh
new file mode 100644
index 0000000..db6abfb
--- /dev/null
+++ b/Classification/resnet50/ansible_workspace/profile.sh
@@ -0,0 +1,9 @@
+#!/bin/bash
+set -ex
+if [ $# -ne 1 ]; then
+  echo "Usage: $0 num_nodes"
+  exit 1
+fi
+NUM_NODES="$1"
+docker_name="cd_test_new"
+ansible hosts -i inventory.ini -m shell -a "docker exec $docker_name bash -c 'cd /workspace/models/Vision/classification/image/resnet50 && bash profile.sh $NUM_NODES'"
\ No newline at end of file
diff --git a/Classification/resnet50/ansible_workspace/set_docker.sh b/Classification/resnet50/ansible_workspace/set_docker.sh
new file mode 100755
index 0000000..7a55f49
--- /dev/null
+++ b/Classification/resnet50/ansible_workspace/set_docker.sh
@@ -0,0 +1,18 @@
+#!/bin/bash
+set -ex
+if [ $# -ne 1 ]; then
+  echo "Usage: $0 filename"
+  exit 1
+fi
+host_file="$1"
+num_hosts=$(wc -l < "$host_file")
+docker_name="cd_test_new"
+
+mapfile -t lines < "$host_file"
+
+for (( i=1; i<${#lines[@]}; i++ )); do
+  line="${lines[$i]}"
+  host_name=$(echo "$line" | awk '{print $1}')
+  ansible $host_name -i $host_file -m shell -a "docker run -itd -e NODE_RANK=$((i-1)) -v /data/dataset/ImageNet:/data/dataset/ImageNet -v /data/home/chende/tools:/workspace/tools --network host --gpus all --shm-size=16g --ulimit memlock=-1 --ulimit core=0 --ulimit stack=67108864 --privileged --ipc host --cap-add=IPC_LOCK --name $docker_name nvcr.io/nvidia/pytorch:24.03-py3 bash"
+done
+ansible hosts -i "$host_file" -m shell -a "docker exec $docker_name bash -c 'bash /workspace/tools/prepare_docker.sh'"
\ No newline at end of file
diff --git a/Classification/resnet50/ansible_workspace/train.sh b/Classification/resnet50/ansible_workspace/train.sh
new file mode 100644
index 0000000..350bf78
--- /dev/null
+++ b/Classification/resnet50/ansible_workspace/train.sh
@@ -0,0 +1,9 @@
+#!/bin/bash
+set -ex
+if [ $# -ne 1 ]; then
+  echo "Usage: $0 num_nodes"
+  exit 1
+fi
+NUM_NODES="$1"
+docker_name="cd_test_new"
+ansible hosts -i inventory.ini -m shell -a "docker exec $docker_name bash -c 'cd /workspace/models/Vision/classification/image/resnet50 && bash train.sh $NUM_NODES'"
\ No newline at end of file
diff --git a/Classification/resnet50/ansible_workspace/update_tools.sh b/Classification/resnet50/ansible_workspace/update_tools.sh
new file mode 100755
index 0000000..cda6deb
--- /dev/null
+++ b/Classification/resnet50/ansible_workspace/update_tools.sh
@@ -0,0 +1,3 @@
+#!/bin/bash
+echo "/data/home/chende/tools"
+ansible hosts -i inventory.ini -m copy -a "src=/data/home/chende/tools dest=/data/home/chende/ mode=0755"
\ No newline at end of file
diff --git a/Classification/resnet50/inventory.ini b/Classification/resnet50/inventory.ini
new file mode 100644
index 0000000..01027fd
--- /dev/null
+++ b/Classification/resnet50/inventory.ini
@@ -0,0 +1,3 @@
+[all]
+of27 ansible_host=192.168.1.27 ansible_ssh_common_args='-o StrictHostKeyChecking=no'
+of28 ansible_host=192.168.1.28 ansible_ssh_common_args='-o StrictHostKeyChecking=no'
diff --git a/Classification/resnet50/tools/args_train_ddp_graph_resnet50.sh b/Classification/resnet50/tools/args_train_ddp_graph_resnet50.sh
new file mode 100755
index 0000000..b4fbd36
--- /dev/null
+++ b/Classification/resnet50/tools/args_train_ddp_graph_resnet50.sh
@@ -0,0 +1,144 @@
+rm -rf core.*
+
+set -ex
+
+
+# bash examples/args_train_ddp_graph.sh ${NUM_NODES} ${DEVICE_NUM_PER_NODE} ${NODE_RANK} ${MASTER_ADDR}
+# ${OFRECORD_PATH} ${TRAIN_BATCH_SIZE} ${EPOCH} ${USE_FP16} ${PYTHON_BIN} ${RUN_TYPE} ${DEBUG_AND_NCCL} ${NSYS_BIN} ${RUN_COMMIT}
+
+# bash examples/args_train_ddp_graph.sh 1 8 0 127.0.0.1 /dataset/79846248 192 50 false python3 ddp false '' 1
+
+NUM_NODES=${1:-1}
+DEVICE_NUM_PER_NODE=${2:-8}
+NODE_RANK=${3:-0}
+MASTER_ADDR=${4:-"127.0.0.1"}
+OFRECORD_PATH=${5:-"/dataset/imagenet/ofrecord"}
+TRAIN_BATCH_SIZE=${6:-192}
+EPOCH=${7:-50}
+USE_FP16=${8:-false}
+PYTHON_BIN=${9:-"python3"}
+RUN_TYPE=${10:-"ddp"} # graph+fp16
+DECODE_TYPE=${11:-"cpu"}
+PRINT_INTERVAL=${12:-100}
+DEBUG_AND_NCCL=${13:-false}
+NSYS_BIN=${14:-""}
+RUN_COMMIT=${15:-"master"}
+ACC=${16:-1}
+VAL_BATCH_SIZE=${17:-50}
+
+
+SRC_DIR=$(realpath $(dirname $0)/..)
+
+AMP_OR="FP32"
+if $USE_FP16; then
+    AMP_OR="FP16"
+fi
+
+TRAN_MODEL="resnet50"
+RUN_TIME=$(date "+%Y%m%d_%H%M%S%N")
+LOG_FOLDER=${SRC_DIR}/test_logs/$HOSTNAME/${NUM_NODES}n${DEVICE_NUM_PER_NODE}g
+mkdir -p $LOG_FOLDER
+LOG_FILENAME=$LOG_FOLDER/${TRAN_MODEL}_${RUN_TYPE}_DC${DECODE_TYPE}_${AMP_OR}_mb${TRAIN_BATCH_SIZE}_gb$((${TRAIN_BATCH_SIZE}*${NUM_NODES}*${DEVICE_NUM_PER_NODE}*${ACC}))_acc${ACC}_${NUM_NODES}n${DEVICE_NUM_PER_NODE}g_${RUN_COMMIT}_${RUN_TIME}
+
+
+export PYTHONUNBUFFERED=1
+echo PYTHONUNBUFFERED=$PYTHONUNBUFFERED
+#export ONEFLOW_COMM_NET_IB_ENABLE=True
+export NCCL_LAUNCH_MODE=GROUP
+echo NCCL_LAUNCH_MODE=$NCCL_LAUNCH_MODE
+echo DEBUG_AND_NCCL=$DEBUG_AND_NCCL
+if $DEBUG_AND_NCCL; then
+    export ONEFLOW_DEBUG_MODE=1
+    echo ONEFLOW_DEBUG_MODE=$ONEFLOW_DEBUG_MODE
+    export NCCL_DEBUG=INFO
+    echo NCCL_DEBUG=$NCCL_DEBUG
+fi
+
+#export ONEFLOW_KERNEL_ENABLE_CUDA_GRAPH=1
+#export ONEFLOW_THREAD_ENABLE_LOCAL_MESSAGE_QUEUE=1
+#export ONEFLOW_KERNEL_DISABLE_BLOB_ACCESS_CHECKER=1
+#export ONEFLOW_ACTOR_ENABLE_LIGHT_ACTOR=1
+#export ONEFLOW_STREAM_REUSE_CUDA_EVENT=1
+
+#export ONEFLOW_STREAM_CUDA_EVENT_FLAG_BLOCKING_SYNC=true
+#export ONEFLOW_VM_WORKLOAD_ON_SCHEDULER_THREAD=1
+
+LEARNING_RATE=$(echo | awk "{print $NUM_NODES*$DEVICE_NUM_PER_NODE*$TRAIN_BATCH_SIZE*$ACC/1000}")
+MOM=0.875
+OFRECORD_PART_NUM=256
+
+EXIT_NUM=-1
+
+if [ ${EPOCH} -lt 10 ];then
+    EXIT_NUM=300
+fi
+CMD=""
+
+if [[ ! -z "${NSYS_BIN}" ]]; then
+    export ONEFLOW_PROFILER_KERNEL_PROFILE_KERNEL_FORWARD_RANGE=1
+    export ONEFLOW_DEBUG_MODE=1
+    # CMD+="${NSYS_BIN} profile --stats true -t nvtx --output ${LOG_FILENAME} "
+    export CUDNN_LOGINFO_DBG=1
+    export CUDNN_LOGDEST_DBG=${SRC_DIR}/cudnn.log
+    CMD+="${NSYS_BIN} profile --stats true --output ${LOG_FILENAME} "
+    EXIT_NUM=30
+fi
+
+
+CMD+="${PYTHON_BIN} -m oneflow.distributed.launch "
+
+CMD+="--nproc_per_node ${DEVICE_NUM_PER_NODE} "
+CMD+="--nnodes ${NUM_NODES} "
+CMD+="--node_rank ${NODE_RANK} "
+CMD+="--master_addr ${MASTER_ADDR} "
+CMD+="--master_port 12345 "
+CMD+="${SRC_DIR}/train.py "
+CMD+="--ofrecord-path ${OFRECORD_PATH} "
+CMD+="--ofrecord-part-num ${OFRECORD_PART_NUM} "
+CMD+="--num-devices-per-node ${DEVICE_NUM_PER_NODE} "
+CMD+="--lr ${LEARNING_RATE} "
+CMD+="--momentum ${MOM} "
+CMD+="--num-epochs ${EPOCH} "
+CMD+="--train-batch-size ${TRAIN_BATCH_SIZE} "
+CMD+="--train-global-batch-size $((${TRAIN_BATCH_SIZE}*${NUM_NODES}*${DEVICE_NUM_PER_NODE}*${ACC})) "
+CMD+="--val-batch-size ${VAL_BATCH_SIZE} "
+CMD+="--val-global-batch-size $((${VAL_BATCH_SIZE}*${NUM_NODES}*${DEVICE_NUM_PER_NODE}*${ACC})) "
+CMD+="--print-interval ${PRINT_INTERVAL} "
+#CMD+="--synthetic-data "
+
+if $USE_FP16; then
+    echo USE_FP16=$USE_FP16
+    CMD+="--use-fp16 --channel-last "
+fi
+
+if [ $EXIT_NUM != -1 ]; then
+    CMD+="--skip-eval "
+fi
+if [ $RUN_TYPE == 'ddp' ]; then
+    CMD+="--ddp "
+else
+    CMD+="--scale-grad --graph "
+    CMD+="--fuse-bn-relu "
+    CMD+="--fuse-bn-add-relu "
+fi
+
+if [ $DECODE_TYPE == 'gpu' ]; then
+    CMD+="--use-gpu-decode "
+fi
+
+echo "Rum cmd ${CMD}"
+
+$CMD 2>&1 | tee ${LOG_FILENAME}.log
+
+echo "Writting log to ${LOG_FILENAME}.log"
+
+if [[ ! -z "${NSYS_BIN}" ]]; then
+    rm ${LOG_FOLDER}/*.sqlite
+    mkdir -p ${LOG_FILENAME}
+    #rm -rf ./log/$HOSTNAME/oneflow.*
+    cp ./log/$HOSTNAME/* ${LOG_FILENAME}/
+    mv ${SRC_DIR}/cudnn.log ${LOG_FILENAME}/cudnn.log
+fi
+
+rm -rf ./log/$HOSTNAME
+echo "done"
diff --git a/Classification/resnet50/tools/extract.py b/Classification/resnet50/tools/extract.py
new file mode 100644
index 0000000..477903d
--- /dev/null
+++ b/Classification/resnet50/tools/extract.py
@@ -0,0 +1,27 @@
+import sys
+import re
+
+# 文件路径
+file_path = sys.argv[1]
+
+# 存储 train 模式下的 throughput
+print(file_path)
+train_throughputs = []
+
+# 正则表达式模式匹配 train 模式下的 throughput
+pattern = re.compile(r'\[train\][^|]*?throughput:\s(\d+\.\d+)')
+
+# 读取文件并提取需要的信息
+with open(file_path, 'r') as file:
+    for line in file:
+        matches = pattern.findall(line)
+        for match in matches:
+            throughput = float(match)
+            train_throughputs.append(throughput)
+
+# 计算平均 throughput
+if train_throughputs:
+    average_throughput = sum(train_throughputs) / len(train_throughputs)
+    print(f'The average throughput for [train] mode is: {average_throughput:.6f}')
+else:
+    print('No [train] mode throughputs found.')
diff --git a/Classification/resnet50/tools/prepare_docker.sh b/Classification/resnet50/tools/prepare_docker.sh
new file mode 100755
index 0000000..84147f5
--- /dev/null
+++ b/Classification/resnet50/tools/prepare_docker.sh
@@ -0,0 +1,14 @@
+#!/bin/bash
+pip config set global.index-url https://pypi.tuna.tsinghua.edu.cn/simple
+python3 -m pip install --upgrade pip
+python3 -m pip install --pre oneflow -f https://oneflow-staging.oss-cn-beijing.aliyuncs.com/branch/master/cu121
+
+cd /workspace 
+cp tools/models.tar.gz ./
+tar -xvf models.tar.gz
+pip install -r models/dev-requirements.txt
+pip install -r models/Vision/classification/image/resnet50/requirements.txt
+
+cp tools/args_train_ddp_graph_resnet50.sh models/Vision/classification/image/resnet50/examples/
+cp tools/train.sh models/Vision/classification/image/resnet50/
+cp tools/profile.sh models/Vision/classification/image/resnet50/
\ No newline at end of file
diff --git a/Classification/resnet50/tools/profile.sh b/Classification/resnet50/tools/profile.sh
new file mode 100755
index 0000000..79e0d0a
--- /dev/null
+++ b/Classification/resnet50/tools/profile.sh
@@ -0,0 +1,7 @@
+NUM_NODES=${1:-1}
+
+if [ "$NODE_RANK" -lt "$NUM_NODES" ]; then
+  bash examples/args_train_ddp_graph_resnet50.sh "$NUM_NODES" 8 "$NODE_RANK" 192.168.1.27 /data/dataset/ImageNet/ofrecord 128 50 true python3 graph gpu 100 false '/usr/local/cuda/bin/nsys' 1
+else
+  echo do nothing
+fi
\ No newline at end of file
diff --git a/Classification/resnet50/tools/train.sh b/Classification/resnet50/tools/train.sh
new file mode 100755
index 0000000..5758c05
--- /dev/null
+++ b/Classification/resnet50/tools/train.sh
@@ -0,0 +1,7 @@
+NUM_NODES=${1:-1}
+
+if [ "$NODE_RANK" -lt "$NUM_NODES" ]; then
+  bash examples/args_train_ddp_graph_resnet50.sh "$NUM_NODES" 8 "$NODE_RANK" 192.168.1.27 /data/dataset/ImageNet/ofrecord 128 50 true python3 graph gpu 100 false '' 1
+else
+  echo do nothing
+fi
\ No newline at end of file