GPU server considerations

vpn

Even on vpn You can't use the browser after , But you can surf the Internet （zoom, Wechat is still available ）,ping ip No problem with the address , but ping There is something wrong with the web page name , representative dns There is a problem . Change at this time dns Just go （ Replaced by Google 8.8.8.8）

Connect （vscode）

1 Download plug-ins ：

2 Click on the plus sign

3 Enter the server address and user name

4 Click on the plus sign , Enter the server

5 Input password

6 openfolder, Locate folder , Left visualization window

Connect (pycharm)

pycharm Because it was written for a long time and it was very small at that time , So the narration may be a little confused …
Simply put, it is to connect the local directory and the server directory + Choose compiler （ It's the environment ）,pycharm Unlike vscode It is a direct visualization of server content , It is equivalent to connecting the local and server , So than vscode There is an additional step of uploading from the local to the server .
1、 stay linux Run in py file ：
python name.py
2、pycharm Configure virtual compiler ：
1）tool-deployment-configuration

+sftp

Enter the server name （ Customize ）

connection-configuration（ It can be used test connection try ）

mappings：
localpath And deploymentpath： They are the local project file directory and the project file in the server

2）file-settings-project-python interpreter-add-ssh interpreter

next-interpreter Write python The location of the interpreter
（linux Into python,
import sys
sys.executable）
commonly bin Drop executable

file-settings-interpreter-show all：

Click pen to change to ssh（ It was configuration）

Be careful , The interpreter cannot run local files when connected to the server , Otherwise, the report will be wrong ：

want pycharm deployment Synchronize local projects to the server

deployment path Write well , And then right-click project,deployment-up to xxx

upload success

deployment Don't set too many paths , There may be conflict

pycharm Remote connection server compilation error ： No such file or directory

reason ：
1、 Maybe the local file was not uploaded to the server ; resolvent ：deployment-mapping Establish a connection and upload ;
2、setting-interpreter Inside mapping No settings

In this way, the setting is completed , spot ok After running successfully

tmux

( Here will be the default ctrlb Changed to ctrla）
First in the terminal ssh [email protected], Sign in
tmux ls See which windows are available
Input tumx Get into tmux page
tmux attach -t number： Get into indx by number The window of
ctrla+| Left and right split screen
ctrla± Split up and down
ctrla+c Open a new window
ctrla+q Turn off split screen / window
ctrla+n Next window /ctrlc+N Last window
ctrl+d Exit but not close
ctrla+r Refresh
ctrla+z Maximize pannel
tmux attach -t Number -d Access to the session And forcibly withdraw from other session
ctrla m Turn off the mouse
ctrla M Open the mouse

gpu command

see cuda edition ：vvcc --version

watch -n 0.1 -d nvidia-smi View dynamic memory
watch -n 0.1 --color gpustat --c View dynamic memory （ Small view ）

nvidia-smi
View static memory

Release gpu Memory （ After the program runs, it still displays gpu Memory occupied ）：
fuser -v /dev/nvidia* # Lookup occupation GPU Resources PID
kill -9 pid

Use specified gpu:
original text
CUDA_VISIBLE_DEVICES=1 python my_script.py
or
os.environ[‘CUDA_VISIBLE_DEVICES’] = ‘3,4,5,6,7’（ The code is written immediately ）

Kill the process by force ： When the program gets stuck / Stop but gpu Memory is not free ：
gpystat -p Check the process pid
kill -9 pid

linux command

~ representative /home/ User name directory
Look at the hidden files under the user name directory ：（ for example bashrc）：~/.bashrc
ls -a Look at the hidden file

# Unrealized
Restore deleted files
here

Document processing

Download and unzip the file
Use wget -O rename url download
decompression ：

Unzip big files （ exceed 2 individual g）
7z x xxx.zip

Local and server file transfer
Upload files locally ：
scp /path/filename [email protected]:/path
Local upload directory ：
scp -r local_dir [email protected]:remote_dir
Server downloads files ：
scp [email protected]:/path/filename /path/filename（ Local directory ）
Server download directory ：
scp -r [email protected]:/path（ Remote directory ） /path（ Local directory ）

Moving files ：
https://blog.csdn.net/u010579482/article/details/72081996
mv sourcePath destPath
example ：
mv /home/lyli/codeRepete/PatchAttack/942d3a0ac09ec9e5eb3a/* /home/lyli/codeRepete/PatchAttack
take 94… The contents of the folder are placed in PatchAttack in

Delete folder ：
rm -rf Directory name
Cannot be in the directory a Delete the directory a

Look at the number of files ：
ll | grep -c “^-”
ll | grep -c “^d”

Check the space occupied by the hard disk
du -lh --max-depth=1 Check whether the subdirectories of the current directory occupy
du -sh Only view the current directory
df -hl Check disk space usage

View remaining memory
free -m Company mb
free -g Company gb

vim

vim file name Get into vim

:q sign out vim
：wq Save and exit

Enter vim ~/.vimrc （ Create... In the home directory vimrc file ）
stay .vimrc Configuration in

Input / key word You can search for keywords , Press enter+n To find the next
Select replication ：ctrl+shift+ cursor
Paste ：shift+insert

putty Connect

Continuous line method ：

Environment related

3090 And cuDA
View servers cuda：nvcc --version
Pay attention to the server cuda The version is best and torch Of cuda Version match , Otherwise it will appear model.cuda() Problems that cannot be implemented

conda info -e : View all conda Environmental Science
python -V see python edition
see torch,cuda And tf edition ：
python

import torch
import tensorflow as tf
toech.__version__
tf.__version__
torch.version.cuda

perhaps pip list grep | torch

Check a path with keywords ：( example ： Look at the pip）
ls ~/anaconda3/envs/tf115/bin/ | grep pip

which pip see pip The path of （ Other orders are the same ）
If pip The path is not right （ Not under this environment path , May cause the current channel There is no package to download ）, function
export PATH=“/home/ user name /anaconda3/bin:$PATH”, And then restart （ in base Switch back to the current environment ） Check which pip
perhaps ：conda install -c conda-forge Package name , This is directly from conda forge download

tail -30 ~/.bashrc Look at the hidden file bashrc After 30 That's ok
echo $PATH Look at the environment variables （ Colon division ）

jupyter problem

jupyter The configuration file ： vim /home/lyli/.jupyter/jupyter_notebook_config.py
or ：~/.jupyter/jupyter_notebook_config.py
Change or add kernal： Reference here
pip install ipykernel
python -m ipykernel install --name your_invironment_name

My environment ：

The following environmental records are created for individuals only
torch1：python3.8.8 ;torch 1.11.0+cu102; tf:1.15.0
test1: python 3.8.10 ; torch 1.11.0+cu102; tf:1.15.5
tf115：python 3.6.9 ; torch 1.10.2 ;tf：1.15.0 （ A bit of a problem. ）
torch1_0: python 3.7.11; torch 1.0.1 ;tf:null
torchEnv: python 3.8.13; torch 1.11.0+cu102; tf: 1.15.5
above cuda All are 10.2
p_3.6： python 3.6.13 ; torch 1.10.2 cuda：11.1