This is deprecated. Old instructions below.
This is a single or multiuser CoCalc server that you can very easily install on your own computer using Docker in a few minutes. If you need something to install on a cluster of servers using Kubernetes, see cocalc-cloud.
GitHub: https://github.com/sagemathinc/cocalc-docker
Docs: https://doc.cocalc.com/docker-image.html
Quickstart on a Mac or Linux server with Docker installed
- Make sure you have at least 25GB disk space free and Docker installed on your computer.
- Run something like this, depending on your requirements:
docker run --name=cocalc -d -v ~/cocalc:/projects -p 443:443 sagemathinc/cocalc-docker- Wait a for the 22GB image to pull, decompress and the container to start, then visit https://localhost.
For other operating systems and way more details, see below.
NOTE: The above will work fine on both x86_64 and aarch64 (e.g., Apple Silicon M1 macs). Cocalc-docker used to have a different Docker image for aarch64, but now there is only on.
If the above doesn't work due to something else already using port 443 or you wanting to serve cocalc on a different port, you could use -p 4043:443 instead. There is extensive Docker documentation online. For example the following runs cocalc-docker on port 7100:
docker run --name=cocalc -d -v ~/cocalc:/projects -p 7100:443 sagemathinc/cocalc-dockerVERSION NOTE: The sagemathinc/cocalc-docker image is currently not backward compatible with sagemathinc/cocalc, since sagemathinc/cocalc uses the deprecated PostgreSQL version 10, whereas cocalc-docker uses PostgreSQL version 14, and we haven't implemented an automated upgrade path yet. That said, using the standard pg_dumpall and psql tools to export and import the database works.
Connecting to https://localhost
The default cocalc-docker container of course has only a self-signed ssl certificate. Browsers have cracked down more and more on allowing connections to such servers. Because cocalc-docker is serving on localhost, you must explicitly tell your browser to allow the connection. Do a Google search for "chrome allow localhost https" to find out how; one result is https://communicode.io/allow-https-localhost-chrome/.
LICENSE AND SUPPORT:
- Much of this code is licensed under the AGPL condition to the commons clause exception. If you would instead like a business-friendly MIT license instead, please contact help@cocalc.com, and we will sell you a 1-year license for $999, which includes some support (you can pay more for significant support). We do have many happy paying customers, and cocalc-docker is popular.
- Visit the CoCalc Docker discussion board for news, updates and more; there is also a more general CoCalc discussion board.
- There is a specialized CoCalc Docker mailing list, and a more general CoCalc mailing list.
- Read the overview docs.
SECURITY STATUS:
- This is not blatantly insecure from outside attack: the database has a long random password, user accounts are separate, encrypted SSL communication is used by default, etc.
- That said, a determined user with a CoCalc account could in theory easily access or change files of other users in the same container, and can definitely overload the server! Open ports are exposed to users for reading/writing to project files, these can be used by authenticated users for accessing any other project's open files. Requests should only connect to the main hub process, which proxies traffic to the raw server with an auth key created by the project's secret key changing on every project startup, see Issue 45. Also see the related issues for adding a user auth token to all requests required for each separate sub module, including JupyterLab server Issue 46 and classical Jupyter in an iframe Issue 47.
- There is no quota on project resource usage, so users could easily crash the server both intentionally or accidentally by running arbitrary code, and could also overflow the storage container by creating excessive files.
- Use this for personal use, behind a firewall, or with an account creation token, so that only other people you trust create accounts. Don't make one of these publicly available with important data in it and no account creation token! See issue 2031. Basically, use this only with people you trust.
- See the open docker-related CoCalc issues.
- The above security and resource problems are all solved by cocalc-cloud.
Install Docker on your computer (e.g., apt-get install docker.io on Ubuntu). Make sure you have at least 25GB disk space free, then type:
docker run --name=cocalc -d -v ~/cocalc:/projects -p 443:443 sagemathinc/cocalc-docker
wait a few minutes for the image to pull, decompress and the container to start, then visit https://localhost. (If you are using Microsoft Windows, instead open https://host.docker.internal/.) It is expected that you'll see a "Your connection is not private" warning, since you haven't set up a security certificate. Click "Show advanced" and "Proceed to localhost (unsafe)".
NOTES:
-
If you get an error about the Docker daemon, instead run
sudo docker .... -
CoCalc will NOT work over insecure port 80. A previous version of these directions suggested using -p 80:80 above and visiting http://localhost, which will not work.
-
If you are using Microsoft Windows (or possibly MacOS!), instead make a docker volume and use that for storage for vastly better performance:
docker volume create cocalc-volume docker run --name=cocalc -d -v cocalc-volume:/projects -p 443:443 sagemathinc/cocalc-docker -
IMPORTANT: If you are deploying CoCalc for use over the web (so not just on localhost), it is probably necessary to obtain a valid security certificate instead of using the self-signed unsafe one that is in your Docker container. See this discussion.
-
If you are using Ubuntu as a host and would like the CoCalc instance to use your host's time and timezone, you can amend the run command as follows, which will use your host's timezone and localtime files inside the container:
docker run --name=cocalc -d -v ~/cocalc:/projects -v "/etc/timezone:/etc/timezone" -v "/etc/localtime:/etc/localtime" -p 443:443 sagemathinc/cocalc-docker -
Here's a new Step-by-Step Guide to Setting up Cocalc-Docker on a Custom Domain with a valid SSL Certificate in case you want to go beyond just running something locally on your laptop, and want to provide an integrated solution for your community.
The above command will first download the image, then start CoCalc, storing your data in the directory ~/cocalc on your computer. If you want to store your worksheets and edit history elsewhere, change ~/cocalc to something else. Once your local CoCalc is running, open your web browser to https://localhost. (If you are using Microsoft Windows, instead open https://host.docker.internal/.)
The docker container is called cocalc and you can refer to the container and use commands like:
$ docker stop cocalc
$ docker start cocalc
You can watch the logs:
$ docker logs cocalc -f
However, these logs sometimes don't work. In that case get a bash shell in the terminal and look at the logs using tail:
$ docker exec -it cocalc bash
$ tail -f /var/log/hub.log
If you want cocalc-docker to serve everything with a custom base path, e.g., at https://example.com/my/base/path then you have to do two things.
docker run -e BASE_PATH=/my/base/path --name=cocalc -d -v ~/cocalc:/projects -p 443:443 sagemathinc/cocalc-dockerThis sets the base path correctly for most of CoCalc, but not for everything, unfortunately.
(2) Rebuild the next.js package
The next package in CoCalc itself is the only thing that hardcodes the basepath. You have to rebuild it exactly once with the BASE_PATH environment variable set, as follows:
~/cocalc-docker/aarch64$ docker run -e BASE_PATH=/my/base/path --name=cocalc -d -v ~/cocalc:/projects -p 443:443 sagemathinc/cocalc-aarch64
034cf17482a467537addd8ef8db0406277a7f76789eac4b03e0535d8e0d9ccfc
~/cocalc-docker/aarch64$ docker exec -it cocalc bash
root@034cf17482a4:/# umask 022
root@034cf17482a4:/# cd /cocalc/src/packages/next
root@034cf17482a4:/cocalc/src/packages/next# echo $BASE_PATH
/my/base/path
root@034cf17482a4:/cocalc/src/packages/next# time npm run build
real 2m12.900s
# Expect this to take a LONG TIME, e.g., up to 10 minutes, though it
# just took 2 minutes on a fast server for me.
Now exit the docker container and restart it to switch to the new version:
~/cocalc-docker/aarch64$ docker stop cocalc; docker start cocalcNow visit: https://localhost:443/my/base/path/ and it should fully work.
We do much of the development of CoCalc itself on https://cocalc.com using a BASE_PATH. So fortunately BASE_PATH functionality does get used frequently.
Projects will stop by default if they are idle for 30 minutes. Admins can manually increase this for any project. If you want to completely disable the idle timeout functionality, set the COCALC_NO_IDLE_TIMEOUT environment variable. Note that the user interface will still show an idle timeout -- it's just that it will have no impact.
docker run -e COCALC_NO_IDLE_TIMEOUT=yes --name=cocalc -d -v ~/cocalc:/projects -p 443:443 sagemathinc/cocalc-dockerBy default, cocalc-docker creates services on 3 ports:
- 22 -- ssh
- 80 -- a simple http redirect that sends the user to https
- 443 -- the main site, served over https, but by default with a self-signed certificate
If you would like the following configuration instead, pass the --env NOSSL=true option when you create the Docker container:
- 22 -- ssh
- 80 -- the main site, served over htttp
docker run --name=cocalc --env NOSSL=true -d -v ~/cocalc:/projects -p 8080:80 sagemathinc/cocalc-dockerYou might want to do this if you are doing your ssl termination via some sort of external server, e.g., haproxy or nginx.
If you're running multiple sites from a single server using an Nginx reverse proxy, a setup like the following could be useful.
Instead of mapping port 443 on the container to 443 on the host, map 443 on the container to an arbitray unused port on the host, e.g. 9090:
docker run --name=cocalc -d -v ~/cocalc:/projects -p 9090:443 sagemathinc/cocalc-dockerIn your nginx sites-available folder, create a file like the following called e.g. mycocalc:
map $http_upgrade $connection_upgrade {
default upgrade;
'' close;
}
server {
listen 443 ssl;
server_name mycocalc.com;
#These need to be obtained independently for example from https://letsencrypt.org/, by running "certbot certonly" on the docker host after DNS is setup
ssl_certificate /etc/letsencrypt/live/mycocalc.com/fullchain.pem;
ssl_certificate_key /etc/letsencrypt/live/mycocalc.com/privkey.pem;
location / {
# push traffic through the proxy to the port you mapped above, in this case 9090, on the localhost:
proxy_pass https://localhost:9090;
# this enables proxying for websockets, which cocalc uses extensively:
proxy_http_version 1.1;
proxy_set_header Upgrade $http_upgrade;
proxy_set_header Connection $connection_upgrade;
proxy_set_header Host $host;
}
}
And soft-link it to your sites-enabled folder, e.g. sudo ln -s /etc/nginx/sites-available/mycocalc /etc/nginx/sites-enabled/mycocalc
If you're using certbot and letsencrypt, you can then get a certificate for your domain using something like sudo certbot --nginx and selecting "mycocalc.com", which will automatically set up an ssl cert and modify your nginx server file.
I have tested a lot in November 2021, and did NOT have any problems with clock skew with Docker Desktop, so this appears to be fixed. If not -- It is critical that the Docker container have the correct time, since CoCalc assumes that the server has the correct time. On a laptop running Docker under MacOS, the clock may get messed up any time you suspend/resume your laptop. This workaround might work for you: https://github.com/arunvelsriram/docker-time-sync-agent/.
I regularly post an Apple Silicon Aarch64 cocalc-docker image. This runs natively. It should also work on other aarch64 Linux systems.
https://hub.docker.com/r/sagemathinc/cocalc-aarch64
Cocalc-docker by default uses a self signed certificate on localhost.
- Firefox -- works fine: With Firefox you can click through some warnings and use CoCalc-docker just fine.
- Chrome -- does NOT work: With Chrome, you can try to use the workaround here, which involves visiting
chrome://flags/#allow-insecure-localhost, but I've found that the websocket connection to the project is still blocked. - Safari -- works fine: With current Safari, you can click through to "accept the risks", and it works really well.
You can run CoCalc locally on your Chromebook as long as it supports Crostini Linux.
-
Install (Crostini) Linux support -- search for Linux in settings and enable.
-
In the Linux terminal, type
sudo su apt-get update && apt-get upgrade && apt-get install tmux dpkg-dev -
Install Docker as here:
sudo su apt-get install -y \ apt-transport-https \ ca-certificates \ curl \ gnupg2 \ software-properties-common && \ curl -fsSL https://download.docker.com/linux/debian/gpg | apt-key add - && \ apt-key fingerprint 0EBFCD88 && \ add-apt-repository \ "deb [arch=amd64] https://download.docker.com/linux/debian \ $(lsb_release -cs) \ stable" && \ apt-get update && apt-get install -y docker-ce -
Install cocalc-docker:
sudo docker run --name=cocalc -d -v /cocalc:/projects -p 443:443 -p 80:80 sagemathinc/cocalc-dockerType
/sbin/ifconfig eth0|grep inetin the terminal, and use whatever ip address is listed there -- e.g., for me it was https://100.115.92.198/
If you're running this docker image on a remote server and want to use ssh port forwarding to connect, type:
ssh -L 8080:localhost:443 username@remote_server
then open your web browser to https://localhost:8080
For enhanced security, make the container only listen on localhost:
docker stop cocalc
docker rm cocalc
docker run --name=cocalc -d -v ~/cocalc:/projects -p 127.0.0.1:443:443 sagemathinc/cocalc-docker
Then the only way to access your CoCalc server is to type the following on your local computer:
ssh -L 8080:localhost:443 username@remote_server
and open your web browser to https://localhost:8080
IMPORTANT: The ssh directions for cocalc.com involving key management, etc., do not apply to cocalc-docker. cocalc.com uses an "ssh gateway", and uniform key management across all of your projects. Cocalc-docker doesn't implement any of that, and just does ssh access directly, in exactly the same standard way as a generic Linux install.
In order to ssh into cocalc-docker, you must expose port 22 of your cocalc-docker container to the outside world. To do that you have to create the container with the option -p <your ip address>:2222:22 (say). Thus, instead of doing:
docker run --name=cocalc -d -v ~/cocalc:/projects -p 443:443 sagemathinc/cocalc-docker
do this instead:
docker run --name=cocalc -d -v ~/cocalc:/projects -p 443:443 -p <your ip address>:2222:22 sagemathinc/cocalc-docker
NOTES:
-
You can use a different port instead of port 2222. That's just an arbitrary port that you'll ssh to (i.e., you pass
-p 2222to ssh.) -
If you have an existing cocalc-docker, and just want to expose port 22 without otherwise changing it, that is a massive pain -- that sort of dynamic reconfiguration is just not something that Docker is any good at. Instead, your best bet is to stop and delete that cocalc-docker and create a new one. Hopefully you are using the
-voption, so all your data is stored on your filesystem, rather than in the Docker container!
Then you can do the following, but it won't succeed until you configure the .ssh directory in your project, as explained below:
~$ ssh projectid@<your ip address> -p 2222
Note that projectid is the hex id string for the project without hyphens. One way to show the project id in this format is to open a .term file in your CoCalc project and run the following command:
~$ echo $USER
To use SSH key authentication with the Docker container, have your private key file in the usual place in the host computer, for example ~/.ssh/.id_ed25519, and copy the matching public key into your project's home directory. For example, you could do the following in a .term in your project:
~$ cd
~$ mkdir .ssh
~$ vi .ssh/authorized_keys
... paste in contents of ~/.ssh/id_cocalc.pub from host computer ...
After doing that, you can then ssh to your project. Here's a less abstract example showing what this looks like.
wstein@studio ~ % docker run --name=cocalc-docker -d -p 127.0.0.1:4043:443 -p 127.0.0.1:2022:22 sagemathinc/cocalc-docker
wstein@studio ~ % more ~/.ssh/id_ed25519.pub
ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAIA1bwpB7b7TVIexZxW003FCbDqzyFurSwZlljmT7sWzo wstein@studio
# I then sign in via my web browser and create a project, and make the above key the
# contents of ~/.ssh/authorized_keys inside my project. After doing that, the following works:
wstein@studio ~ % ssh 65b5a3c0e4d046329854f3d3db725f0b@localhost -p 2022
The authenticity of host '[localhost]:2022 ([127.0.0.1]:2022)' [...]
Are you sure you want to continue connecting (yes/no/[fingerprint])? yes [...]
Welcome to Ubuntu 22.04.1 LTS (GNU/Linux 5.10.124-linuxkit aarch64)
[...]
~$ hostname
6456fb9c78a8
~$ echo $USER
65b5a3c0e4d046329854f3d3db725f0b
~$ more .ssh/authorized_keys
ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAIA1bwpB7b7TVIexZxW003FCbDqzyFurSwZlljmT7sWzo wstein@Williams-Mac-Studio.localTo use rsync to copy files to your project, make sure to include the -e 'ssh -p 2222' option (more discussion).
Get a bash shell insider the container, then connect to the database and make a user (me!) an admin as follows:
$ sudo docker exec -it cocalc bash
root@17fecb49c5c2:/# cd /cocalc/src/scripts
root@17fecb49c5c2:/cocalc/src/scripts# ./make-user-admin wstein@gmail.com
UPDATE 1Obviously, you should really make the user you created (with its email address) an admin, not me!Refresh your browser, and then you should see an extra admin tab and the top of your browser window; you can also open any project by directly visiting its URL, and change the idle timeout and always running settings. In the Admin tab you can search for users, impersonate any user, ban users, configure dozens of things about CoCalc, send a notification that all signed in users see, and more. One thing admin users can't do is get a root shell -- for that you have to use sudo docker exec -it cocalc bash (of course, CoCalc is just Ubuntu linux, so you could make it so a specific project can become root via sudo).
Note that the make-user-admin script is in /cocalc/src/scripts. Take a look at it:
root@17fecb49c5c2:/cocalc/src/scripts# more ./make-user-admin
...
echo "update accounts set groups='{admin}' where email_address='$1'" | psqlAs you can see, aside from some error checking, the entire script is just a 1-line PostgreSQL query. If you know basic SQL, you can very easily do all kinds of interesting things.If you type psql as root, you'll get the PostgreSQL shell connected to the database for CoCalc. Type \d to see the tables, and \d tablename for more about a particular table. For example,typing \d accounts shows all the fields in the accounts table, and groups is one of them.Here's where in the source code ofCoCalc itself all of these database tables are defined. In any case, being aware of all this can be very helpfulif you want to do some batch action, e.g., :
- delete all accounts that are old or inactive
- query to get the status of projects or accounts
You can also make it so that running sudo su in a CoCalc terminal allows a project to gain root access. First as above, from outside of CoCalc, dodocker exec -it cocalc bash, then type visudo:
$ docker exec -it cocalc bash
root@931045eda11f:/# visudo
Then run this echo command, but replace 0630f773c01847e79c0863c0118fe0de by the project id with all dashes removed:
echo '0630f773c01847e79c0863c0118fe0de ALL=(ALL) NOPASSWD:ALL' >> /etc/sudoers
Alternatively, you can get the username by typing whoami in a terminal from the project:
~$ whoami
0630f773c01847e79c0863c0118fe0de
After the above echo command, go to a terminal in your project and try sudo su:
~$ sudo su
root@121119037fd4:/projects/0630f773-c018-47e7-9c08-63c0118fe0de# ls /root
run.py
Bam, you're root!
NOTE: A project having sudo access is completely unrelated to a user in CoCAlc being an admin. Neither implies the the other.
NOTE: You can also use visudo to edit the file (which is better), or EDITOR=emacs visudo to edit it using emacs.
WARNING: Obviously having a user able to run as root in a Docker container introduces additional security issues.
Sign in as a user that is an admin (see the previous section above). Click on the Admin tab at the top, search for the user, and then click the "Password" toggle, and click "Request Password Reset Link...".
This does NOT set the password. It just makes a password reset link, which you send your user via some communications channel that works. You may need this because:
- You do not have email setup. It is possible to setup Sendgrid so your cocalc-docker image sends out email, but we haven't documented that yet...
- You have email setup, but it sometimes fails for users with aggressive spam filtering.
After making your main account an admin as above, search for " Registration Tokens" in the Admin tab. Create one or more tokens and share them with people who you want to use your server. Nobody else will be able to make an account.
By default users are NOT allowed to share files publicly, and the server at your_server/share is disabled. You can enable the share server and public file sharing in Admin --> Site Settings --> "Allow public file sharing".
Similar to public file sharing, users are not allowed to make an account without entering an email address and password. You can further restrict users by requiring a registration token or via disabling "Allow email signup" in Admin --> Site Settings. On the other hand, you can allow anybody to use your server without even creating an account by going to Admin --> Site Settings --> "Allow anonymous signup". You probably don't want to do this.
If docker exec -it cocalc bash doesn't seem to give you the right terminal height, e.g. content is only displayed in the uppper part of the terminal, this workaround may help when launching bash:
docker exec -e COLUMNS="`tput cols`" -e LINES="`tput lines`" -it cocalc bash
More information on this issue is in moby issue 33794.
In order to build and run CoCalc on an SELinux box, first set SELinux to permissive:
$ setenforce 0
Tell docker and SELinux to "play nicely":
$ chcon -Rt svirt_sandbox_file_t cocalc
return SELinux to enabled:
$ setenforce 1
-- via discussion
See the docs directory for some additional documentation.
If you started the container as above, there will be a directory ~/cocalc on your host computer that contains all data and files related to your projects and users -- go ahead and verify that it is there before upgrading. It might look like this:
Williams-MacBook-Pro:~ wstein$ ls cocalc
be889c14-dc96-4538-989b-4117ffe84148 postgres conf
The directory postgres contains the database files, so all projects, users, file editing history, etc. The directory conf contains some secrets and log files. There will also be one directory (like be889c14-dc96-4538-989b-4117ffe84148) for each project that is created.
CoCalc-Docker includes a PostgreSQL server. However, you can also use your own external PostgreSQL server.
WARNING (APRIL 2023): The sagemathinc/cocalc Docker image is currently not backward compatible with sagemathinc/cocalc-docker, since sagemathinc/cocalc uses the deprecated PostgreSQL version 10, whereas cocalc-docker uses PostgreSQL version 14, and we haven't implemented an automated upgrade path yet. Here are user contributed instructions for manually upgrading the database, though.
New images are released regularly, as you can see on the SageMath, Inc. Dockerhub page.
To get the newest image, do this (which will take some time):
docker pull sagemathinc/cocalc-docker
Once done, you can delete and recreate your CoCalc container: (This will not delete any of your project or user data, which you confirmed above is in ~/cocalc.)
docker stop cocalc
docker rm cocalc
docker run --name=cocalc -d -v ~/cocalc:/projects -p 443:443 sagemathinc/cocalc-docker
Now visit https://localhost to see your upgraded server.
Instead of upgrading the Docker image, you could upgrade the source code of cocalc to the latest master version (or any other commit or branch) as follows.
First become root in your container: docker exec -it cocalc bash, then:
root@...:~# umask 022
root@...:~# cd /cocalc/src
root@...:/cocalc/src# git pull
root@...:/cocalc/src# npm run build
root@...:/cocalc/src# cd /root; python -c "import run; run.start_hub()"This could take about 15 minutes. It could randomly fail if some npm package is temporarily not available; if that happens, try again. Upgrading this way does not upgrade any system-wide Ubuntu packages or configuration, so it might result in a broken Docker container. In that case, your data should be fine, and you can upgrade as described in the section above.
There is a minimal version of cocalc-docker that is built to run in personal mode, in which absolutely everything in the container runs as a single non-root user named user.
The CoCalc Docker images at Docker Hub contain a subset of all the software in at cocalc.com. At present, the images are about 12 GB while the cloud service has hundreds of GB of packages and libraries.
Suppose you'd like to add software to your local CoCalc instance after installing and starting the Docker container. Here's an example of how to add an install of texlive-full, in case you need more than the minimal texlive installation in the published image:
The Docker image is Ubuntu 18.04. You can do
sudo docker exec -it [container name] bash
to become root in the container, then do
apt-get install texlive-full
to install the package.
Note that the texlive-full package is over 3 GB. So you will need the additional disk space to install it, and it could take several minutes to over an hour to install, depending on your connection to the internet and the speed of your computer.
Additional notes:
- Be sure to type
umask 022first before you install software if you are using a method other thanapt-get. This step is needed to ensure that permissions are set properly. The default umask is 007. If you usepip3orpip2without setting the umask to 022, the package gets installed, but it is not visible to normal users as a result. - Most instructions for adding packages to Ubuntu 18.04 should work for CoCalc-Docker, for example
pip installfor Python 2 packages, andpip3 installfor Python 3 packages. - Whenever you upgrade your CoCalc image from Docker Hub as described in Upgrade above, you will need to repeat the above steps.
- See the contrib subdirectory here for scripts contributing by people that are running their own cocalc-docker servers. For example, there is a Python script for removing old accounts and old or deleted projects.
The cocalc-docker images are not some black box images that are built in some mysterious way. You can see exactly what recipe is used to build them by looking at Dockerfile. Moreover, you can modify Dockerfile if you want and build your own image.
It is easy (but time consuming) to build the cocalc-docker image from scratch. We do this regularly using the update-the-build-stage-0.sh scripts, e.g., for x86_64:
$ make cocalc-dockerSome reasons to build your own image:
- You want to run the most up-to-date version of the cocalc source code, or your own special branch. Though you can also easily do that starting with our premade image, as explained in the section on upgrading.
- You want to change what software is installed in your cocalc-docker image, e.g., remove Julia and add something else like Tensorflow that we don't include by default.
Building the image no longer rebuilds sage, since that would add hour(s) to the build. Instead, we use a build of Sage from here, so the build of cocalc-docker is much quicker (it's mostly just installing binaries and building Cocalc).
This section assuming that your docker host has a GPUs and the nvidia-docker2 runtime is installed properly. For more information please see the NVIDIA Container Toolkit project.
Test of the docker with GPUs support should give a similar output:
(base) [root@gput401 cocalc-docker]# docker run --runtime=nvidia -e NVIDIA_VISIBLE_DEVICES=all --rm nvidia/cuda:latest nvidia-smi
Tue Jun 16 17:52:16 2020
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 440.64.00 Driver Version: 440.64.00 CUDA Version: 10.2 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
|===============================+======================+======================|
| 0 Tesla T4 Off | 00000000:00:06.0 Off | 0 |
| N/A 31C P0 16W / 70W | 0MiB / 15109MiB | 0% Default |
+-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+
| Processes: GPU Memory |
| GPU PID Type Process name Usage |
|=============================================================================|
| No running processes found |
+-----------------------------------------------------------------------------+
If the test running w/o problems, you can start to rebuild your own cocalc docker image with tensorflow or pytorch support:
cp Dockerfile Dockerfile-gpu
Simply change first line in the Dockerfile-gpu
#ARG MYAPP_IMAGE=ubuntu:22.04
ARG MYAPP_IMAGE=nvcr.io/nvidia/tensorflow:23.08-tf2-py3
to instead start with the official Tensorflow image. For the official PyTorch image, use:
#ARG MYAPP_IMAGE=ubuntu:22.04
ARG MYAPP_IMAGE=nvcr.io/nvidia/pytorch:23.10-py3
Then rebuild your image:
docker build -t cocalc-gpu -f Dockerfile-gpu .
Run it with GPU support.
docker run -it --gpus all --name=cocalc-gpu -d -v ~/cocalc_test:/projects -p 443:443 -p 0.0.0.0:2222:22 --rm cocalc-gpu bash