Skip to content

Latest commit

 

History

History
50 lines (33 loc) · 3.56 KB

run.md

File metadata and controls

50 lines (33 loc) · 3.56 KB
Raw

Running Containers

Let's say that you found a container image from the Package List or DockerHub, or built your own container - the normal way to run an interactive Docker container on your Jetson would be using docker run like this:

$ sudo docker run --runtime nvidia -it --rm --network=host CONTAINER:TAG

That's actually a rather minimal command, and doesn't have support for displays or other devices, and it doesn't mount the model/data cache (/data). Once you add everything in, it can get to be a lot to specify by hand. Hence, we have some helpers that provide shortcuts.

The jetson-containers run launcher can be run from any directory and forwards its command-line to docker run, with some added defaults - including the above flags, mounting the /data cache, and mounting various devices for display, audio, and video (like V4L2 and CSI cameras)

$ jetson-containers run CONTAINER:TAG                   # run with --runtime=nvidia, default mounts, ect
$ jetson-containers run CONTAINER:TAG my_app --abc xyz  # run a command (instead of interactive mode)
$ jetson-containers run --volume /path/on/host:/path/in/container CONTAINER:TAG  # mount a directory

The flags and arguments to jetson-containers run are the same as they are to docker run - anything you specify will be passed along.

autotag

To solve the issue of finding a container with package(s) that you want and that's compatible with your version of JetPack/L4T, there's the autotag tool. It locates a container image for you - either locally, pulled from a registry, or built from source:

$ jetson-containers run $(autotag pytorch)   # find pytorch container to run for your version of JetPack/L4T

What's happening here with the $(autotag xyz) syntax, is that Bash command substitution expands the full container image name and forwards it to the docker run command. For example, if you do echo $(autotag pytorch) it would print out something like dustynv/pytorch:r35.2.1 (assuming that you don't already have the pytorch image locally).

You can of course use autotag interspersed along with other command-line arguments to launch the container:

$ jetson-containers run $(autotag pytorch) my_app --abc xyz  # run a command (instead of interactive mode)
$ jetson-containers run --volume /path/on/host:/path/in/container $(autotag pytorch)  # mount a directory

Or with using docker run directly:

$ sudo docker run --runtime nvidia -it --rm --network=host $(./autotag pytorch)

This is the order in which autotag searches for container images:

  1. Local images (found under docker images)
  2. Pulled from registry (by default hub.docker.com/u/dustynv)
  3. Build it from source (it'll ask for confirmation first)

When searching for images, it knows to find containers that are compatible with your version of JetPack-L4T. For example, if you're on JetPack 4.6.x (L4T R32.7.x), you can run images that were built for other versions of JetPack 4.6. Or if you're on JetPack 5.1 (L4T R35), you can run images built for other versions of JetPack 5.1 (and likewise for JetPack 6.0 and newer)