Brev fallback setup¶
Use this page only when a local laptop is not enough:
- you do not have access to an NVIDIA GPU for the CUDA run;
- your laptop cannot install the workshop environments in time;
- the room network is struggling and instructors ask the group to switch to prepared cloud machines.
If your laptop can run the setup in Before you start, stay local. That is the default path for the workshop. Brev is a fallback, not an extra prerequisite.
Wait for the instructor coupon
Do not create a Brev instance yet. Creating cloud GPU machines can use paid credits, and the workshop machines should be created with the coupon from the instructors. Wait until the instructors share the coupon and tell you which instance type to use.
1. Install the Brev CLI¶
Brev is the NVIDIA cloud development environment we can use for the fallback machines. Install its CLI with Pixi:
Windows users
Brev supports Windows through WSL. Run the Brev CLI commands from your WSL Ubuntu terminal, not from PowerShell.
2. Create or log in to your Brev account¶
Create an account at https://login.brev.nvidia.com/signin, then log in from your terminal:
Wait for the coupon
Log in now, but do not create an instance yet. Wait until the instructors share the Brev coupon.
The CLI asks for your email address and opens a browser for authentication. When login succeeds, you should see your Brev account in the browser.
Select the organization for the workshop:
3. Create the fallback instance¶
From the root of this workshop repository, run:
brev create "$(whoami)-roscon-pixi" \
--startup-script @docs/code/brev/setup_brev.sh \
--type massedcompute_L40S
The startup script installs Pixi, clones this repository onto the instance, and pre-downloads both ROS environments with PyTorch. On a compatible NVIDIA GPU machine, the preferred CUDA platform selects the GPU build. That means you can keep working even if the local room Wi-Fi is slow.
If you are not inside a fresh clone of the workshop repository, download the script first:
curl -sLO https://raw.githubusercontent.com/prefix-dev/roscon-2026-declarative-ros-workspaces-with-pixi/main/docs/code/brev/setup_brev.sh
brev create "$(whoami)-roscon-pixi" \
--startup-script @./setup_brev.sh \
--type massedcompute_L40S
Choosing the instance type
Use the instance type announced by the instructors. For CUDA exercises it must have an NVIDIA GPU and a recent driver. If we only need a bandwidth fallback, a CPU instance may be enough.
4. Connect to the instance¶
Open it in VS Code:
Or use a terminal:
The repository is cloned here:
Check Pixi and the pre-warmed environment:
For Exercise 1's CUDA run, the instance must have an NVIDIA GPU and a compatible driver. Run this inside the Brev terminal, from the repository root:
nvidia-smi
pixi run --manifest-path solutions/01-ros-workspace/pixi.toml --platform cuda-linux-64 cuda-check
Expect the GPU name, compute capability and GPU result: 8.0.
This executes a tensor operation on CUDA and exits; it fails rather than falling back to CPU.
If it fails, ask an instructor to check the instance type, driver and selected platform.
Then try the ROS node on the same device:
pixi run --manifest-path solutions/01-ros-workspace/pixi.toml --platform cuda-linux-64 build
pixi run --manifest-path solutions/01-ros-workspace/pixi.toml --platform cuda-linux-64 brain
Look for thinking on: cuda, then stop it with Ctrl+C.
The node publishes without a turtlesim window, so this works over SSH.
Return to Exercise 1 for the explanation and Jetson commands.
5. Know the limits¶
A Brev shell is excellent for dependency solving, CUDA checks and non-graphical ROS commands.
Graphical apps such as turtlesim are still easiest on your own laptop.
If the room switches fully to Brev, follow the instructor's live guidance for any GUI steps.
6. Clean up¶
When you are done, delete the instance so it stops using credits: