Skip to content

creating_custom_brev.md

Latest commit

 

History

History
53 lines (43 loc) · 4.78 KB

creating_custom_brev.md

File metadata and controls

53 lines (43 loc) · 4.78 KB

Creating a custom GPU instance using Brev

The goal of this guide is to walk through the steps to create a custom GPU instance on NVIDIA Brev using the Brev Launchable interface.

Brev provides access to a wide variety of GPU instances on difference cloud environment. While this provides you flexibility on pricing and avaiablity, it makes it difficult to test this launchable on every combination of compute environment and cloud environment. As such, some tinkering may be required to get the instance to work on the specific configuration you select. To use one of our recommended environments that is tested to work, please check out the Quickstart Instructions.

Warning

Our Docker Compose YAML files have a known issue when using GCP resources on Brev. We recommend using another cloud provider until this is resolved.

Option 1 - Create a custom instance using the Brev Launchable Creator

  1. Go to Brev’s Launchable Creator (requires account)
  2. When asked How would you like to provide your code files? select "I have code files in a GitHub repository". Enter the URL of this repository.
  3. When asked What type of runtime environment do you need? select "With container(s)"
  4. Under Choose a Container Configuration select "Docker Compose" and click on the toggle to select "I have an existing docker-compose.yaml file".
  5. Under Upload Docker Compose select "Provide GitHub/Gitlab URL" and provide a link to one of the Docker Compose YAML files in the docker/brev directory. There is a README.md in that directory with instructions on which YAML to select. Please take note that Blackwell GPUs require CUDA13 while other GPU instances usually will work on CUDA12. Note, you need to pass a link to the file in GitHub, not to the raw.github.com file (e.g. docker-compose-nb-2602-cuda13.yaml). Click "Validate".
  6. On the next page, when asked Do you want a Jupyter Notebook experience? select "No, I don't want Jupyter (Not Recommended)". We will provide Jupyter in the Docker compose already.
  7. In the section title Do you need to expose services? make sure that ports 8888, 8787, and 8786 are open. Name port 8888 jupyter so Brev can treat it as a jupyterlab based instance and provide an Open Notebook button.
  8. Select your desired compute environment. Make sure you select sufficient disk size to download the datasets you want to work with. We recommend at least 128GB, unless you want to run the 11M cell notebook, which you will need atleast 200GB. Note, you will not be able to resize the instance once created.
  9. Create a name for your launchable, and deploy.

Note

Git is not installed by default in this container, but it can be installed using

apt update
apt install git -y

Option 2 - Install rapids-singlecell manually in a rapids/notebook container

A second option to set up rapids-singlecell is to start by Creating a Custom RAPIDS Brev Instance, manually git cloning this single-cell-analysis-blueprint repository, and manually installing the scverse rapids-singlecell library.

  1. Follow the instructions for Creating a RAPIDS Brev Instance. Remember the RAPIDS version for Step 3. Launch the instance.
  2. Download this repository into the instance.
  3. Download the rapids-singlecell repository into the instance. Follow their installation instructions. It is important to have the same RAPIDS version and CUDA Toolkit version as your installed RAPIDS versions. For Brev, we currently default to CUDA Toolkit 13 to support Blackwell GPUs and CUDA Toolkit 12 for other GPUs. Therefore, you can install rapids-singelcell using `mamba env create -f conda/rsc_.yml'
    • Example to install rapids-singelcell for 26.02
    mamba env create -f conda/rsc_rapids_26.02.yml

You can check this by going to the command line and running nvcc --version inside the instance. You'll get an output like this:

```bash
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2025 NVIDIA Corporation
Built on Fri_Feb_21_20:26:18_PST_2025
Cuda compilation tools, release 12.8, V12.8.93
Build cuda_12.8.r12.8/compiler.35583870_0
```
  1. Activate the mamba environment and run the notebooks/your code. 
    mamba activate rapids-singelcell