Skip to content

Get Started with Cosmos Reason1 on Brev: Inference and Post-Training

Author: Saurav Nanda Organization: NVIDIA

This guide walks you through setting up NVIDIA Cosmos Reason1 on a Brev H100 GPU instance for both inference and post-training workflows. Brev provides on-demand cloud GPUs with pre-configured environments, making it easy to get started with Cosmos models.

Overview

Brev.dev is a cloud GPU platform that provides instant access to high-performance GPUs like the H100. This guide will help you:

  1. Set up a Brev instance with H100 GPU
  2. Configure the environment for Cosmos Reason1
  3. Run inference on the Reason1 model
  4. Perform post-training (SFT) on custom datasets

Prerequisites

The cheat code: Launchables

Launchables are an easy way to bundle a hardware and software environment into an easily shareable link. Once you've dialed in your Cosmos setup, a Launchable is the most convenient way to save time and share your configuration with others.

Note: Cosmos and Brev are evolving. You may encounter minor UI and other differences in the steps below over time.

Step 1: Create a Brev Instance

  1. Log in to your Brev account
  2. Find the Launchable section of the Brev website. Launchables Menu
  3. Click the Create Launchable button. Create Launchable Button

  4. Enter the Cosmos Reason GitHub URL https://github.com/nvidia-cosmos/cosmos-reason1

  5. Choose H100 GPU instance with 80GB VRAM. H100 Instance

  6. If you don't need Jupyter, remove it. Tell Brev which ports to open if you plan to setup some other custom server(if any). Jupyter

  7. Add a setup script for Cosmos Reason. See sample setup script for an example. Setup Script

  8. Name your Launchable and configure access.(usually takes 2-3 minutes). create

Step 2: Connect to Your Instance

Once your instance is ready, Brev will provide SSH connection details. instance

Option 1: Open Jupyter Notebook

Notebook

Option 2: Copy the SSH command from your Brev dashboard

brev login --token <YOUR_TOKEN>

Open a terminal locally

brev shell sample-reason1-fa3124

OR Open in Code Editor

brev open sample-reason1-fa3124 cursor

Step 3: Authenticate Hugging Face CLI

The Hugging Face Token is required to download the Cosmos Reason1 model:

# Authenticate with Hugging Face
~/.local/bin/hf auth login

When prompted, enter your Hugging Face token. You can create a token at https://huggingface.co/settings/tokens.

Important: Make sure you have access to the Cosmos-Reason1-7B model. Request access if needed.

Step 4: Run Inference and Post-Training

Now you're ready to run inference with Cosmos Reason1!

Follow the steps provided in the Cosmos Reason GitHub repo to run the inference and post-training examples.

Troubleshooting

Model Download Issues

If the model fails to download:

  1. Verify your Hugging Face authentication: ~/.local/bin/hf whoami
  2. Ensure you have access to the Cosmos-Reason1-7B model
  3. Check your internet connection
  4. Try downloading manually: huggingface-cli download nvidia/Cosmos-Reason1-7B

SSH Connection Issues

If you lose SSH connection:

  1. Brev instances may pause after inactivity
  2. Check your Brev dashboard for instance status
  3. Restart the instance if needed
  4. Reconnect using the SSH command

Resource Management

Stopping Your Instance

To avoid unnecessary charges:

  1. Go to your Brev dashboard
  2. Select your instance
  3. Click "Stop" or "Delete" when done

Saving Your Work

Before stopping your instance:

# Save model checkpoints to cloud storage (e.g., S3, GCS)
# Or download them to your local machine
scp -r ubuntu@<your-instance-ip>:~/cosmos-reason1/examples/post_training_hf/outputs ./local-outputs

Additional Resources

Support

For issues related to: