Cosmos predict 2.5 on dgx spark

Omniverse Learning and Resources Tutorials & Guides
1

I have dgx spark and like to run cosmos predict 2.5 inference.
Is there an example of how to do this?

Is there a docker container with cosmos predict 2.5 installed?

Thanks,

2

You can run Cosmos Predict 2.5 inference on DGX Spark either by using the official Docker workflow from the Cosmos docs or by cloning the GitHub repo and running the provided inference script; there is no “DGX‑Spark‑only” image, but you build and run a standard Cosmos Predict 2.5 container on Spark. docs.nvidia

Official docs and examples

  • NVIDIA provides an inference guide in the Cosmos Predict 2.5 repo at docs/inference.md, which walks through running base models for text, image, and video inputs. github
  • The main product documentation site has a Predict2.5 section with links to installation, setup, and usage instructions. docs.nvidia
  • The Cosmos Cookbook site includes quick‑start inference recipes you can adapt to DGX Spark (scripts, recommended flags, and example prompts). nvidia-cosmos.github

Docker container for Predict 2.5

The standard workflow is to build and run the Cosmos Predict 2.5 container yourself, rather than pulling a prebuilt NVIDIA NGC image specifically for Spark. github

Typical steps from the installation docs: docs.nvidia

  1. Clone the repo on your DGX Spark:
    • git clone https://github.com/nvidia-cosmos/cosmos-predict2.5.git (path from docs). github
  2. From the repo root, build the image (note the --ulimit flag the docs recommend):
    • docker build --ulimit nofile=131071:131071 -f Dockerfile . -t cosmos-predict-2.5 docs.nvidia
  3. Run the container with GPU access:
    • docker run --gpus all --rm -v .:/workspace -v /workspace/.venv -it cosmos-predict-2.5 docs.nvidia

Inside the container you then follow the repo’s setup and inference guides (environment variables, model download, and the command‑line interface). github

Running inference (high level)

  • The inference guide shows how to: github
    • Set up prerequisites and environment variables (e.g., model cache via HF_HOME). github
    • Choose a Predict 2.5 checkpoint (base vs domain‑specific). github
    • Run example commands that take text, image, or video as input and generate future‑frame video trajectories. huggingface
  • A beginner tutorial video also walks step‑by‑step through installing Cosmos Predict 2.5, downloading weights, and running first predictions; this is applicable on DGX Spark as long as Docker and drivers are set up. youtube

DGX Spark specifics

  • DGX Spark is fully capable of running the standard Cosmos Predict 2.5 stack; you do not need a custom image, only a correct CUDA/driver + Docker + NVIDIA Container Toolkit setup as described in your Spark HOW‑TO/inference stack docs. forums.developer.nvidia
  • You can mount your Spark storage (datasets, output directory) into /workspace when launching the container to keep I/O local to the DGX. docs.nvidia

Would you prefer a concrete example command line (with typical flags and volumes) tailored to a single DGX Spark node with, say, 4 or 8 GPUs, or are you planning to run multi‑node distributed inference?

3

@Richard3D

Thanks for detailed instructions and I will try it later.
Do you have any idea about performance of predict 2.5 2B on dgx spark in terms of speed?

Thanks,

4

No sorry. I do not have any specific metrics like that.