For AI agents: The documentation index is at https://docs.digitalocean.com/llms.txt. Markdown versions of pages use the same URL with index.html.md in place of the HTML page (for example, append index.html.md to the directory path instead of opening the HTML document).

AMD Primus

This 1-Click image ships AMD Primus, AMD’s high-performance training framework for large-scale foundation-model training on AMD Instinct GPUs (MI300X). It bakes in the AMD-published Primus training container (rocm/primus:v26.4) with the full ROCm stack (PyTorch, TransformerEngine, AITER) and the Megatron-LM and TorchTitan backends, plus the Primus repository with submodules pinned to a release tag.

Everything is baked into the snapshot at build time, so the droplet is instant-on: no multi-gigabyte image pull or submodule clone at first use.

Two ways to use it

1. JupyterLab notebook on port 80 (fresh per-droplet token), with a guided quickstart notebook.
2. primus-cli inside the container (docker exec -it rocm /bin/bash) for a real Llama-2 7B mock-data pretraining example (single-GPU-ready).

See the Primus documentation for backends, configuration, and multi-node training.

Software Included

Deploying this Offering using the Control Panel

Click the Deploy to DigitalOcean button to deploy this offering. If you aren’t logged in, this link will prompt you to log in with your DigitalOcean account.

[](https://cloud.digitalocean.com/gpus/new?appId=4e4e7343039010cd43c436c6&image=AMD Primus™ v26.4 on Ubuntu 24.04&type=applications)

Getting Started After Deploying AMD Primus™

After the droplet boots, open http://YOUR_DROPLET_IP in a browser; the JupyterLab token is printed in the SSH login MOTD (and at /root/.digitalocean_passwords). Open 00-primus-quickstart.ipynb for a guided run. This droplet has one MI300X, so pin GPUS_PER_NODE=1: SSH in, run docker exec -it rocm /bin/bash, then from /workspace/Primus launch a single-GPU Llama-2 7B mock-data pretrain with ./primus-cli direct --env GPUS_PER_NODE=1 --env PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True -- train pretrain --config <single-GPU config> (the notebook builds the config for you). On an 8x MI300X droplet, drop GPUS_PER_NODE=1 and use the stock configs as-is.