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).

vLLM 0.27.1 and vLLM-Omni on AMD ROCm 7.14

This 1-Click image ships two serving engines side by side on an AMD ROCm 7.14 host stack running Ubuntu 24.04, preconfigured for AMD Instinct GPUs.

vLLM 0.27.1 is a high-throughput, memory-efficient inference and serving engine for large language models, exposing an OpenAI-compatible API. It delivers state-of-the-art serving throughput with PagedAttention and continuous batching.

vLLM-Omni 0.27.0rc1 extends that engine to omni-modal models, covering text, audio, image and video, and ships its own vllm-omni command line interface. It is built from source against ROCm inside this image and is version-paired with vLLM 0.27.1.

Both engines live inside a single Docker container named rocm, alongside a JupyterLab environment with example notebooks. The login banner prints the JupyterLab URL and token plus ready-to-paste serve commands for both engines.

A note on ROCm versions, which are deliberately different and are not a defect: the host runs ROCm 7.14 for the driver and amd-smi, while the vLLM container carries ROCm 7.2.3, because upstream vLLM does not publish a ROCm 7.14 container image.

This image was validated on a live AMD Instinct MI300X. Verified on that node: a real vLLM completion served over the OpenAI-compatible API, a successful vllm_omni import reporting a ROCm platform, the vllm-omni CLI present, and a real GPU matrix multiplication.

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=341959c4771a11d5799358b6&image=vLLM 0.27.1 + vLLM-Omni (ROCm 7.14) 0.27.1 on Ubuntu 24.04&type=applications)

Getting Started After Deploying vLLM 0.27.1 + vLLM-Omni (ROCm 7.14)

After the droplet boots, SSH in as root. The login MOTD prints the JupyterLab URL and token, plus copy-paste commands for both engines.

Serve a standard LLM with vLLM on port 8000:

docker exec -d rocm vllm serve Qwen/Qwen3-8B --port 8000
curl http://localhost:8000/v1/chat/completions -H 'Content-Type: application/json' \
  -d '{"model":"Qwen/Qwen3-8B","messages":[{"role":"user","content":"Hello"}]}'

Serve an omni-modal model with vLLM-Omni on port 8091. Note that vLLM-Omni has its own CLI, vllm-omni, and is not a flag on vllm:

docker exec -d rocm vllm-omni serve Qwen/Qwen3-Omni-30B-A3B-Instruct --port 8091

Gated or large weights need a Hugging Face token: export HF_TOKEN=....

Open an interactive shell with docker exec -it rocm /bin/bash. Inspect the GPU stack with amd-smi monitor and rocminfo.

UFW is enabled and only ports 22, 80 and 443 are open. Ports 8000 and 8091 are reachable from inside the droplet. Open them deliberately with ufw, or tunnel: ssh -L 8000:localhost:8000 root@YOUR_DROPLET_IP.