Fixing Radeon Cloud SSH Access

Fixing Radeon Cloud SSH Access | Generated by AI

Home 2026.07

Question: Why switching to Jupyter image and installing SSH server fixed access on Radeon Cloud?

Answer:

Yes, your understanding is basically correct. On Radeon Cloud, different container images are not just different packages — they define the whole runtime environment.

The flow is roughly:

Radeon Cloud
      |
      v
GPU VM / container
      |
      +-- selected image
              |
              +-- JupyterLab?
              +-- SSH server?
              +-- vLLM?
              +-- model files?
              +-- tools?

The vLLM image is optimized for serving models, not necessarily interactive development. It may start:

vllm serve ...

but not include:

jupyter lab
sshd
terminal UI

The Jupyter image is more like a development workstation:

JupyterLab
 ├── Notebook
 ├── Terminal
 ├── Python environment
 ├── ROCm tools
 └── SSH (if installed)

The Radeon Cloud examples also use Jupyter as the entry point for vLLM experiments: for example, their Hermes + vLLM workshop instructs users to open a Jupyter terminal and start the vLLM server there. (Radeon Cloud)

So your workaround:

switch image -> Jupyter notebook
        |
        v
apt install openssh-server
        |
        v
ssh into GPU machine

is a normal cloud workflow.

A typical setup:

sudo apt update
sudo apt install -y openssh-server

sudo mkdir -p /run/sshd

sudo service ssh start

ss -tlnp | grep :22

Then from your laptop:

ssh user@<gpu-instance-ip>

After SSH works, you can run:

tmux new -s vllm

vllm serve \
  Qwen/Qwen3-8B \
  --host 0.0.0.0 \
  --port 8000

The architecture becomes:

Laptop
  |
  | ssh
  v
Radeon Cloud GPU instance
  |
  +-- tmux
        |
        +-- vLLM server :8000
        |
        +-- ROCm
        |
        +-- AMD GPU

For hackathon work, I would usually choose:

Jupyter/Base image → exploration, coding, debugging, SSH, agents
vLLM image → final inference deployment/demo

because during development you need shell access much more often than a preconfigured inference server. Radeon Cloud provides both notebook-style environments and vLLM-focused templates. (Radeon Cloud)

One thing to check: after installing SSH, make sure the container keeps the process alive. Some cloud containers restart when the main process exits. Running sshd under the platform’s expected startup mechanism is safer than manually launching it.

Back Donate