Gemma 3

Deploy Gemma 3 from Google DeepMind on Spheron GPU instances using vLLM. Gemma 3 is released under the Gemma Terms of Use, which permits commercial use and modification after accepting Google's license agreement.

Recommended hardware

Model	Recommended GPU	Instance Type	Notes
Gemma-3-4B-IT	Any 8GB GPU	Spot	Lightweight
Gemma-3-12B-IT	RTX 4090 (24GB)	Dedicated or Spot	Full precision
Gemma-3-27B-IT	A100 80GB	Dedicated	Full precision
Gemma-3-27B-IT (INT4)	RTX 4090 (24GB)	Dedicated or Spot	AWQ 4-bit

Manual setup

Use these steps to set up the server manually after SSH-ing into your instance. This works on any provider regardless of cloud-init support.

Step 1: Connect to your instance

ssh <user>@<ipAddress>

Replace <user> with the username shown in the instance details panel (e.g., ubuntu for Spheron AI instances) and <ipAddress> with your instance's public IP.

Step 2: Install vLLM

sudo apt-get update -y
sudo apt-get install -y python3-pip
pip install vllm

Step 3: Start the server

Run the server in the foreground to verify it works:

HF_TOKEN=<your-hf-token> python3 -m vllm.entrypoints.openai.api_server \
  --model google/gemma-3-12b-it \
  --port 8000 \
  --dtype bfloat16

Press Ctrl+C to stop. Replace <your-hf-token> with your HuggingFace token. For Gemma-3-27B on A100 80GB, replace the model with google/gemma-3-27b-it.

Step 4: Run as a background service

To keep the server running after you close your SSH session, create a restricted token file and a systemd service:

# Store the token in a file readable only by root
sudo mkdir -p /etc/vllm
sudo install -m 600 /dev/null /etc/vllm/hf-token
echo "HF_TOKEN=<your-hf-token>" | sudo tee -a /etc/vllm/hf-token > /dev/null
 
sudo tee /etc/systemd/system/vllm-gemma3.service > /dev/null << 'EOF'
[Unit]
Description=Gemma 3 vLLM Inference Server
After=network.target
 
[Service]
Type=simple
EnvironmentFile=/etc/vllm/hf-token
ExecStart=/usr/bin/python3 -m vllm.entrypoints.openai.api_server \
  --model google/gemma-3-12b-it \
  --port 8000 \
  --dtype bfloat16
Restart=on-failure
RestartSec=10
 
[Install]
WantedBy=multi-user.target
EOF
 
sudo systemctl daemon-reload
sudo systemctl enable vllm-gemma3
sudo systemctl start vllm-gemma3

Replace <your-hf-token> with your HuggingFace token. Using EnvironmentFile= with chmod 600 prevents other local users from reading the token via systemctl show.

Accessing the server

SSH tunnel

ssh -L 8000:localhost:8000 <user>@<ipAddress>

Usage example

from openai import OpenAI
 
client = OpenAI(base_url="http://localhost:8000/v1", api_key="not-needed")
 
response = client.chat.completions.create(
    model="google/gemma-3-12b-it",
    messages=[{"role": "user", "content": "Explain the Apache 2.0 license in one paragraph."}],
)
print(response.choices[0].message.content)

Cloud-init startup script (optional)

If your provider supports cloud-init, you can paste this into the Startup Script field when deploying to automate the setup above.

Gemma-3-12B (RTX 4090)

#cloud-config
write_files:
  - path: /etc/systemd/system/vllm-gemma3.service
    content: |
      [Unit]
      Description=Gemma 3 vLLM Inference Server
      After=network.target
 
      [Service]
      Type=simple
      EnvironmentFile=/etc/vllm/hf-token
      ExecStart=/usr/bin/python3 -m vllm.entrypoints.openai.api_server \
        --model google/gemma-3-12b-it \
        --port 8000 \
        --dtype bfloat16
      Restart=on-failure
      RestartSec=10
 
      [Install]
      WantedBy=multi-user.target
runcmd:
  - apt-get update -y
  - apt-get install -y python3-pip
  - pip install vllm
  - mkdir -p /etc/vllm
  - install -m 600 /dev/null /etc/vllm/hf-token
  - echo "HF_TOKEN=<your-hf-token>" >> /etc/vllm/hf-token
  - systemctl daemon-reload
  - systemctl enable vllm-gemma3
  - systemctl start vllm-gemma3

Replace <your-hf-token> with your HuggingFace token.

Gemma-3-27B (A100 80GB)

Replace the model in ExecStart with google/gemma-3-27b-it.

What's next

vLLM Inference Server: vLLM configuration details
Phi-4 & Phi-4 Multimodal: Another efficient small model option
Instance Types: GPU selection for small models
Cost Optimization: Spot instances for Gemma 3