Source: llm/gemma3

Run and Serve Gemma3 with SkyPilot and vLLM#

SkyPilot is a framework for running AI and batch workloads on any infra, offering unified execution, high cost savings, and high GPU availability.

On Mar 12, 2025, Google released Gemma 3, a collection of lightweight, state-of-the-art open models built from the same research and technology that powers Google’s Gemini 2.0 models.

Gemma 3 is available in four sizes (1B, 4B, 12B, and 27B) and introduces these key capabilities:

  • Multimodal capabilities: Support for vision-language input with text output (1B, 4B, 12B, and 27B models)

  • 128K token context window: 16x larger input context for analyzing more complex data

  • Multilingual support: Over 140 languages supported across all model sizes

  • Advanced features: Function calling, structured output, and improved reasoning capabilities

  • Optimized for devices: Can run on a range of hardware from single GPUs to larger clusters

In preliminary evaluations, Gemma 3 outperforms many larger models in its size class, making it ideal for a wide range of applications from personal projects to production systems.

This guide walks through how to run and host Gemma 3 models on any infrastructure ranging from local GPU workstations, Kubernetes clusters, and public clouds (15+ clouds supported).

SkyPilot supports a variety of LLM frameworks and models. In this guide, we use vLLM, an open-source library for fast LLM inference and serving, as an example.

Note: The provided YAML configuration is set up for the Gemma 3 4B instruction-tuned model, but can be easily modified for other Gemma 3 variants.

Step 0: Setup#

Install SkyPilot on your local machine:

pip install 'skypilot-nightly[all]'

To set up and verify your infra, run

sky check

See docs for details.

You also need to set up your huggingface account and the Gemma3 usage agreement.

Step 1: Run it with SkyPilot#

Now it’s time to run deepseek with SkyPilot. The instruction can be dependent on your existing hardware.

sky launch gemma3.yaml \
  -c gemma-3 \
  --env MODEL_NAME=google/gemma-3-4b-it \
  --gpus L4:1
  --env HF_TOKEN=xxxx

Step 2: Get Results#

Get a single endpoint that load-balances across replicas:

export ENDPOINT=$(sky status --ip gemma-3)

Query the endpoint in a terminal:

curl http://$ENDPOINT:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "google/gemma-3-4b-it",
    "messages": [
      {
        "role": "system",
        "content": "You are a helpful assistant."
      },
      {
        "role": "user",
        "content": "Who are you?"
      }
    ]
  }' | jq .

Get response

{
  "id": "chatcmpl-525ee12bfe3a456b9e764639724e1095",
  "object": "chat.completion",
  "created": 1741805203,
  "model": "google/gemma-3-4b-it",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "reasoning_content": null,
        "content": "I'm Gemma, a large language model created by the Gemma team at Google DeepMind. I’m an open-weights model, which means I’m widely available for public use! \n\nI can take text and images as inputs and generate text-based responses. \n\nIt’s nice to meet you!",
        "tool_calls": []
      },
      "logprobs": null,
      "finish_reason": "stop",
      "stop_reason": 106
    }
  ],
  "usage": {
    "prompt_tokens": 20,
    "total_tokens": 87,
    "completion_tokens": 67,
    "prompt_tokens_details": null
  },
  "prompt_logprobs": null
}

Deploy the Service with Multiple Replicas#

The launching command above only starts a single replica(with 1 node) for the service. SkyServe helps deploy the service with multiple replicas with out-of-the-box load balancing, autoscaling and automatic recovery. Importantly, it also enables serving on spot instances resulting in 30% lower cost.

The only change needed is to add a service section for serving specific configuration:

service:
  replicas: 2
  readiness_probe:
    path: /v1/chat/completions
    post_data:
      model: $MODEL_NAME
      messages:
        - role: user
          content: Hello! What is your name?
      max_tokens: 1

And run the SkyPilot YAML with a single command:

sky serve up -n gemma-serve gemma.yaml

Included files#

gemma3.yaml

resources:
  accelerators:  {L4:1, L4:4, A100:1, A100-80GB:1}
  ports:
    - 8000
  disk_tier: best
  use_spot: true

envs: 
  HF_TOKEN:
  MODEL_NAME: google/gemma-3-4b-it
  MAX_MODEL_LEN: 4096

setup: |
  uv pip install git+https://github.com/huggingface/transformers.git@0013ba6
  uv pip install vllm --extra-index-url https://wheels.vllm.ai/c0c25e25fa93ee7c3f279abbba5597c0fafa74ee

run: |
  echo 'Starting vllm openai api server...'
  python -m vllm.entrypoints.openai.api_server \
    --host 0.0.0.0 \
    --tensor-parallel-size $SKYPILOT_NUM_GPUS_PER_NODE \
    --model $MODEL_NAME \
    --max-model-len $MAX_MODEL_LEN

service:
  replicas: 1
  # An actual request for readiness probe.
  readiness_probe:
    path: /v1/chat/completions
    post_data:
      model: $MODEL_NAME
      messages:
        - role: user
          content: Hello! What is your name?
      max_tokens: 1