Ollama

Ollama runs open large language models locally, backed by remote compute. This image bundles the Ollama server with a built-in web chat interface — open the page, pick a model, and start chatting. The Ollama HTTP API is served from the same port for your own apps.

Usage

1. Deploy

dxflow workflow create --identity ollama hub://ollama

# Start with defaults, or tune per run with --override
dxflow workflow start ollama
dxflow workflow start ollama \
    --override env.app.PASSWORD=my-strong-pass \
    --override env.app.STARTUP_MODEL=qwen2.5:1.5b

# Publish the web port on an HTTPS link
dxflow workflow start ollama --link

2. Open the interface

Open your browser at http://localhost:8080 and sign in as dxflow with the password you set in PASSWORD. The chat UI lists the installed models — pick one and start a conversation. The streaming response renders as it is generated. A start given --link publishes port 8080 at an HTTPS URL printed on the start line, serving the chat UI and the API from anywhere.

3. Use the API

The Ollama HTTP API is proxied under the same port at /api, behind the same credential, so your own tools can call it:

curl -u dxflow:my-strong-pass http://localhost:8080/api/chat -d '{
  "model": "smollm2:135m",
  "messages": [{ "role": "user", "content": "Hello!" }]
}'

Configuration

Attach a GPU with resources.gpu: nvidia so Ollama offloads the model to it, or remove the line to run on CPU.

name: ollama
tags:
    - ai
steps:
    - name: app
      runtime: docker
      mode: parallel
      image: ghcr.io/dxflow-ai/ollama:latest
      volumes:
          - name: volume
            host: ./volume
            container: /volume
      ports:
          - name: web
            host: "8080"
            container: "8080"
      env:
          - PASSWORD=dxflow
          - STARTUP_MODEL=smollm2:135m
      resources:
          cpu: "4"
          memory: 8G
          gpu: nvidia
      link: web

[volume]
app.volume = ./volume

[port]
app.web = 8080

[env]
app.PASSWORD = dxflow
app.STARTUP_MODEL = smollm2:135m

[resource]
app.cpu = 4
app.memory = 8G
app.gpu = nvidia

{
    "arch": ["amd64", "arm64"],
    "image": "ghcr.io/dxflow-ai/ollama:latest",
    "version": "0.34.0",
    "size": {
        "amd64": "3.6G",
        "arm64": "2.8G"
    },
    "minimum": {
        "cpu": 4,
        "memory": "8G",
        "storage": "50G"
    }
}

Notes

• STARTUP_MODEL is pulled on startup and selected in the UI (default smollm2:135m, preloaded into the image). Pull more models any time from a terminal with ollama pull <name>.
• The web interface is a React app (served by nginx) that reverse-proxies to the local Ollama server on 11434 — the UI calls it under /ollama/api/*, and the standard API is also exposed directly at /api/*, so the browser and the API share port 8080.
• GPU vs CPU: with a GPU attached, Ollama detects it and loads as many layers into VRAM as fit — check nvidia-smi while a model answers. Without one it runs on CPU automatically, which suits the small models; for 7B and up, keep the card and give the step more memory.
• Set a strong PASSWORD; it defaults to dxflow, which every reader of this page knows. nginx checks it as HTTP basic auth for the user dxflow across the whole port, so the UI and the API share one credential.