English | Русский

CDLP-OlamaUncensored

Roadwarrior setup for a private Ollama GPU host and one peer host on the same private network.

This repo is for a simple two-host layout:

• one GPU Ubuntu host running Ollama
• one peer Ubuntu host allowed to reach it over private IP only

The scripts are tuned for DigitalOcean, but usually work on similar Ubuntu or Debian-family hosts too.

Included files

• roadwarrior-ollama-gpu-host.sh Installs or updates Ollama on the GPU host, configures systemd, enables private-network serving, optionally locks access with UFW, pulls the model, and can create a derived local model with a Modelfile.
• roadwarrior-ollama-peer-host.sh Verifies that the peer host can reach the GPU host over the private network, checks /api/tags, optionally checks /api/chat, and prints the exact OpenClaw onboarding command for the remote Ollama endpoint.

Quick start

Run this on the GPU Ubuntu host:

curl -fsSL https://raw.githubusercontent.com/vektort13/CDLP-OlamaUncensored/main/roadwarrior-ollama-gpu-host.sh | bash

Run this on the peer Ubuntu host:

curl -fsSL https://raw.githubusercontent.com/vektort13/CDLP-OlamaUncensored/main/roadwarrior-ollama-peer-host.sh | bash

GPU host profiles

The GPU-host script includes ready presets so you do not need to enter model tuning values by hand.

Profile	Base model	Derived model	Use when
best-general	qwen3.5:27b	qwen3.5-local	Best overall default for most 24GB+ cards.
best-coding	gemma4:31b	gemma4-coder-local	Best built-in preset here for coding and reasoning.
low-vram	dolphin3:8b	dolphin3-local	Conservative option for lower VRAM hosts.
manual	Any local tag	Optional	Full control over model, keep-alive, local-only mode, and Modelfile settings.

All preset profiles default to the same Roadwarrior behavior:

• private-network serving only
• UFW limited to one peer host on TCP/11434
• local-only Ollama mode enabled
• model pull enabled
• local smoke test enabled

VRAM recommendations

Use this as a practical starting point, not a theoretical maximum-fit chart.

VRAM	Start with	Why
16GB	dolphin3:8b	Safest fit with room for runtime overhead and a smaller context.
24GB	qwen3.5:27b	Best current general-purpose single-GPU default.
48GB	gemma4:31b	Strong coding and reasoning preset with more context headroom.
80GB	gemma4:31b	Strongest preset included here unless you intentionally want to go manual.

If you want something between the presets, gpt-oss:20b is still a good alternate when you want a more developer-oriented model in roughly the 16GB to 24GB range.

Recommended topology

• GPU host: Ubuntu droplet with NVIDIA GPU and Ollama
• Peer host: Ubuntu box that runs OpenClaw or any other client
• Network: same VPC or other private network
• Exposure: private IP only

If a DigitalOcean Cloud Firewall is attached to the GPU droplet, add the same inbound TCP/11434 allow rule there that the GPU-host script applies in UFW.

What the peer-host script checks

The peer-host script verifies:

• the GPU host is reachable over private IP
• /api/tags works from the peer host
• /api/chat works too if the GPU host already has a local model pulled
• the exact openclaw onboard --custom-base-url ... command you should run next

It uses the native Ollama API URL and intentionally does not use /v1.

OpenClaw note

These scripts do not install OpenClaw.

The peer-host script only helps you validate the private Ollama path and then either:

• prints the exact openclaw onboard command to use
• runs that command directly if openclaw is already installed on the peer host

Notes

• These scripts expect an interactive TTY.
• The GPU-host script is tuned for one peer host, not a shared public endpoint.
• The GPU-host script binds Ollama to 0.0.0.0:11434, so the firewall rule is what keeps the service private.
• The default derived models are intentionally tuned for direct local use and lighter filtering.
• Derived Modelfiles are stored under /opt/roadwarrior-ollama/modelfiles.
• Large model pulls can take time, especially on a fresh host.