homelab/active/container_litellm/litellm.md

Podman litellm

• Podman litellm
  • Setup litellm Project
  • Install litellm
    • Create the ai user
    • Write the litellm compose spec
      • A Note on Volumes
    • Convert litellm compose spec to quadlets
    • Create the litellm.env file
    • Start and enable your systemd quadlet
    • Expose litellm
  • Using LiteLLM
    • Adding Models
    • Testing Models
  • Backup litellm
  • Upgrade litellm
    • Upgrade Quadlets
  • Uninstall
  • Notes
    • SELinux

Setup litellm Project

• Copy and rename this folder to active/container_litellm
• Find and replace litellm with the name of the service.
• Create the rootless user to run the podman containers
• Write the compose.yaml spec for your service
• Convert the compose.yaml spec to a quadlet
• Install the quadlet on the podman server
• Expose the quadlet service
• Install a backup service and timer

Install litellm

Create the ai user

# SSH into your podman server as root
useradd ai
loginctl enable-linger $(id -u ai)
systemctl --user --machine=ai@.host enable podman-restart
systemctl --user --machine=ai@.host enable --now podman.socket
mkdir -p /home/ai/.config/containers/systemd

Write the litellm compose spec

See the docker run command here

Edit the compose.yaml at active/container_litellm/compose/compose.yaml

A Note on Volumes

Named volumes are stored at /home/litellm/.local/share/containers/storage/volumes/.

Convert litellm compose spec to quadlets

Run the following to convert a compose.yaml into the various .container files for systemd:

# Generate the systemd service
podman run \
--security-opt label=disable \
--rm \
-v $(pwd)/active/container_litellm/compose:/compose \
-v $(pwd)/active/container_litellm/quadlets:/quadlets \
quay.io/k9withabone/podlet \
-f /quadlets \
-i \
--overwrite \
compose /compose/compose.yaml

# Copy the files to the server
export PODMAN_SERVER=ai-ai
scp -r active/container_litellm/quadlets/. $PODMAN_SERVER:/home/ai/.config/containers/systemd/

Create the litellm.env file

Should look something like:

LITELLM_MASTER_KEY="random-string"
LITELLM_SALT_KEY="random-string"

UI_USERNAME="admin"
UI_PASSWORD="random-string"

Then copy it to the server

export PODMAN_SERVER=ai
scp -r active/container_litellm/config.yaml $PODMAN_SERVER:/home/ai/litellm_config.yaml
ssh $PODMAN_SERVER chown -R ai:ai /home/ai/litellm_config.yaml

Start and enable your systemd quadlet

SSH into your podman server as root:

ssh ai
machinectl shell ai@
systemctl --user daemon-reload
systemctl --user restart litellm
journalctl --user -u litellm -f
# Enable auto-update service which will pull new container images automatically every day
systemctl --user enable --now podman-auto-update.timer

Expose litellm

1. If you need a domain, follow the DDNS instructions
2. For a web service, follow the Caddy instructions
3. Finally, follow your OS's guide for opening ports via its firewall service.

Using LiteLLM

Adding Models

// qwen3.5-35b-a3b-thinking
{
    "temperature": 1,
    "top_p": 0.95,
    "presence_penalty": 1.5,
    "extra_body": {
        "top_k": 20,
        "min_p": 0,
        "repetition_penalty": 1,
        "chat_template_kwargs": {
            "enable_thinking": true
        }
    }
}

// qwen3.5-35b-a3b-coding
{
    "temperature": 0.6,
    "top_p": 0.95,
    "presence_penalty": 0,
    "extra_body": {
        "top_k": 20,
        "min_p": 0,
        "repetition_penalty": 1,
        "chat_template_kwargs": {
            "enable_thinking": true
        }
    }
}

// qwen3.5-35b-a3b-instruct
{
    "temperature": 0.7,
    "top_p": 0.8,
    "presence_penalty": 1.5,
    "extra_body": {
        "top_k": 20,
        "min_p": 0,
        "repetition_penalty": 1,
        "chat_template_kwargs": {
            "enable_thinking": false
        }
    }
}

Testing Models

# List models
curl -L -X GET 'https://aipi.reeseapps.com/v1/models' \
-H 'Content-Type: application/json' \
-H 'Authorization: Bearer sk-1234'

curl -L -X POST 'https://aipi.reeseapps.com/v1/chat/completions' \
-H 'Content-Type: application/json' \
-H 'Authorization: Bearer sk-1234' \
-d '{
    "model": "gpt-4o-mini", # 👈 REPLACE with 'public model name' for any db-model
    "messages": [
        {
            "content": "Hey, how's it going",
            "role": "user"
        }
    ],
}'

Backup litellm

Follow the Borg Backup instructions

Upgrade litellm

Upgrade Quadlets

Upgrades should be a repeat of writing the compose spec and installing the quadlets

export PODMAN_SERVER=
scp -r quadlets/. $PODMAN_SERVER$:/home/litellm/.config/containers/systemd/
ssh litellm systemctl --user daemon-reload
ssh litellm systemctl --user restart litellm

Uninstall

# Stop the user's services
systemctl --user disable podman-restart
podman container stop --all
systemctl --user disable --now podman.socket
systemctl --user disable --now podman-auto-update.timer

# Delete the user (this won't delete their home directory)
# userdel might spit out an error like:
# userdel: user litellm is currently used by process 591255
# kill those processes and try again
userdel litellm

Notes

SELinux

https://blog.christophersmart.com/2021/01/31/podman-volumes-and-selinux/

:z allows a container to share a mounted volume with all other containers.

:Z allows a container to reserve a mounted volume and prevents any other container from accessing.