charts

Draw Things gRPC Server Helm Chart

What is Draw Things?

Draw Things is an on-device Stable Diffusion app for macOS and iOS. The official draw-things-grpc-server-cli container exposes the same inference pipeline behind a gRPC API, so the phone/Mac app can offload generation to a beefy GPU box.

This chart deploys that gRPC server on Kubernetes, paired with a persistent volume for model storage and an NVIDIA GPU request. Behaviour is driven by the bjw-s-labs common library, so all knobs live in values.yaml.

Key features:

• 🎨 Same engine as the app — the Draw Things app talks to it natively over gRPC
• 🚀 GPU acceleration — one nvidia.com/gpu requested by default
• 💾 Persistent model library — checkpoints survive pod restarts via a PVC at /grpc-models
• 🧱 bjw-s common library — declarative, opinionated, no custom templates to maintain

TL;DR

helm repo add obeone https://charts.obeone.cloud
helm install draw-things obeone/draw-things

Prerequisites

• Kubernetes 1.31+
• Helm 3.18+
• An NVIDIA GPU node with the NVIDIA device plugin installed (so nvidia.com/gpu resources are schedulable). Depending on your runtime, you may also need to set defaultPodOptions.runtimeClassName: nvidia.

Installing the chart

helm install draw-things obeone/draw-things

The chart will:

1. Create a Deployment with one replica (single-GPU singleton).
2. Provision a 50 GiB PVC mounted at /grpc-models.
3. Expose a ClusterIP service on port 7859.

To override common knobs:

helm install draw-things obeone/draw-things \
  --set persistence.models.size=200Gi \
  --set defaultPodOptions.runtimeClassName=nvidia

Uninstalling the chart

helm uninstall draw-things

The PVC is retained by default — delete it explicitly if you want a clean slate:

kubectl delete pvc -l app.kubernetes.io/name=draw-things

Configuration

Example values.yaml

defaultPodOptions:
  runtimeClassName: nvidia

controllers:
  main:
    containers:
      main:
        resources:
          limits:
            nvidia.com/gpu: 1
          requests:
            nvidia.com/gpu: 1

persistence:
  models:
    enabled: true
    size: 200Gi
    accessMode: ReadWriteOnce
    globalMounts:
      - path: /grpc-models

service:
  main:
    ports:
      grpc:
        port: 7859

Parameters

Image

Key	Type	Default	Description
controllers.main.containers.main.image.repository	string	drawthingsai/draw-things-grpc-server-cli	Image repository
controllers.main.containers.main.image.tag	string	``	Image tag (defaults to appVersion: latest)
controllers.main.containers.main.image.pullPolicy	string	IfNotPresent	Image pull policy
controllers.main.containers.main.command	list	["gRPCServerCLI"]	Container command
controllers.main.containers.main.args	list	["/grpc-models"]	Container args

GPU & runtime

Key	Type	Default	Description
defaultPodOptions.runtimeClassName	string	unset	NVIDIA RuntimeClass name. Set to nvidia if your cluster needs explicit runtime opt-in.
controllers.main.containers.main.resources.requests.nvidia.com/gpu	int	1	GPUs requested (must equal limits)
controllers.main.containers.main.resources.limits.nvidia.com/gpu	int	1	GPUs allocated

Service

Key	Type	Default	Description
service.main.type	string	ClusterIP	Service type
service.main.ports.grpc.port	int	7859	gRPC service port
service.main.ports.grpc.protocol	string	TCP	Service port protocol

Persistence

Key	Type	Default	Description
persistence.models.enabled	bool	true	Enable PVC for /grpc-models
persistence.models.size	string	50Gi	PVC size
persistence.models.accessMode	string	ReadWriteOnce	PVC access mode
persistence.models.globalMounts	list	[/grpc-models]	Mount paths for the models volume

Ingress

Key	Type	Default	Description
ingress.main.enabled	bool	false	Enable ingress. gRPC requires end-to-end h2c support.
ingress.main.hosts	list	[]	Ingress hosts
ingress.main.tls	list	[]	Ingress TLS

Loading models

The container looks up checkpoints in /grpc-models. Three ways to populate it:

1. kubectl cp — fine for a quick test:

   POD=$(kubectl get pod -l app.kubernetes.io/name=draw-things -o name | head -n1)
   kubectl cp ./my-model.safetensors default/${POD#pod/}:/grpc-models/
   

2. Init container (downloader) — declare an extra container in controllers.main.initContainers that wgets checkpoints into /grpc-models at boot time.

3. Pre-populated PVC — point persistence.models.existingClaim at a PVC you already filled out-of-band (NFS share, restored snapshot, etc.).

Connecting from the Draw Things app

kubectl port-forward svc/draw-things 7859:7859

Then in the macOS / iOS app, add a remote server pointing to 127.0.0.1:7859.

Troubleshooting

Pod stuck in Pending — 0/N nodes are available: insufficient nvidia.com/gpu

The cluster has no schedulable GPU. Verify with:

kubectl get nodes -o json | jq '.items[].status.allocatable | with_entries(select(.key | contains("nvidia")))'

Install / fix the NVIDIA device plugin before continuing.

Container starts but immediately exits

Check logs:

kubectl logs -l app.kubernetes.io/name=draw-things

A frequent cause is an empty /grpc-models directory — populate it (see above) or the server will refuse to start.

Source code

• https://github.com/drawthingsai/draw-things-community
• https://hub.docker.com/r/drawthingsai/draw-things-grpc-server-cli
• https://github.com/obeone/charts/

Requirements

Repository	Name	Version
https://bjw-s-labs.github.io/helm-charts	common	5.0.1

Maintainers

Name	Email	Url
obeone	obeone@obeone.org