Providers

OpenModal supports multiple providers through a pluggable interface. All code outside providers/ is provider-agnostic — your code doesn't change when you switch backends.

Local (Docker)

Run everything on your local machine using Docker. No cloud account needed.

openmodal --local run examples/hello_world.py

Feature	How it works
f.remote()	docker run with your function
f.map()	Parallel calls to the same container
Sandboxes	docker run -d + docker exec
GPU	--gpus all with NVIDIA Container Toolkit
Volumes	Local directories (~/.openmodal/volumes/)
Image build	docker build locally (no registry)
Scaling	Single machine

Best for: development, testing, university labs with GPU machines but no cloud budget.

If you request a GPU that doesn't match your hardware, OpenModal tells you and suggests using cloud instead.

GCP (default)

Uses GKE (Google Kubernetes Engine) for all workloads. Auto-scales with spot GPUs and supports scale-to-zero.

openmodal run examples/hello_world.py
openmodal deploy examples/vllm_serving.py
openmodal run examples/sandbox.py

Auto-creates the GKE cluster on first run (~5 min one-time setup).

AWS

Single EKS backend for all workloads — Karpenter auto-provisions the right instance type based on your GPU request.

openmodal --aws run examples/hello_world.py

Feature	How it works
f.remote()	EKS pod with your function
Sandboxes	EKS pod + kubectl exec
GPU	Karpenter provisions spot GPU nodes (p5, g5, g6)
Volumes	Init container sync (S3)
Image build	Local docker build + ECR push
Scale-to-zero	KEDA ScaledObject
Scaling	Karpenter auto-provisioning

Auto-creates the EKS cluster on first run (~15 min one-time setup).

Azure

Single AKS backend for all workloads — KEDA comes as a built-in addon, so there's nothing extra to install.

openmodal --azure run examples/hello_world.py

Feature	How it works
f.remote()	AKS pod with your function
Sandboxes	AKS pod + kubectl exec
GPU	NC/ND-series VMs
Volumes	Init container sync (Azure Blob Storage)
Image build	Local docker build + ACR push
Scale-to-zero	KEDA (native AKS addon)
Scaling	AKS node autoprovisioning

Auto-creates the AKS cluster on first run (~5 min one-time setup).

Adding a new provider

All providers implement CloudProvider from providers/base.py:

class CloudProvider(ABC):
    def create_instance(spec, image_uri, name) -> (name, ip)
    def delete_instance(name)
    def list_instances(app_name) -> [{name, status, ip}]
    def wait_for_healthy(ip, port, timeout) -> bool

    def build_image(dockerfile_dir, name, tag) -> image_uri
    def image_exists(image_uri) -> bool

    def create_sandbox_pod(name, image_uri, timeout, **kwargs)
    def exec_in_pod(pod_name, *args, workdir, env) -> ContainerProcess
    def copy_to_pod(pod_name, local_path, remote_path)
    def copy_from_pod(pod_name, remote_path, local_path)

    def ensure_volume(name) -> uri
    def stream_logs(instance_name) -> Popen | None
    def machine_spec_str(gpu) -> str
    def instance_name(app, func, suffix) -> str

To add a provider: create providers/<name>/ implementing CloudProvider, then register it in providers/__init__.py.