Container Backends
SForge runs work and judge containers on a backend: local Docker by default, k8s for a Kubernetes cluster, or e2b for E2B cloud Sandboxes. One sforge run uses one backend for all its tasks; choose it with --backend or backend: in an experiment file.
| Backend | Use it for | Needs |
|---|---|---|
docker | Trying tasks, task development, small runs | A local Docker daemon |
k8s | Full-suite and other large runs | kubectl access to a cluster |
e2b | Running without Docker or a cluster | An E2B account and Templates for the tasks |
WARNING
The Docker backend is for a few tasks at a time: around 20 concurrent tasks exhaust even a large host. Use k8s for the full suite.
The Judge Server
Every run has a judge server. It grades submissions in fresh judge containers, runs game sessions and forwards the agent's model requests, so the sandboxes must be able to reach it. sforge run serves it in-process:
| Backend | Sandboxes reach it at |
|---|---|
| Docker | http://host.docker.internal:<port>, set up automatically |
| Kubernetes, E2B | http://<--judge-advertise>:<--judge-port>: an IP of this machine the pods or Sandboxes can reach |
To share one judge between several sforge run processes or hosts, start sforge serve and pass --judge-url to the runs. They must share SFORGE_ADMIN_SECRET, which is automatic on one machine with the same log directory.
Resource limits (--work-cpu-limit, --work-mem-limit, --judge-cpu-limit, --judge-mem-limit) override the task's cpus and memory; on Kubernetes they become Pod requests and limits.
Docker
The default backend. Images come from local Docker, via sforge pull or sforge build:
sforge pull --task ad_placement_optimization
sforge run --task ad_placement_optimization --agent claude-codeKubernetes
Each container becomes a Pod, labeled app=sforge. You need:
kubectlaccess, and a namespace where you can create Pods and NetworkPolicies;- nodes that can pull the task images (by default from the benchmark's
registry,docker.io/seededgefor EdgeBench); - an IP of this host that pods can reach, for
--judge-advertise; - a CNI that enforces NetworkPolicy, for isolated tasks.
export SFORGE_K8S_NAMESPACE=sforge
sforge run --backend k8s --task ad_placement_optimization --agent claude-code \
--judge-advertise 10.0.0.12 --judge-port 8080Pods pull <SFORGE_K8S_IMAGE_REGISTRY>/<image>:<tag>. To pull from a registry near the cluster, or to run tasks you have changed, push the images there and point SFORGE_K8S_IMAGE_REGISTRY at it:
sforge push --task ad_placement_optimization --registry registry.example.com/sforge
export SFORGE_K8S_IMAGE_REGISTRY=registry.example.com/sforgeThe other SFORGE_K8S_* variables are listed in Environment Variables. For the full suite, see All Tasks (Kubernetes).
E2B
Work, judge and game containers run as E2B Sandboxes. The judge server still runs on your host, so --judge-advertise must be an address reachable from the internet.
E2B starts Sandboxes from Templates rather than images. SForge derives each Template name from the image key (edgebench.work.ad_placement_optimization:7b2f533ff7d4 becomes edgebench-work-ad-placement-optimization:7b2f533ff7d4) and looks for it in your E2B team. SFORGE_E2B_TEMPLATE_NAMESPACE selects Templates published under a shared namespace, and SFORGE_E2B_TEMPLATE_MAP maps image keys to other names. The E2B example shows how to build Templates.
A Template's CPU and memory are fixed when it is built. Your E2B plan bounds them, the Sandbox lifetime and the concurrency; keep --timeout well below the plan's maximum Sandbox lifetime.
For Volcano Engine veFaaS, which speaks the e2b protocol, set SFORGE_E2B_PROVIDER=vefaas and E2B_DOMAIN; SFORGE_E2B_TEMPLATE_MAP then maps image keys to function ids.