Template Guide

Managed Workspace Templates

Launch curated GPU app workspaces from the portal, track startup progress, reopen active sessions, and use the proxied dashboard once the primary service is ready.

Current catalog

•LoRA Pilot is the available end-to-end Stable Diffusion workspace, with ControlPilot as the primary dashboard.
•OneTrainer, SwarmUI, and vLLM Server are preview managed app templates for dedicated fine-tuning, generation, and LLM serving workflows.
•Every managed template launch creates a workspace app session, a linked managed job, and a progress page.
•Template sessions are grouped under /portal/templates so users can reopen progress, open dashboards, or stop running sessions.

What templates are

Templates are curated launch lanes for managed GPU app workspaces. They are different from custom jobs: the image, command, primary UI port, workspace paths, and launch notes are preconfigured so users can start an app session instead of hand-writing a runtime contract.

•Use /portal/templates for curated managed app launches and session history.
•Use /portal/jobs/new when you need a raw custom image, command, and routing intent.
•The template catalog is backed by the orchestrator template definitions and exposed through the portal.

Templates in portal Custom job submit

Launch flow

Launching a managed template creates a workspace app session, submits a linked managed job, and sends the browser to the session progress page. That page polls the orchestrator until capacity is found, the container starts, and the primary dashboard becomes reachable.

•Provisioning means Jungle Grid is requesting managed capacity from the provider pool.
•Starting dashboard means a container or endpoint exists, but the primary app service has not passed readiness yet.
•Running with dashboard ready means the session has a Jungle Grid dashboard URL and the Open dashboard action is enabled.
•Stopping a session completes the linked job and schedules the managed provider reservation for deletion.

Progress route shape

/portal/templates/{templateSlug}/sessions/{sessionId}

Template sessions

The Template sessions section on /portal/templates groups active and past managed app launches by template. Each session row shows the session ID, linked job, job state, dashboard readiness, runtime, cost, and available actions.

•View progress opens the launch timeline and lifecycle activity for that session.
•Open dashboard appears only after the primary UI is reachable through the Jungle Grid dashboard proxy.
•Stop session is available to users with jobs:write permission while the session is still active.
•Session history is workspace-scoped, so switching workspaces changes the visible template sessions.

Template sessions

LoRA Pilot

LoRA Pilot is the main available creator workspace. It runs the notrius/lora-pilot:stable image and starts ControlPilot as the primary dashboard, with shared persistent workspace storage under /workspace.

•Status: available.
•Default workload type: fine-tuning, optimized for balanced placement.
•Primary UI: ControlPilot on port 7878 at /.
•Workspace path: /workspace; output path: /workspace/outputs.
•Supporting services include ComfyUI on port 5555 and Kohya SS on port 6666 inside the same workspace.
•Models, datasets, outputs, configs, and logs are expected to live under /workspace.

Runtime

image: notrius/lora-pilot:stable
command: /bin/bash
args:
  - -lc
  - /opt/pilot/bootstrap.sh && exec /usr/bin/supervisord -n -c /etc/supervisor/supervisord.conf

LoRA Pilot source LoRA Pilot docs

OneTrainer

OneTrainer is a preview training-focused workspace for LoRA, embeddings, full fine-tuning, and dataset tooling. It uses the AI-Dock OneTrainer container adapter and exposes the AI-Dock service portal as the primary UI.

•Status: preview.
•Default workload type: fine-tuning, optimized for speed.
•Runtime image: ghcr.io/ai-dock/onetrainer:latest-cuda.
•Primary UI: Service Portal on port 1111 at /.
•Workspace path: /workspace; output path: /workspace/outputs.
•Best suited for dedicated fine-tuning sessions rather than broad generation workflows.

Runtime

image: ghcr.io/ai-dock/onetrainer:latest-cuda
command: init.sh
args:
  - WORKSPACE=/workspace
environment:
  WORKSPACE: /workspace
  SERVICE_PORTAL_PORT_HOST: "1111"

OneTrainer source OneTrainer wiki

SwarmUI

SwarmUI is a preview generation workspace for prompt-to-image, image and video experimentation, workflow automation, and model switching. It is intended for interactive generation rather than a raw ComfyUI-only deployment.

•Status: preview.
•Default workload type: inference, optimized for balanced placement.
•Runtime image: mchaker/swarmui:latest.
•Primary UI: SwarmUI on port 7801 at /.
•Best suited for generation sessions and workflow experimentation.

Runtime

image: mchaker/swarmui:latest
command: /bin/bash
args:
  - -lc
  - cd /SwarmUI 2>/dev/null || cd /workspace/SwarmUI 2>/dev/null || cd /app/SwarmUI 2>/dev/null || cd /app; ./launch-linux.sh --host 0.0.0.0 --port 7801 --launch_mode none

SwarmUI source SwarmUI docs

vLLM Server

vLLM Server is a preview LLM inference template that launches the official vLLM OpenAI-compatible server image. It serves a selected Hugging Face model through a per-session API key and exposes the ready /v1 API base URL on the session progress page.

•Status: preview.
•Default workload type: inference, optimized for speed.
•Runtime image: vllm/vllm-openai:latest.
•Primary API: OpenAI-compatible server on port 8000.
•Default model: Qwen/Qwen3-0.6B; users can enter another Hugging Face model ID at launch.
•Use a Hugging Face credential for gated or private models.

Runtime

image: vllm/vllm-openai:latest
command: /bin/bash
args:
  - -lc
  - vllm serve "$VLLM_MODEL_ID" --host 0.0.0.0 --port 8000 --dtype auto --api-key "$VLLM_API_KEY" --generation-config vllm
environment:
  VLLM_MODEL_ID: Qwen/Qwen3-0.6B

OpenAI client

from openai import OpenAI

client = OpenAI(
    base_url="https://ready-vllm-endpoint/v1",
    api_key="jg_vllm_...",
)

response = client.chat.completions.create(
    model="Qwen/Qwen3-0.6B",
    messages=[{"role": "user", "content": "Hello"}],
)

vLLM source vLLM OpenAI server docs

Troubleshooting managed template launches

Most template issues fall into three buckets: provider capacity, image/container startup, or dashboard proxy readiness. The progress page and lifecycle logs should identify which bucket the current session is in.

•Provider scarcity shows as waiting for capacity or provisioning retries while Jungle Grid searches for an available GPU.
•Docker Hub unauthenticated pull limits should mark the linked job and app session failed instead of showing a successful launch.
•If a GPU is found but the app is still starting, the progress page should move from provisioning into image/container or dashboard startup phases.
•If a direct RunPod link works but the Jungle Grid dashboard does not, check the stored upstream host, primary port, and dashboard proxy logs.
•Dashboard 404s for app-local /api, /views, /css, /js, /assets, /static, /images, or /favicon.ico paths are handled by rewriting those paths through the session proxy.

When reporting a template issue, include the session ID, linked job ID, current phase, and the latest lifecycle log lines from the progress page.