Forge Spaces: Install, Launch, Stop, and Remove Mini-Apps

It is an experimental launcher for local mini-apps adapted from Hugging Face Spaces. Each entry has its own wrapper, pinned Space snapshot, interface, dependencies, models, and local Gradio server. It is not a gallery of web links and not the normal txt2img pipeline.

The inspected clean Original Forge commit exposes nine: Animagine XL 3.1, BiRefNet, Florence-2, GeoWizard, IC-Light, IDM-VTON, IllusionDiffusion, PhotoMaker V2, and Sapiens Normal. A tenth forge_space_example directory is disabled by configuration and is not counted.

It creates the Space huggingface_space_mirror directory and downloads the repo_id at the revision in space_meta.json. If the built-in wrapper has requirements.txt, Install also runs pip against the active Forge Python environment.

Install fetches the pinned Space snapshot. forge_app.py can then call from_pretrained, hf_hub_download, or direct URLs for base models, adapters, processors, and checkpoints. Those secondary fetches are a separate phase.

The pinned snapshot lives under the built-in Space directory in huggingface_space_mirror. Hugging Face model cache is redirected to the Forge models/diffusers directory. Relative outputs can be created inside the mirror because Launch changes the process working directory there.

No. The manager deletes the Space mirror and explicitly warns that models/diffusers can remain because those cache entries may be reused. It also does not uninstall Python packages added to the shared Forge environment.

It can. The manager recursively removes the mirror, and some wrappers use relative output directories while the process working directory is that mirror. Export files you care about before clicking Uninstall.

For a Space with requirements.txt, the inspected UI deliberately labels the enabled Install button “Reinstall” whenever the mirror exists. That label alone is not an error. Use the console and Launch result to determine whether installation is complete.

The installed check only tests whether huggingface_space_mirror exists. A partial snapshot, failed pip step, unavailable secondary model, dependency mismatch, unsupported device operation, or remote-code error can therefore fail later. Capture the first traceback before changing anything.

Not reliably from Install status alone. A Space may need its pinned snapshot, secondary model repositories, direct checkpoint URLs, and remote Python modules during first Launch or first use. Offline operation is only established after that exact Space and task complete with network access blocked.

Do not assume so. These wrappers construct their own Diffusers or Transformers pipelines and resolve named repositories through the shared Hugging Face cache. The checkpoint selected in the main Forge header usually does not control them.

The manager starts at the configured Forge port, tests ports in sequence, and chooses the first free one. If the main UI owns 7860, the first Space commonly receives 7861, but the actual console URL is authoritative.

No. Browser closure is not wired to the manager state. Return to Forge → Spaces, press Terminate, and wait until the entry returns to “Installed, Ready to Launch”.

No. The wrapper executes Python in the same process, temporarily changes process-wide cwd and sys.path, can install packages into the active environment, and shares GPU and model-cache state. Some wrappers explicitly enable trust_remote_code.

The inspected Space launch call does not forward the main Gradio auth or TLS arguments. It reuses the server binding name, so --listen or a custom server name can expose the separate Space port without the protection you expected from the main UI. Keep Forge local unless you have independently secured every port.

The UI does not enforce a global one-Space limit, but the implementation shares process-wide cwd, sys.path, module names, model cache, and GPU management. This audit does not certify concurrent use; run one at a time when diagnosing or protecting work.

No support matrix is established by the inspected code. Several wrappers call CUDA directly, use float16, or state that CPU is unsupported. Treat a successful Install as no hardware guarantee and preserve the exact first error.

There is no verified end-user “paste any Space URL” workflow. Forge identifies an extension as a Space only when it has both space_meta.json and forge_app.py, and the wrapper must integrate with the local spaces runtime. Copying arbitrary remote code into extensions is a developer and security task, not a normal install step.