Automatically create a LLM plugin from a gradio app

[freddyaboulton]

• [x] I have searched to see if a similar issue already exists.

Is your feature request related to a problem? Please describe.
The gradio-tools library showed that there is interest in using gradio apps as tools for LLMs. I'd like to take one of the experiments I was working on in that library and bring it to the gradio_client library to maximize usage and impact: automatically creating LLM plugins from gradio apps.

My proposal is the following:

Add a deploy_plugin method from a gradio_client.Client instance with the following signature:

    def deploy_plugin(self,
        api_names: List[str],
        org_name: str | None = None,
        version: str = "v1",
        hf_token: str | None = None,
        private: bool = True,
        email: str | None = None,
        llm: Literal["chatgpt"] = "chatgpt"):

Note: email is needed by ChatGPT plugins.

Under the hood - this will create a new space that serves the necessary metadata needed for ChatGPT to recognize the space as a ChatGPT plugin. Right now I'm mentioning ChatGPT because that's the LLM with the most mature plugin ecosystem that I know of but it should be possible to support other LLMs (e.g. huggingchat when they allow plugins). The llm argument will make that configuration possible.

Then users can do the following once there is a space they want to deploy as a plugin:

client = grc.Client("freddyaboulton/video-creator")
client.deploy_plugin(api_names=['/generate'], private=False)

We can even make this a configuration parameter to gradio deploy, e.g.

gradio deploy-plugin freddyaboulton/video-creator --api-names generate  --private false

The benefits of doing this approach:

• Users interested in using gradio as plugins pay for their own compute. I also thought about modifying gradio core so that all gradio apps can be plugins but that may spam spaces with automated calls that will degrade the experience of users using the UI. There has been concern about this in the past with just the current gradio_client
• ChatGPT plugins timeout after 60 seconds so this will incentivize users to duplicate spaces to keep runtime low
• Create an ecosystem of plugins on the hub. The hub can be the place to share/discover plugins. We can create a gallery space to showcase the plugins created by the community.

[abidlabs]

This is awesome, I'm a big fan of the duplication + deployment approach -- we can use the same approach to deploying Gradio apps to other platforms as well.

    def deploy_discord(self,
        api_names: List[str],
        discord_token: str
        org_name: str | None = None,
        version: str = "v1",
        hf_token: str | None = None,
        private: bool = True):

and

gradio deploy-discord freddyaboulton/chat