Override model with frozen settings

[rkfg]

I run a semi-private instance for friends with frozen settings so that no one can accidentally or maliciously break things. However, this doesn't allow swapping models for just one generation. I changed the code to allow overriding settings and wrote an extra network extension that adds a tag like <model:hash:name> to the prompt when a model card is clicked at the checkpoints tab. It mostly works fine but breaks if SDXL is used in img2img, looks like this model swap happens too late (unlike when settings are overridden from PNG Info for example) and the SD 1.5 code path is used with an SDXL model. At least that's what I figured from TypeError: DiffusionWrapper.forward() got an unexpected keyword argument 'crossattn'.

Long story short, I need to add the Model hash parameter to Override settings dropdown when the model is clicked at the Checkpoints tab instead of trying to swap it immediately. Is it possible? I found that this "override" dropdown component can only be used from Gradio while the checkpoint cards use pure HTML/JS and there's seemingly no way to call Gradio from JS. The Gradio docs are really, really bad in this regard and basically any advanced usage like JS/Python interop isn't covered anywhere.

If this is too big of a problem, maybe there's another way to run a public web UI instance with most of the settings blocked but still allow people to use different models, parameters and other stuff without interfering with each other? Basically apply the parameters (clip skip etc.), load the desired model, generate, return results, revert the parameters back to some default that I set, then process the next job and so on.

[rkfg]

After much struggle I found a way to connect back to Python from JS, and that is via events. I create a hidden text input, set the value and emulate the input event because otherwise Gradio doesn't update the internal value. In short, the code is like this:

In ui.py, create_override_settings_dropdown I added this:

def update(x: str, dd):
        params = dict([e.split(': ')[0], e.split(': ')[1]] for e in dd)
        parts = x.split('|')
        params["Model"] = parts[1]
        params["Model hash"] = parts[0]
        return gr.Dropdown.update(value=[e + ': ' + params[e] for e in params.keys()], choices=params)

    control = gr.Textbox(elem_id="override_settings_control", visible=False)
    control.input(fn=update, inputs=[control, dropdown], outputs=[dropdown])

Not very foolproof but okay. It merges the value in form of hash|modelname with the existing overrides.
Then in modules/ui_extra_networks_checkpoints.py I changed the onclick handler to this:

            "onclick": '"' + f"""var ta = get_uiCurrentTabContent().querySelector('#override_settings_control textarea');
                ta.value = '{checkpoint.shorthash}|{checkpoint.name_for_extra}';
                ta.dispatchEvent(new Event('input'));
                get_uiCurrentTabContent().querySelector('.extra-networks > .tab-nav > button:nth-child(1)').click()""" + '"',

It finds the textbox from above, sets the value and emulates user input, after that it switches to the Generation tab.

Goal achieved but at what cost...