Changing checkpoint from another app

[Kilvoctu]

Hello,

I have a Python discord bot I'm using as a frontend for the webui.
While it has the basic functionality of sending prompt request to the webui and getting an image back, I've been trying to figure out how to have this bot send a request to swap the checkpoint.
The use case is so Discord users can generate their image using their desired checkpoint from available list.

Is anyone able to point me in the right direction on how to develop this? Thanks.

[ReiFan49]

You may need to provide the api_name on given element.
The element key is options_templates["sd_model_checkpoint"], you need to modify the component_args part to have api_name key in it.

Example on bb295f5 snapshot:

stable-diffusion-webui/modules/shared.py

Line 243 in bb295f5

"sd_model_checkpoint": OptionInfo(None, "Stable Diffusion checkpoint", gr.Dropdown, lambda: {"choices": modules.sd_models.checkpoint_tiles()}, refresh=sd_models.list_models),

from

OptionInfo(None, "Stable Diffusion checkpoint", gr.Dropdown, lambda: {"choices": modules.sd_models.checkpoint_tiles()}, refresh=sd_models.list_models)

into

OptionInfo(None, "Stable Diffusion checkpoint", gr.Dropdown, lambda: {"choices": modules.sd_models.checkpoint_tiles(), "api_name": "sd_model"}, refresh=sd_models.list_models)

However, the current model is focused on handling "one selected model at a time". So, changing them at will may require further queueing.

Let's say:

User A > model "wd1.4"
User B > model "sd1.4"
User C > model "sd1.3"

When User A queues the a prompt, their stored model setting will be prepended to the API call so the model will change right before generating an image.

With an example queue of: "User A", "User B", "User A", "User C", "User B". The API call formed will be in this order:

1. Set model to wd1.4
2. Sent prompt of fischl and kirima sharo singing
3. Set model to sd1.4
4. Sent prompt of beautiful moon in the sky
5. Set model to wd1.4
6. etc.

This is still untested but the base theory is that you need to have internal queueing on your bot to change model over and over without sticking to fn_index that always changes upon git pulling the repo.

[aliencaocao]

Ah, i actually went to use devtools and planned to do the same as a last resort. Thanks for the reply, i will check out the repo you linked, and also have commented on the feature request.

Do you have experience with how to keep 2 models in VRAM at the same time? If not, the loading is quite long. I have 12GB and I believe it is enough to fit.

[Kilvoctu]

Yeah you'll have enough. The setting is this one

As you can see, cached checkpoins use the RAM instead. They become cached after loading each checkpoint once, then swapping will be almost instant. On 8GB VRAM and 16GB RAM, my machine can handle 2 checkpoints and have memory to spare

[aliencaocao]

Is there any way to change the default for this or do I have to set it everytime I launch the UI?

[Kilvoctu]

That particular setting will stick to whatever you set the slider to. I set mine to 2 a long time ago and haven't needed to touch it again.

[aliencaocao]

Ok, thank you