Add Multi-GPU function

[gurilagardnr]

Would you consider adding the ability to load the model to GPU-1/2/3 for multi-gpu users to help us keep all the things loaded in vram?

[fpgaminer]

Is there another node that does this, so I can see the best way to implement that? If not, I'll probably just add a separate node with a field for selecting which GPU # to load the model onto.

[diodiogod]

Is there another node that does this, so I can see the best way to implement that? If not, I'll probably just add a separate node with a field for selecting which GPU # to load the model onto.

https://github.com/city96/ComfyUI_ExtraModels have the override/force clip and Vae, IDK if that is what the OP is asking for.

[without-ordinary]

https://github.com/pollockjj/ComfyUI-MultiGPU

Comfy-Org/ComfyUI#7063
https://github.com/comfyanonymous/ComfyUI/tree/worksplit-multigpu

Edit: I found that is uses multiple GPUs as-is but there is no control over which GPUs are used, it uses all GPUs that are made visible to ComfyUI. This is both cool and a little annoying (only because there is currently no means to control which GPUs are used). I think all that would be needed is the ability to pass a value to device_map.

[without-ordinary]

What I might possible do is run JoyCaption via vllm outside of ComfyUI and access it via API calls. Having an API node for that would be rather nice.