Kindly add the Force Set Device for TextEncoder!

[Iory1998]

Regarding the Flux models, in ComfyUI, there is the feature to set/force loading text-encoders to the CPU rather the GPU, which helps with Vram management a lot. It does impact speed a bit, but it's a about a few seconds more, but the benefit is that I can use higher precision models and save time by unloading the text-encoders from Vram each time there is a change in the model or LoRA.

[lllyasviel]

you should move this PR to #1591 (comment) -> https://github.com/Juqowel/GPU_For_T5

[Iory1998]

Thank you. I did that.