RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu!

[abc77654]

您好，请教下：我按照步骤和工作流提示下载完成模型和工作流，运行遇到了错误：
RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu! (when checking argument for argument mat2 in method wrapper_CUDA_addmm)

[Vigilence]

From what I remember, comfyui now puts the vae on the cpu to save vram. Not sure if this is the issue here, but it looks like it could be. The dev might be able to add an option for us to choose what loads the model (cpu, gpu) to fix the issue.

[mingyi456]

@Vigilence It should not be due to the vae, right? I do not use comfyui myself, but my experience from watching dozens of comfyui tutorial videos tells me that the only difference between this workflow and a standard Flux workflow is swapping the load diffusion model node with the load df11 model. So the issue should be with the df11 model weights being "called" before they are transferred to the gpu, I believe.

[mingyi456]

@abc77654 Are you still facing the same problem? I just set up comfyui with the df11 custom node, and since you did not provide your workflow (I checked your image, and it does not seem to have an embedded workflow), I manually reproduced your layout (which seems to be the stock df11 workflow, just rearranged) from scratch, and I could not reproduce your issue.

I also tried the stock df11 workflow and also substituted the df11 node into the stock SPRO workflow, and they all work without issues. So the issue might be with your setup, but I am not sure what exactly it is.

What are the command line arguments you used to launch comfyui?