VRAM management could improve

[andreszs]

Please take no offense, but the demo workflow is probably one of the most unstable I've ever tried and all problems seem to be memory-related.

Summary of errors

1. SamplerCustomAdvanced: Allocation on device <- requires lowering settings, then it will appear again and can only be avoided by constantly restarting comfyui after each successful prompt. Unloading models does not help.
2. Expected all tensors to be on the same device, but found at least two devices, cpu and cuda:0! (when checking argument for argument mat2 in method wrapper_CUDA_bmm) -> this one appears in subsequent prompts
3. mat1 and mat2 must have the same dtype, but got Float and BFloat16 -> introduced by disabling mixed_precision to avoid the previous error

It's simply impossible to benchmark this workflow since it runs only once per comfyui session.