Support for loading a smaller model for sample generation?

[jnz86]

I have a 12GB card. The trainer says it’s working but the output is terrible, separate issue.

I wanted to turn on samples. But I’m using the FP16 Hunyuan checkpoint with swapped blocks. This checkpoint will never run on its own on my machine. It will always Out Of Memory even if it’s a 256x256x1 generation.

So just like the sample settings that point to VAE and Encoders, would it be possible to point to a different model like the FP8 or a GGUF as a checkpoint for the sample generation?