That doesn't seem relevant here to me though.

To clarify, training on the GPU in RStudio or python notebooks works fine. Models run and they save perfecty.

Theres only issues when running it within zorro, the second and all models afterwards fail.