Different Results Same Checkpoint

Hey folks, I trained a segmentation model on a GPU, saved the checkpoint and inference works great.

I use the same model and checkpoint on my PC using the CPU and the results (segmentation masks) are a lot poorer than they were in the GPU.

Any ideas what could be going on here and how to resolve?

Generally, some numerical differences with different hardware are expected as different implementations will incur some differences due to operation reordering. Could you try inference on the same data (e.g., the same image) and check how much the resulting values differ between CPU and GPU?