CPU execution fails due to unconditional use of torch.amp.autocast

[orangeSi]

Hi Profluent team,

Thank you for open-sourcing Protein2PAM — it is a very useful and well-designed tool.

I encountered an issue when running Protein2PAM in a CPU-only environment. The program crashes with the following error:

RuntimeError: Currently, AutocastCPU only support Bfloat16 as the autocast_cpu_dtype

This happens because torch.amp.autocast is always enabled in oracle.evaluate(), even when device_type == "cpu". However, PyTorch CPU autocast only supports bfloat16, while the current implementation uses float32 (or an incompatible default), which leads to this runtime error.

Thank you very much for your work and for maintaining this repository!

Best regards,
Si