About using bf16_mode in inference

Hello,

In the latest release of the Aurora library I noticed a new parameter called bf16_mode. The documentation explains that it is only used for fine-tuning, but I tried to enable bf16_mode during inference with the default checkpoint (0.25° fine-tuned version). Over several days of testing, GPU memory usage dropped noticeably and I have not observed any obvious degradation in the forecast output.

Because my GPU has limited VRAM (<24 GB), I would like to know whether using bf16_mode for inference has any hidden drawbacks, like numerical, or otherwise, compared with standard FP32 inference. Could you please run or share a small benchmark and clarify whether bf16_mode is supported for inference? If not, can it be set by default on fine-tuning instead providing this parameter on Aurora class?

Start with the Aurora class and the documentation describing bf16_mode for fine-tuning, then inspect how the default checkpoint is used during inference. Run a small comparison of bf16_mode and FP32 inference, including GPU memory and forecast output, and document whether inference is supported or whether the parameter should default during fine-tuning.