降低训练显存占用可能性

[codexq123]

有可能利用fp8训练，将显存占用降低到24g吗

[Passenger12138]

If your hardware supports the FP8 format, you can certainly use FP8 for training and adopt mixed precision, where the large model operates in BF16 and the LoRA module in FP8, to reduce memory usage. However, it's important to note that the LoRA module itself has relatively few parameters, so using FP8 may not yield significant benefits in this case.