faster fine-tune whisper suggestion

[ILG2021]

faster-whisper is a wonderful solution for faster inference but fine tune still need a better solution. I have rent a A6000 to fine-tune a 5.5 hours dataset, it cost 7 hours to train and need 44.8G rams by using the method of Sanchit Gandhi, I don't know what will cause if I add amout of the dataset. I have search a method named peft + lora + dnd to do faster train, but it doesn't support windows and can not convert to ctranslate2. So anyone gives some suggestion, thank you.

[monk1337]

You can try LLMtuner, It provides peft + lora etc in few lines of code.

https://github.com/promptslab/LLMtuner/tree/main