DPO training speed is so slow

[DuroyGeorge]

My device is 16+512 M4 macbook air. The notebook says it will take about 30 minutes on m3 macbook air, however, I have run the train code for 30 minutes but just finish no more than 20 steps, which is only about 1/6 of the whole training.

[rasbt]

Hi there, and hm, this is weird. Besides having an older chip, the only difference is that I have the 24 GB version. So one thought I am having is maybe in your case your laptop is using a lot of swap because it is running out of memory?

Maybe have the "Activity Monitor" open and double-check that you are not exceeding the available RAM (and if it's exceeding it, try to close as many apps as possible and only keep the code notebook running)

Let me know how it goes, I am curious to hear what you find

[DuroyGeorge]

The following is the truth: I transferred the models to MPS. I had assumed that MPS was similar to Apple's CUDA, capable of accelerating the training speed. However, in reality, this is not the case. When I adhered to the original code and conducted the training on the CPU, all tasks were completed smoothly.
Thank you for your response. Your work has been of great assistance to me.

[rasbt]

Glad that it works for you now. The yellow & red memory pressure indicators might actually explain the slower performance as there could be excessive memory swapping. And technically you are right, MPS should be preferred. But it's not quite as mature and CPU and CUDA support related to the discussion here: #625