accelerate support for train_finetune.py with one GPU

[yl4579]

Request per #96 (comment). Current finetuning script is quite slow because it doesn't support accelerate due to #7, but a lot of people are finetuning using a single GPU, so we don't really have to worry about DP or DDP and can directly turn to accelerate which supports mixed precision that saves time and RAM. It would be great if someone could implement this feature for a single GPU. For multi-GPU we are still trying to solve #7 .

[korakoe]

I'm doing a few tests now... but I believe I may have implemented this, following this guide

[korakoe]

After testing, it doesn't appear that VRAM usage drops by very much, however, training speed is seemingly faster from what I'm observing. I could be missing a few things, however.

I'll do a quick training run just to see

UPDATES:

• After the first epoch, VRAM usage seems to drop by 1GB, from 13.6GB to 12.6GB on Google Colab
• After a certain amount of time, the above statement becomes untrue, jumping back up to its initial value
• after saving a model it jumps by 1GB, from 13.6GB to 14.6GB, the cause in unknown

[yl4579]

@korakoe Did you do 16 bit mixed precision? Does it involve diffusion training or SLM adversarial training?

[korakoe]

@korakoe Did you do 16 bit mixed precision? Does it involve diffusion training or SLM adversarial training?

I'm using fp16 mixed precision, Diffusion training but not SLM (due to Colab, however the implementation does work)

[korakoe]

after it jumps to 14GB it drops to 11GB, so I'm assuming that accelerate is doing some form of offloading?

[yl4579]

@korakoe I think the main concern in #96 (comment) is the speed, so if the speed is faster I guess it's fine, even if VRAM drop is not that big.