CUDA OOM When Using Flash Attention

[HaniItani]

trafficstars

Hello,

Thank you for sharing your awesome work!

I'm trying to train Vicuna on my own dataset. I walked through the installation process from source. I had to install pytorch with cuda 11.7.0 support instead of 11.6. My server only supports cuda 11.2.2/11.4.4/11.5.2/11.7.0/11.8.0 but not 11.6.

When I try to train the 13B model with flash attention, I get CUDA OOM error even when the per_device_train_batch_size is set to 1. I think there might be a memory leak. I also tried building flash attention from source and still got the same error.

I know this is probably a flash attention problem, but do you have any insights? Any guidance will be very much appreciated.

Best regards, Hani

[HaniItani]

I just tried training the 7B model and it works fine with both flash attention and without. It would be great if I can train models>7B with flash attention. I think using flash attention will enable me to train the 13B model only. Beyond that I'll need to look into memory efficient tuning methods such as LoRA

[zhisbug]

How many GPUs do you use?

[HaniItani]

@zhisbug, thank you for your reply.

I was using 4A100-80GB GPUs but then realized the repository uses 8.

I have a questions regarding the batch size. The run code uses 8GPUs with per_device_train_batch_size set to 4 and gradient_accumulation_steps set to 1. Correct me if I'm wrong, but the total batch size would be 8x4x1=32, but the repo says it uses a global batch size of 128.

I checked the Alpaca repo and they set the gradient_accumulation_steps to 8 for their 7B model so the global batch size would be 4x4x8 =128. Am I missing something?

Your input will be very much appreciated!

[zhisbug]

@Michaelvll might be the right person to clarify the training config

[zhisbug]

@HaniItani is your problem solved?

[HaniItani]

Yes, I figured it out. Thank you very much.

[alwayshalffull]

@HaniItani what did you do to fix it? i'm also running into OOM training 13B models with flash attention, wondering if it's the same problem. thanks!

[HaniItani]

Hi @alwayshalffull , Flash Attention was not the problem, I had the wrong parameters. The 13B model requires 8 80GB-A100 GPUs to finetune with a per device train batch size of 4 as reflected in scripts/train_vicuna_13b.sh. If you're getting OOM when saving the model, please check this issue. Hope this helps.