philschmid / deep-learning-pytorch-huggingface

MIT License
624 stars 145 forks source link

Regarding the OOM issues with fine-tuning Flan-T5-xl #58

Open Alsodream opened 2 months ago

Alsodream commented 2 months ago

I tried to do full fine-tuning on the Flan-T5-XL, but I have always faced the issue of OOM. I used 5 A5000 cards, each with 24GB, which should be acceptable in theory. However, I still have OOM. Do I have to use Deepspeed. In the explanation, I saw the word 'DS unload'. 'yes' means that Deepspeed was not used, right? Do any friends also conduct similar experiments? Can you tell me the reason? 1