Cuda out of memory, 0 bytes free
lavellanedaaubay opened this issue · comments
I suggest you directly fine-tune my pre-trained model, which will significantly reduce your training time. If an OOM is reported, you can freeze some of the parameters of the model by adding the following code at line 116 in bart_model.py.
unfreeze_layers = ['layers.0', 'layers.1', 'layers.2', 'layers.3', 'layers.4', 'layers.5', 'layers.6',
'layers.7', 'layers.8']
for name, param in self.model.named_parameters():
for ele in unfreeze_layers:
if ele in name:
param.requires_grad = False
Thanks for your answer it worked we don't have the OOM error anymore. But we have trouble downloading the pretrained model. First we have a proxy issue and when we download it aside the pytorch_model.bin is changed into a zip and it won't work.
We are downloading it from here: https://huggingface.co/NTUYG/ComFormer/tree/main
Maybe you need install git fls, then run the follows:
git lfs install
git clone https://huggingface.co/NTUYG/ComFormer
Then you can get the model file.
If you are in China, I can upload it in Baidu Netdisk.
Thanks a lot for your answers everything work now. I will contact you if I get in other issues.