thunlp
Received a fatal error！！
test fewshot.py, get the below fatal error.
pytorch 1.10.0
for step, inputs in enumerate(train_dataloader):
if use_cuda:
inputs = inputs.cuda()
logits = prompt_model(inputs)
labels = inputs['label']
loss = loss_func(logits, labels)
loss.backward() # it causes the fatal error.
torch.nn.utils.clip_grad_norm_(prompt_model.parameters(), 1.0)
tot_loss += loss.item()
optimizer1.step()
one of the variables needed for gradient computation has been modified by an inplace operation: [torch.cuda.FloatTensor [20, 1076]], which is output 0 of SoftmaxBackward0, is at version 1; expected version 0 instead. Hint: the backtrace further above shows the operation that failed to compute its gradient. The variable in question was changed in there or anywhere later. Good luck!
Browsed much materials, pytorch 1.4 or below can support it. wish somebody can fix it.
I run on pytorch 1.9.0. But it is strange that it requires 1.4 or below to solve the bug. Can you start a new virtual environment and test it?
Is it correct on pytorch 1.9.0? To create a virtual env is difficult. Pytorch 1.4 or below is not compatible with cuda 11.7. Guess need to downgrade my cuda? (not cudatoolkit）
After verification, pytorch 1.9 does not work! pytorch 1.9, cuda 11.7, python 3.8. nvidia 3090
For torch 1.4, we found that transofrmers complains torch>=1.5.0 is required for a normal functioning of this module, but found torch==1.4.0+cu92. Could you please tell us what configuration you are using ?
openprompt 1.0.1 pytorch 1.9, cuda 11.6/7 for fewshot.py, it doesn't work. For pytorch 1.4, it is impossible, because transformer 4.2 needs the version 1.5 of pytorch. please verify the combination: pytorch 1.9 or above, cuda 11.6/7 openprompt 1.01 ( up-to-date)
