RuntimeError: unsupported operation: some elements of the input tensor and the written-to tensor refer to a single memory location. Please clone() the tensor before performing the operation.

[collabgear]

Hello,

When starting the train_gencad.py as mentioned in the README, I am getting the subject error.
Below is the full cli with stack trace:

GenCAD/GenCAD$ python train_gencad.py csr -name test -gpu 0

/media/collabgear/SSD-M2-1TB/PycharmProjects/GenCAD/GenCAD/model/cond_ldm.py:368: FutureWarning: torch.cuda.amp.autocast(args...) is deprecated. Please use torch.amp.autocast('cuda', args...) instead.
@autocast(enabled = False)

[INFO] GPU: 0

Total trainable params: 6.72 million

EPOCH[1/10]: 38%|
| 28955/76265 [13:45<24:56, 31.61it/s, loss_cmd=1.42, loss_args=5.49]Traceback (most recent call last):
File "/media/collabgear/SSD-M2-1TB/PycharmProjects/GenCAD/GenCAD/train_gencad.py", line 187, in
train_model(args.model, args)
File "/media/collabgear/SSD-M2-1TB/PycharmProjects/GenCAD/GenCAD/train_gencad.py", line 57, in train_model
ae_trainer.train(train_loader=train_loader, val_loader=val_loader, val_loader_all=val_loader_all, ckpt=args.ckpt_path)
File "/media/collabgear/SSD-M2-1TB/PycharmProjects/GenCAD/GenCAD/trainer/autoencoder_trainer.py", line 188, in train
outputs, loss_dict = self.validate_one_step(val_data)
File "/media/collabgear/SSD-M2-1TB/PycharmProjects/GenCAD/GenCAD/trainer/autoencoder_trainer.py", line 86, in validate_one_step
loss_dict = self.loss_fn(outputs)
File "/home/collabgear/miniconda3/envs/gencad_env/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1751, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
File "/home/collabgear/miniconda3/envs/gencad_env/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1762, in _call_impl
return forward_call(*args, **kwargs)
File "/media/collabgear/SSD-M2-1TB/PycharmProjects/GenCAD/GenCAD/utils/loss.py", line 26, in forward
padding_mask = get_padding_mask(tgt_commands, seq_dim=-1, extended=True) * visibility_mask.unsqueeze(-1)
File "/media/collabgear/SSD-M2-1TB/PycharmProjects/GenCAD/GenCAD/utils/model_utils.py", line 44, in get_padding_mask
torch.narrow(padding_mask, seq_dim, 3, S-3).add(torch.narrow(padding_mask, seq_dim, 0, S-3)).clamp(max=1)
RuntimeError: unsupported operation: some elements of the input tensor and the written-to tensor refer to a single memory location. Please clone() the tensor before performing the operation.

I had to use Pytorch of the version 2.71 as the most recent one does not support my GPU ( due to CUDA support ).
Could it be related?

[3dewberries]

For anyone who might encounter the same issue: I have cloned the repository and fixed the issue there: 3dewberries/GenCAD@d7a38b8

Training is running fine now.

Will keep experimenting.