pytorch-fast-transformers version

[YuanChang98]

Hello, I have replicated the conda environment following environment.md and downloaded the pre-trained model checkpoints, when I tried to run run_finetune_xxxx.sh I got this error: AttributeError: module 'torch.linalg' has no attribute 'qr' the corresponding file is fast_transformers/feature_maps/fourier_features.py, line 37. I looked up the pytorch documentation for module torch.linalg and found that this module does not have the function qr for pytorch1.7.1 and below. To save trouble, I didn't upgrade pytorch but downgraded pytorch-fast-transformers to version 0.3.0 which uses torch.qr instead and it worked. So I‘m wondering if there is any version mismatch between pytorch and pytorch-fast-transformers or just because I did't follow the guide well. Some of my conda environments are below: CUDA version 11.0 pytorch==1.7.1 pytorch-lightning==1.1.5 pytorch-fast-transformers==0.4.0 then changed to 0.3.0 cudatoolkit==11.0.221 apex==0.1 Thank you.

[AaronFeller]

I have inference running on the model and I didn't seem to have any problems once I ensured CUDA and Torch could talk to each other (matching versions). I also installed the GPU version of torch even though the current server I'm using does not have GPUs.

cuda==11.8.0 torch==2.0.1+cu118 torchaudio==2.0.2+cu118 torchvision==0.15.2+cu118 pytorch-lightning1.1.5 pytorch-fast-transformers==0.4.0 apex==0.1

[zhanxw]

For some reason, pytorch-fast-transformers uses torch.linalg.qr. You need to uninstall and install pytorch-fast-transformer. pip uninstall pytorch-fast-transformers pip install pytorch-fast-transformers==0.4.0

The pip installed version uses torch.qr: fast_transformers/feature_maps/fourierfeatures.py line 37: `Q, = torch.qr(block)`

This problem can also occur if you install pytorch-fast-transformers by compiling its source codes.