Question about the suitable version of xformers

[hnuzhy]

Hi, thanks a lot for releasing you code of awesome works RVT and RVT-2.
When I installed RVT-2 and run pip install -e '.[xformers]', I still met a bug as below:
AttributeError: '_OpNamespace' object has no attribute '_efficient_attention_forward'.
This bug is related to a non-defined torch function torch.ops.aten._efficient_attention_forward, which is imported by xformers.
I used your recommended torch version 1.12.1. And I guess this problem occurred due to the rapid updating of the lib xformers.
Actually, in the setup.py file, it always pulls down the lateset version.
Thus, I would like to ask what is the matched version of xformers you used when running RVT-2.

[imankgoyal]

Hi, @hnuzhy

I have used this version for xformers 0.0.19+70161e5.d20230419

python3 -c "import xformers; print(xformers.__version__)"
0.0.19+70161e5.d20230419

I hope it works for you. Please comment if it does so that it can help others who are trying the same.

Best,
Ankit

[hnuzhy]

Hi, @hnuzhy

I have used this version for xformers 0.0.19+70161e5.d20230419

python3 -c "import xformers; print(xformers.__version__)"
0.0.19+70161e5.d20230419

I hope it works for you. Please comment if it does so that it can help others who are trying the same.

Best, Ankit

OK. Thank you for your reply.
Actually, after trying, I used the v0.0.22, and it worked with PyTorch 1.12.1.
Maybe you can modified this line into
"xformers @ git+https://github.com/facebookresearch/xformers.git@v0.0.22#egg=xformers" or
"xformers @ git+https://github.com/facebookresearch/xformers.git@v0.0.19#egg=xformers"
for installing a matched xformers.

[hnuzhy]

By the way, I also met the version problem when installing PyTorch3D. Because its installation script also contained the @stable suffix, which keeps changing and may not be matched with PyTorch 1.12.1. Although I wanted to install RVT-2, it seems that I should also have to install the rvt/libs/peract locally, which needs to use the PyTorch3D. Would you mind show us the right version of it? I have tried Version 0.7.1 and Version 0.7.2, but they all failed.

[hnuzhy]

Continue from previous post, I found a workable installation way of PyTorch3D from here

$ conda install -c fvcore -c iopath -c conda-forge fvcore iopath
$ conda install -c bottler nvidiacub
$ conda install pytorch3d -c pytorch3d

Hope this will be useful for other followers~

[imankgoyal]

Thanks for sharing this info @hnuzhy! I will add it to the Readme for others to discover.