xformers `Version: 0.0.26.post1` is clashing with torch 2.3.1

[MoezGholami]

🐛 Bug

installing xformers pip install xformers (which at the moment points to 0.0.26.post1 uninstalls torch and triton 2.3.1 and replaces it with 2.3.0. This is causing conflict with the packages that do require 2.3.1 such as "torchvision 0.18.1+cu121" and "torchaudio 2.3.1+cu121".

Command

To Reproduce

$ pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu121
[skipped output]
$ pip install xformers==0.0.26.post1 # (currently equivalent to pip install xformers)
[...]
Collecting xformers
  Using cached xformers-0.0.26.post1-cp310-cp310-manylinux2014_x86_64.whl (222.7 MB)
Collecting torch==2.3.0
  Using cached torch-2.3.0-cp310-cp310-manylinux1_x86_64.whl (779.1 MB)
[...]
  Attempting uninstall: triton
    Found existing installation: triton 2.3.1
    Uninstalling triton-2.3.1:
      Successfully uninstalled triton-2.3.1
      [...]
  Attempting uninstall: torch
    Found existing installation: torch 2.3.1+cu121
    Uninstalling torch-2.3.1+cu121:
[...]
ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
torchvision 0.18.1+cu121 requires torch==2.3.1, but you have torch 2.3.0 which is incompatible.
torchaudio 2.3.1+cu121 requires torch==2.3.1, but you have torch 2.3.0 which is incompatible.

Expected behavior

xformers working with latest version of torch.

[ashleykleynhans]

I also need xformers to be compatible with torch 2.3.1 but I would say that this is a feature request rather than a bug.

[comaniac]

Second on this issue.

[dmaljovec]

This is not as simple as changing this line to:

            for torch_version in ['2.3.0', '2.3.1']:

Is it? If I understand the tests correctly, my guess is this is already tested with a later version given the use of the nightly channel from Pytorch used here.

[dmaljovec]

FWIW, I am experimenting with using the "install from source" directions and it is at least able to resolve an environment. I have yet to test that everything works together, but I am at least able to generate a lockfile with uv combining xformers@git+https://github.com/facebookresearch/xformers.git@main#egg=xformers and torch>=2.3.1.

[comtest99]

ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
xformers 0.0.26.post1 requires torch==2.3.0, but you have torch 2.3.1+cu121 which is incompatible.

[danthe3rd]

Hi,
We don't have a release for 2.3.1, we will make a release for 2.3.2 tho or 2.4.0. In the meantime, you can either downgrade pytorch, or install xFormers from source (long).

[youkaichao]

From what I heard from @albanD , they will not release 2.3.2 . Does this mean we need to wait for 2.4.0?

[danthe3rd]

Good point - in that case we will try to add a release for 2.3.1 then

[chi0tzp]

Good point - in that case we will try to add a release for 2.3.1 then

Thanks @danthe3rd, much appreciated! Any ETA for the 2.3.1 build? Cheers!

[danthe3rd]

We'll try to get dev wheels/conda packages early next week, and hopefully soon after that some stables for 2.3.1

[chi0tzp]

Cheers @danthe3rd!

[danthe3rd]

Version 0.0.27 just got published, and is compatible with PT 2.3.1