xformers installs incompatible torch version (2.7.1) despite having torch 2.5.1+cu121 installed for CUDA 12.1

[chnisar515]

Custom Node Testing

• I have tried disabling custom nodes and the issue persists (see how to disable custom nodes if you need help)

Your question

Hi team,

I'm currently setting up ComfyUI in a clean virtual environment using Python 3.10.11, Windows 11, and a NVIDIA RTX 5070 (Ada, 12GB) with CUDA 12.1 support. I encountered a critical dependency conflict while attempting to install xformers alongside the CUDA-enabled version of PyTorch.

My Environment:

OS: Windows 11 64-bit
Python: 3.10.11 (inside a venv)
GPU: NVIDIA GeForce RTX 5070 (Ada)
CUDA version: 12.1 (confirmed via PyTorch and nvidia-smi)
pip: 24.0
ComfyUI: Git clone of master
Torch target: torch==2.5.1+cu121 from the official PyTorch CUDA 12.1 index

Installation Steps I Followed:
python -m venv venv venv\Scripts\activate

Installed the correct CUDA-enabled torch stack:
pip install torch==2.5.1+cu121 torchvision==0.20.1+cu121 torchaudio==2.5.1+cu121 --extra-index-url https://download.pytorch.org/whl/cu121

Verified torch and CUDA compatibility:
python -c "import torch; print(torch.__version__, torch.cuda.is_available())"

Then I attempted to install xformers:
pip install xformers

and output is below

(venv) C:\ComfyUI\ComfyUI>pip install torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 --extra-index-url https://download.pytorch.org/whl/cu121
Looking in indexes: https://pypi.org/simple, https://download.pytorch.org/whl/cu121
Collecting torch==2.5.1
Using cached https://download.pytorch.org/whl/cu121/torch-2.5.1%2Bcu121-cp310-cp310-win_amd64.whl (2449.4 MB)
Collecting torchvision==0.20.1
Using cached https://download.pytorch.org/whl/cu121/torchvision-0.20.1%2Bcu121-cp310-cp310-win_amd64.whl (6.1 MB)
Collecting torchaudio==2.5.1
Using cached https://download.pytorch.org/whl/cu121/torchaudio-2.5.1%2Bcu121-cp310-cp310-win_amd64.whl (4.1 MB)
Requirement already satisfied: filelock in c:\comfyui\comfyui\venv\lib\site-packages (from torch==2.5.1) (3.18.0)
Requirement already satisfied: typing-extensions>=4.8.0 in c:\comfyui\comfyui\venv\lib\site-packages (from torch==2.5.1) (4.14.1)
Requirement already satisfied: networkx in c:\comfyui\comfyui\venv\lib\site-packages (from torch==2.5.1) (3.4.2)
Requirement already satisfied: jinja2 in c:\comfyui\comfyui\venv\lib\site-packages (from torch==2.5.1) (3.1.6)
Requirement already satisfied: fsspec in c:\comfyui\comfyui\venv\lib\site-packages (from torch==2.5.1) (2025.5.1)
Collecting sympy==1.13.1 (from torch==2.5.1)
Using cached https://download.pytorch.org/whl/sympy-1.13.1-py3-none-any.whl (6.2 MB)
Requirement already satisfied: numpy in c:\comfyui\comfyui\venv\lib\site-packages (from torchvision==0.20.1) (2.2.6)
Requirement already satisfied: pillow!=8.3.*,>=5.3.0 in c:\comfyui\comfyui\venv\lib\site-packages (from torchvision==0.20.1) (11.3.0)
Requirement already satisfied: mpmath<1.4,>=1.1.0 in c:\comfyui\comfyui\venv\lib\site-packages (from sympy==1.13.1->torch==2.5.1) (1.3.0)
Requirement already satisfied: MarkupSafe>=2.0 in c:\comfyui\comfyui\venv\lib\site-packages (from jinja2->torch==2.5.1) (3.0.2)
Installing collected packages: sympy, torch, torchvision, torchaudio
Successfully installed sympy-1.13.1 torch-2.5.1+cu121 torchaudio-2.5.1+cu121 torchvision-0.20.1+cu121

(venv) C:\ComfyUI\ComfyUI>pip install xformers
Collecting xformers
Using cached xformers-0.0.31.post1-cp39-abi3-win_amd64.whl.metadata (1.1 kB)
Requirement already satisfied: numpy in c:\comfyui\comfyui\venv\lib\site-packages (from xformers) (2.2.6)
Collecting torch==2.7.1 (from xformers)
Using cached torch-2.7.1-cp310-cp310-win_amd64.whl.metadata (28 kB)
Requirement already satisfied: filelock in c:\comfyui\comfyui\venv\lib\site-packages (from torch==2.7.1->xformers) (3.18.0)
Requirement already satisfied: typing-extensions>=4.10.0 in c:\comfyui\comfyui\venv\lib\site-packages (from torch==2.7.1->xformers) (4.14.1)
Collecting sympy>=1.13.3 (from torch==2.7.1->xformers)
Using cached sympy-1.14.0-py3-none-any.whl.metadata (12 kB)
Requirement already satisfied: networkx in c:\comfyui\comfyui\venv\lib\site-packages (from torch==2.7.1->xformers) (3.4.2)
Requirement already satisfied: jinja2 in c:\comfyui\comfyui\venv\lib\site-packages (from torch==2.7.1->xformers) (3.1.6)
Requirement already satisfied: fsspec in c:\comfyui\comfyui\venv\lib\site-packages (from torch==2.7.1->xformers) (2025.5.1)
Requirement already satisfied: mpmath<1.4,>=1.1.0 in c:\comfyui\comfyui\venv\lib\site-packages (from sympy>=1.13.3->torch==2.7.1->xformers) (1.3.0)
Requirement already satisfied: MarkupSafe>=2.0 in c:\comfyui\comfyui\venv\lib\site-packages (from jinja2->torch==2.7.1->xformers) (3.0.2)
Using cached xformers-0.0.31.post1-cp39-abi3-win_amd64.whl (100.2 MB)
Using cached torch-2.7.1-cp310-cp310-win_amd64.whl (216.1 MB)
Using cached sympy-1.14.0-py3-none-any.whl (6.3 MB)
Installing collected packages: sympy, torch, xformers
Attempting uninstall: sympy
Found existing installation: sympy 1.13.1
Uninstalling sympy-1.13.1:
Successfully uninstalled sympy-1.13.1
Attempting uninstall: torch
Found existing installation: torch 2.5.1+cu121
Uninstalling torch-2.5.1+cu121:
Successfully uninstalled torch-2.5.1+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.
torchaudio 2.5.1+cu121 requires torch==2.5.1+cu121, but you have torch 2.7.1 which is incompatible.
torchvision 0.20.1+cu121 requires torch==2.5.1+cu121, but you have torch 2.7.1 which is incompatible.
Successfully installed sympy-1.14.0 torch-2.7.1 xformers-0.0.31.post1

(venv) C:\ComfyUI\ComfyUI>

Logs

Other

No response

[Mayukh-Haldar]

I am not an expert but I found two solutions to this problem. First of all, your CUDA version is 12.1 for which xformers builds specific to torch 2.5.1 are not readily available. So I tried to investigate the issue.

1.) So i ran this command in cmd in a python virtual env (I am using windows with CUDA Version 12.9).
pip install torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 xformers --index-url https://download.pytorch.org/whl/cu121
I included xformers in the same line as it will then look for the compatible version of xformers library in the index url for pytorch 2.5.1 with cuda 12.1. But unfortunately after installing torch 2.5.1, torchvision 0.20.1, torchaudio 2.5.1 and multiple versions of xformers, it raised the following error message:

INFO: pip is looking at multiple versions of sympy to determine which version is compatible with other requirements. This could take a while.
INFO: pip is looking at multiple versions of torchaudio to determine which version is compatible with other requirements. This could take a while.
INFO: pip is looking at multiple versions of torchvision to determine which version is compatible with other requirements. This could take a while.
INFO: pip is looking at multiple versions of to determine which version is compatible with other requirements. This could take a while.
INFO: pip is looking at multiple versions of torch to determine which version is compatible with other requirements. This could take a while.
ERROR: Cannot install torch==2.5.1, torchaudio==2.5.1+cu121, torchvision==0.20.1+cu121, xformers==0.0.22.post4, xformers==0.0.22.post7, xformers==0.0.23, xformers==0.0.23.post1, xformers==0.0.24, xformers==0.0.25, xformers==0.0.25.post1, xformers==0.0.26, xformers==0.0.26.post1, xformers==0.0.27, xformers==0.0.27.post1 and xformers==0.0.27.post2 because these package versions have conflicting dependencies.

The conflict is caused by:
The user requested torch==2.5.1
torchvision 0.20.1+cu121 depends on torch==2.5.1+cu121
torchaudio 2.5.1+cu121 depends on torch==2.5.1+cu121
xformers 0.0.27.post2 depends on torch==2.4.0
The user requested torch==2.5.1
torchvision 0.20.1+cu121 depends on torch==2.5.1+cu121
torchaudio 2.5.1+cu121 depends on torch==2.5.1+cu121
xformers 0.0.27.post1 depends on torch==2.4.0
The user requested torch==2.5.1
torchvision 0.20.1+cu121 depends on torch==2.5.1+cu121
torchaudio 2.5.1+cu121 depends on torch==2.5.1+cu121
xformers 0.0.27 depends on torch==2.3.1
The user requested torch==2.5.1
torchvision 0.20.1+cu121 depends on torch==2.5.1+cu121
torchaudio 2.5.1+cu121 depends on torch==2.5.1+cu121
xformers 0.0.26.post1 depends on torch==2.3.0
The user requested torch==2.5.1
torchvision 0.20.1+cu121 depends on torch==2.5.1+cu121
torchaudio 2.5.1+cu121 depends on torch==2.5.1+cu121
xformers 0.0.26 depends on torch==2.2.2
The user requested torch==2.5.1
torchvision 0.20.1+cu121 depends on torch==2.5.1+cu121
torchaudio 2.5.1+cu121 depends on torch==2.5.1+cu121
xformers 0.0.25.post1 depends on torch==2.2.2
The user requested torch==2.5.1
torchvision 0.20.1+cu121 depends on torch==2.5.1+cu121
torchaudio 2.5.1+cu121 depends on torch==2.5.1+cu121
xformers 0.0.25 depends on torch==2.2.1
The user requested torch==2.5.1
torchvision 0.20.1+cu121 depends on torch==2.5.1+cu121
torchaudio 2.5.1+cu121 depends on torch==2.5.1+cu121
xformers 0.0.24 depends on torch==2.2.0
The user requested torch==2.5.1
torchvision 0.20.1+cu121 depends on torch==2.5.1+cu121
torchaudio 2.5.1+cu121 depends on torch==2.5.1+cu121
xformers 0.0.23.post1 depends on torch==2.1.2
The user requested torch==2.5.1
torchvision 0.20.1+cu121 depends on torch==2.5.1+cu121
torchaudio 2.5.1+cu121 depends on torch==2.5.1+cu121
xformers 0.0.23 depends on torch==2.1.1
The user requested torch==2.5.1
torchvision 0.20.1+cu121 depends on torch==2.5.1+cu121
torchaudio 2.5.1+cu121 depends on torch==2.5.1+cu121
xformers 0.0.22.post7 depends on torch==2.1.0
The user requested torch==2.5.1
torchvision 0.20.1+cu121 depends on torch==2.5.1+cu121
torchaudio 2.5.1+cu121 depends on torch==2.5.1+cu121
xformers 0.0.22.post4 depends on torch==2.1.0

To fix this you could try to:

1. loosen the range of package versions you've specified
2. remove package versions to allow pip attempt to solve the dependency conflict

ERROR: ResolutionImpossible: for help visit https://pip.pypa.io/en/latest/topics/dependency-resolution/#dealing-with-dependency-conflicts

So i found that xformers 0.0.27.post2 depends on torch==2.4.0. So then, I uninstalled the existing packages and then ran the following installation command:

pip install torch==2.4.0 torchvision==0.19.0 torchaudio==2.4.0 xformers --index-url https://download.pytorch.org/whl/cu121

It did install xformers 0.0.27.post2, but it came up with another issue. I made a test python file in the directory where the venv was created and named it main.py, and wrote the following code:
import xformers
import torch
print(xformers.__version__, torch.__version__)

Then I executed the file with python main.py and it raised the following issue:

A module that was compiled using NumPy 1.x cannot be run in
NumPy 2.1.2 as it may crash. To support both 1.x and 2.x
versions of NumPy, modules must be compiled with NumPy 2.0.
Some module may need to rebuild instead e.g. with 'pybind11>=2.12'.

If you are a user of the module, the easiest solution will be to
downgrade to 'numpy<2' or try to upgrade the affected module.
We expect that some modules will need time to support NumPy 2.

Then finally I installed numpy 1.26.4 with pip install numpy==1.26.4

Which finally solved the issue.
The final output of python main.py was 0.0.27.post2 2.4.0+cu121

2.) Else, what you can do is to upgrade your CUDA version with the help of the link CUDA Toolkit Archive and install xformers with torch for either 12.4 or 12.6 depending on your CUDA version.

[github-actions]

This issue is being marked stale because it has not had any activity for 30 days. Reply below within 7 days if your issue still isn't solved, and it will be left open. Otherwise, the issue will be closed automatically.

[M-Bag]

I've been having the same issue, but while this method makes absolutely no sense to me why pip didn't work, but pip3 did is beyond me, however, I had to install pytorch, xformers together. Then to get sage-attn, flash_attn, and triton all together, I had to download the wheels for my specific versions to my local machine.

pip3 download triton-windows==3.3.1.post19 sageattention==1.0.6 flash_attn-2.7.4.post1+cu128 --dest H:\pipcache\wheels

Then installed them with pip3.

pip3 install --no-index --find-links=H:\pipcache\wheels sage-attn flash_attn triton-windows

One difference is I'm using pytorch with CUDA support for 12.8, so I had to do a little digging to find out which versions would play nice together, but they are functional now! Hope that helps.

M<