conda cannot install latest version (2.6.1) with pytorch 2.5.*

[themaigod]

😵 Describe the installation problem

conda cannot install latest version (2.6.1) with pytorch 2.5.*. It seems that conda channel did not upload file pyg-2.6.1-py312_torch_2.5.0_cu124.tar.bz2.
I could not find correct file in https://anaconda.org/pyg/pyg/files?page=1&version=2.6.1

Environment

• PyG version: 2.6.1
• PyTorch version: 2.5.1
• OS: Windows
• Python version: 3.12
• CUDA/cuDNN version: 12.4
• How you installed PyTorch and PyG (conda, pip, source): conda
• Any other relevant information (e.g., version of torch-scatter): None

[akihironitta]

@themaigod I'd suggest installing via pip for now.

[themaigod]

@akihironitta Yes I will do that. But it is wired. Upload such a file on conda channel in my opinion, is not a hard thing. So I guess probably there is an automatically issue. By the way, if it will not be a quick fix, I suggest the document can update to mention it.

[akihironitta]

@themaigod Other PyG core might also have different opinions, but when I worked on PyTorch Lightning (that even has no C++/CUDA dependency), it was often deprioritised when automation for publishing releases on conda channel had issues due to the small number of downloads via conda. PyTorch has recently decided to move away from their official pytorch conda channel because of the maintenance cost and the much smaller number of conda users than that of pip pytorch/pytorch#138506.

I suggest the document can update to mention it.

Although I think this issue and other conda-related installation GitHub issues are one way to let users know of the current state, feel free to contribute your suggestion!

---

Just to understand the needs, what makes you use conda install when you can use install a pure Python package? I use conda for managing multiple envs but not for installing packages just like this comment: pytorch/pytorch#138506 (comment).

[themaigod]

@akihironitta Well, I consider this question as a researcher from school. As for pytorch and related packages, I think cuda version is an important issue, especially when working on a fixed cuda version, like public servers in lab. Conda give a simple way by cudatoolkit. However, the installation from pip of some packages sometimes cannot recoginize the true version of pytorch. So, the consequent actions will be better in conda.

In addition, the dependency between packages are complex, thus you need to downgrade some packages manually if you are using pip. It is widely happened in reproducing the results of some papers with their author codes. And actually, it is very hard to find a valid combination, while conda may be able to do it automatically.

However, I feel that conda recently meet some problems, as a user. For example, I create a new python environment, and directly install pytorch by conda. It will fail in the first attempt, and spend a lot of time on the second or the third attempt to find the solution. The following actions of installation on other packages will become more tough, and even take hours or fail. So I am using a replacement like mamba.

[themaigod]

@akihironitta I just finished read #138506 . I am concern about that, because I believe it will increase the cost of users to build stable environment in some situations. And more importanatly, it will increase the barrier of beginners. So I think it will influence the whole environmemnt not obviously but widely.
As comment, compatibility on the all kinds of ways is the key of pytorch success--I believe that it is one of the reasons of pytorch over tensorflow, and another reason is the consistency of API. So tensorflow 2.0 did not make a change--dynamic mode is important but not the key.