Issue installing 3.0.2 cuda-enabled version on windows linux subsystem (WSL2) distro Ubuntu-24.04 #4663
-
|
Dear DP developers, I'm currently setting up deepmd-kit on my office computer, which has a 4060Ti GPU and should be capable of running small-scale test simulations. The system is running Windows 11 (latest version) with the corresponding NVIDIA Game-ready drivers and Ubuntu 24.04 on WSL2. Upon launching WSL, I verified that nvidia-smi correctly detects my 4060Ti GPU and reports CUDA 12.8 availability within WSL. However, when attempting to install deepmd-kit 3.0.2 using the command: "conda create -n deepmd deepmd-kit lammps horovod" Conda only installs the CPU version, despite my system supporting CUDA. I tried setting CONDA_OVERRIDE_CUDA=12.6 and installing NVIDIA’s CUDA toolkit 12.6 within WSL, but neither resolved the issue: To troubleshoot, I tried installing older versions of deepmd-kit like 3.0.1 and 3.0.0. Surprisingly, these versions installed correctly with CUDA support:
This suggests a potential issue with the conda-forge packaging process for deepmd-kit 3.0.2, or possibly a deeper conda dependency resolution mechanism that I'm unaware of. Also, I'm uncertain about whether this is an issue specific to WSL, or it could happen in other Linux systems. Could someone confirm if this is a packaging issue? If it's not easily fixable, would it be possible to add a notice in the official documentation for WSL users, suggesting that they should downgrade to an older version if CUDA support is needed? Appreciated! |
Beta Was this translation helpful? Give feedback.
Replies: 3 comments 7 replies
-
|
Could you run |
Beta Was this translation helpful? Give feedback.
-
|
Seems to give me deepmd-kit-3.0.2-cuda126py312 now. I'll try finish this installation and see if things run normally ASAP, while keep you posted. |
Beta Was this translation helpful? Give feedback.
-
|
Hi Jinzhe, Up to now, deepmd 3.0.2 installed by explicitly specifying deepmd-kit=3.0.2=cuda* has successfully passed most of my daily usage scenario, including dp --pt freeze, dp --pt test, lammps, ASE calculator and dp evaluator. During these tests, my 4060Ti GPU can be correctly loaded and utilized. Based on that, there might be some issue with conda packaging in 3.0.2 making it unable to install the GPU version by default. Could you please have a further check or add a note in the documentation? Best. |
Beta Was this translation helpful? Give feedback.
-
|
Hey community, |
Beta Was this translation helpful? Give feedback.
-
|
Can you please advise the exact command to download the deepmd-kit GPU-version? I used |
Beta Was this translation helpful? Give feedback.
-
|
Just like Jinzhe said, |
Beta Was this translation helpful? Give feedback.
-
|
I used this code, but an error occurred：
|
Beta Was this translation helpful? Give feedback.
-
|
Can you try checking your cuda version with nvidia-smi command? It seems that you don't have a cuda driver on your system. |
Beta Was this translation helpful? Give feedback.
-
|
Thank you very much for your help. I found the cause in this document: https://conda-forge.org/docs/user/tipsandtricks/#installing-cuda-enabled-packages-like-tensorflow-and-pytorch. Since I'm using an HPC platform, I need to add |
Beta Was this translation helpful? Give feedback.
-
|
I can reproduce this result, which seems to be caused by a bug in conda. When I use mamba instead of conda, it installs the GPU version by default. |
Beta Was this translation helpful? Give feedback.
I can reproduce this result, which seems to be caused by a bug in conda. When I use mamba instead of conda, it installs the GPU version by default.