please help me

[Atai0610]

/home/zhaopingtai/miniconda3/envs/aihwkit/lib/python3.10/site-packages/torch/cuda/init.py:235: UserWarning:
NVIDIA GeForce RTX 5060 Laptop GPU with CUDA capability sm_120 is not compatible with the current PyTorch installation.
The current PyTorch install supports CUDA capabilities sm_50 sm_60 sm_61 sm_70 sm_75 sm_80 sm_86 sm_89 sm_90 compute_90.
If you want to use the NVIDIA GeForce RTX 5060 Laptop GPU GPU with PyTorch, please check the instructions at https://pytorch.org/get-started/locally/

[PabloCarmona]

What version of Pytorch and CUDA are you using?

[Atai0610]

[PabloCarmona]

Did you try using Pytorch 2.10 or greater? Also, are you using the latest version fo the aihwkit(1.1.0)?

[Atai0610]

我因为安装问题真的要崩溃了
我先安装的pytorch 版本如图

然后用pip install aihwkit 安装
是不能用gpu的

如果用 conda install -c conda-forge aihwkit-gpu 安装 又会自动配置pytorch2.6 头疼的很

[PabloCarmona]

Hey @Atai0610, did this get resolved after the guidance on the #775 discussion?

[delcenjo]

That warning is the core problem: the PyTorch you have installed was built for older CUDA architectures (up to sm_90) and does not include kernels for your GPU. The RTX 5060 (Blackwell, compute capability sm_120) needs a PyTorch built against CUDA 12.8 or newer. With an older build there is no kernel image for sm_120, so every CUDA operation fails even though the card is detected.

There are two parts to getting aihwkit running on this GPU:

1. PyTorch. Install a CUDA 12.8 (or newer) build into your conda env:

pip uninstall -y torch
pip install torch --index-url https://download.pytorch.org/whl/cu128

Then verify:

python -c "import torch; print(torch.__version__, torch.cuda.is_available(), torch.cuda.get_device_name(0))"

You should see a cu128 version, True, and the 5060.

2. aihwkit's CUDA extension. aihwkit's GPU support is a compiled CUDA extension, so it also has to contain kernels for sm_120. The prebuilt GPU packages are compiled for a fixed set of architectures and generally do not include Blackwell yet, so on a 50-series card you will most likely need to build aihwkit from source against a CUDA 12.8+ toolkit, with the build configured to target the sm_120 architecture (passing the CUDA architectures to the CMake build, e.g. -DCMAKE_CUDA_ARCHITECTURES=120). Build it in the same environment that has the cu128 PyTorch so the two match.

If you do not specifically need the analog GPU kernels, aihwkit also runs on CPU, which sidesteps the compatibility issue entirely while you sort out the GPU build.

To confirm which build you currently have, it helps to post the output of:

python -c "import torch; print(torch.__version__, torch.version.cuda)"