CUDA capability sm_86 is not compatible with the current PyTorch installation.

[sangnekim]

Hello.
I'd like to request about environment settings.

I set up conda env following README.md like below

# Optional solution for step 2: install libs step by step
conda create -n pointmlp python=3.7 -y
conda activate pointmlp
conda install pytorch==1.10.1 torchvision==0.11.2 cudatoolkit=10.2 -c pytorch -y
pip install cycler einops h5py pyyaml==5.4.1 scikit-learn==0.24.2 scipy tqdm matplotlib==3.4.2
pip install pointnet2_ops_lib/.

Then, I met CUDA capability sm_86 is not compatible with the current PyTorch installation. and code did not work anymore.

I checked environment.yml and I found that pytorch=1.8.1 and cudatoolkit=11.1. These are different from versions in README.md. I set up env following environment.yml and then everything goes well.

Thus, I think above installation guide should be modified like this.

# Optional solution for step 2: install libs step by step
conda create -n pointmlp python=3.7 -y
conda activate pointmlp
pip install torch==1.8.1+cu111 torchvision==0.9.1+cu111 torchaudio==0.8.1 -f https://download.pytorch.org/whl/torch_stable.html
pip install cycler einops h5py pyyaml==5.4.1 scikit-learn==0.24.2 scipy tqdm matplotlib==3.4.2
pip install pointnet2_ops_lib/.

[ma-xu]

@shkim960520
Thanks a lot for your kind suggestions. Looks like this error will only be caused by particular GPUs, like 3070, 3080 and 3090. Could you please double-check if your GPUs are 30X0?

I will add the modified guide in the readme file. Thanks.

[sangnekim]

Thanks for reply.

My current GPUs are 3090. However, I met similar error on NVIDIA-A100 GPUs.
Although I've not run codes on A100 gpus with torch=1.8.1 & cudatoolkit=11.1, I think it works

[ma-xu]

Thanks for reply.

My current GPUs are 3090. However, I met similar error on NVIDIA-A100 GPUs. Although I've not run codes on A100 gpus with torch=1.8.1 & cudatoolkit=11.1, I think it works

Seems this will happen to all Ampere GPUs, see https://discuss.pytorch.org/t/need-help-trouble-with-cuda-capability-sm-86/120235/6. For these GPUs, users may need CUDA>=11.0

[ma-xu]

closed since no further discussions. pls feel free to re-open it if necessary.