OSError: Could not find compatible tinycudann extension for compute capability 89

[Lucklycat]

i install is ok ,but when i run a train.py code ,there havea problem:

i am try add the TCNN_CUDA_ARCHITECTURES=86,but there is no effect.

please give me some advice.

my env: os: unbutu Graphics card：RTX4090 cuda: 11.6

[Tobe2d]

+1

[tancik]

Also running into the same issue. What is the recommended software versions when using at 4090?

I have tried,

• CUDA 11.7, nvidia-driver 525, pytorch 1.12.1+cu116 -> No compute capability 89
• CUDA 11.7, nvidia-driver 515, pytorch 1.12.1+cu116 -> Installs, but can't find the 4090
• CUDA 12, nvidia-driver 525, pytorch 1.12.1+cu116 -> Can't compile when pytorch cuda (11) / and system cuda (12) don't match

It does install if I use TCNN_CUDA_ARCHITECTURES=86, but this will be less efficient, right?

[Tom94]

Unfortunately, this seems like a case of compatibility hell to me -- ~I don't think there's a way to compile for compute capability 89 and have a PyTorch version with CUDA <11.8. I'm surprised that your second bullet installs at all~. EDIT: actually, if you install CUDA 11.8 on your system while using PyTorch with 11.6 or 11.7, you should be able to. Compute capability 89 is supported starting with CUDA 11.8.

In any case, from personal experiments in C++-land, tiny-cuda-nn is more-or-less equally fast when running with TCNN_CUDA_ARCHITECTURES=86 on a 4090 as it is when running 89, so you don't need to worry about this.

What I'll do on my end (probably after the SIGGRAPH deadline) is to hardcode CUDA version vs. compute capability compatibilities into setup.py (ugh) to make sure compilation falls back to the latest supported version rather than outright failing. Sorry about that one -- the CMake script already has this implemented, but it's not present on the Python side yet.

[tancik]

Thanks @Tom94 for the pointers, I had not realized that compute capability support did not start until CUDA 11.8.

The following worked for me! CUDA 11.8, nvidia-driver 520, pytorch 1.13.1

[Tobe2d]

CUDA 11.8 did work for me too

[yuntaeJ]

In my RTX4090 pc,

CUDA 11.8, nvidia-driver 525, pytorch 2.0.0

works for me. I installed it using below command:

conda install pytorch==2.0.0 torchvision==0.15.0 torchaudio==2.0.0 pytorch-cuda=11.8 -c pytorch -c nvidia