lanctot
commented
1 year ago My Question now is whether i can use newer CUDA drivers for example 12.2 with latest Cudnn. In the global_variables.sh i could only see the options 10.2 and lower which is very old. Did somebody test newer CUDA drivers?
According NVIDIA Toolkit theres not even a 10.2 Cuda version which is supported by Ubuntu 22 (which is the recommended OS for this project) just 18 and lower. https://developer.nvidia.com/cuda-10.2-download-archive?target_os=Linux&target_arch=x86_64&target_distro=Ubuntu
You mean this? https://github.com/google-deepmind/open_spiel/blob/a95a9d6589a57e9c7587aa228785aae16f7597f3/open_spiel/scripts/global_variables.sh#L83
Those are external links. They are hosted on pytorch.org. All you need to do is find the one you want. You can get those URLs by visiting this page (https://pytorch.org/get-started/locally/) and pick the selections that fit your environment. E.g. I just found the link for CUDA 12.1 on Linux.
If this whole approach is just there to be viewed by externals and not really maintained, is there maybe a better recommended alternativ to using ./alphazero_torch_example ?
From the README:
Important note: this implementation was a user contribution (see https://github.com/deepmind/open_spiel/pull/319), and is not regularly tested nor maintained by the core team. This means that, at any time, it may not build or work as originally intended due to a change that will not have been caught by our tests. Hence, if bugs occur, please open an issue to let us know so we can fix them.
Unfortunately, we simple don't have the resources to maintain this, so you will have to customize it yourself (though I don't see why it wouldn't work with the correct version of LibTorch unless it's not backward-compatible).
We could use some help from the community, though! For example, things like this:
I had to also remove the dqn_torch examples and all references to this in every CMake file to successfully run all tests. Some files from those examples are trying to import a game for the test suite which isnt existing in this repo anymore.
Could I ask you to submit a PR with those fixes? That would not only help us out, but every other user would benefit as well. So we'd be quite grateful for the help.
CasparQuast
TheSQLGuru
Hello,
im working on a game implementation and want to train an Agent for the Game with the AlphaZero approach. I managed to compile and run tests with the workaround described https://github.com/google-deepmind/open_spiel/issues/966 there.
I had to also remove the dqn_torch examples and all references to this in every CMake file to successfully run all tests. Some files from those examples are trying to import a game for the test suite which isnt existing in this repo anymore.
My Question now is whether i can use newer CUDA drivers for example 12.2 with latest Cudnn. In the global_variables.sh i could only see the options 10.2 and lower which is very old. Did somebody test newer CUDA drivers?
According NVIDIA Toolkit theres not even a 10.2 Cuda version which is supported by Ubuntu 22 (which is the recommended OS for this project) just 18 and lower. https://developer.nvidia.com/cuda-10.2-download-archive?target_os=Linux&target_arch=x86_64&target_distro=Ubuntu
If this whole approach is just there to be viewed by externals and not really maintained, is there maybe a better recommended alternativ to using ./alphazero_torch_example ? My goal is just to use my RTX 3070TI with CUDA/Cudnn to accelerate the training process which seemed to be straight forward at first sight, but appears to be quite hard to even setup. Is there anyone who recently experimented with this or can help me with some tipps? Thanks in advance :+1: