CUDA_HOME error when installing gsplat v1.4

[MartinEthier]

I have a conda env that has pytorch 2.1 installed with cuda 11.7. I am able to pip install gsplat v1.3 or older without any issues. When I try pip installing gsplat v1.4, I get the same error as #197. Error: "OSError: CUDA_HOME environment variable is not set. Please set it to your CUDA install root.". His issue was that he did not have cuda installed with pytorch but I have it installed properly and I am able to use pytorch with my GPU using this env.

[FantasticOven2]

Hey, can you try to run nvcc --version in your conda env and what's the output?

[martinResearch]

I believe the cause could be that since version1.4 the whl file is not uploaded on pypi.org anymore and only the source distribution is released on pypi.org (see https://pypi.org/project/gsplat/#files). As a result pip install will use the source distribution and try to compile gsplat during the installation (while it was using the whl file and compiling at runtime the first time gsplat is used in previous versions). You will need to install cuda ans setup the CUDA_PATH variable in your environment before installing gsplat if you want to install from source. This could be made clearer in the doc. Alternatively you can install from the pre-compiled wheels as explained int the readme but only wheel for python 3.10 are released at the moment

[martinResearch]

Notre that the following doc is outdated.

The easiest way is to install from PyPI. In this way it will build the CUDA code on the first run (JIT).
pip install gsplat

because this will try to install from source and compile the cuda code during the installation since version1.4