Editing Cupy installation instructions #47
Conversation
Most users will not want to compile Cupy from source. I have expanded the Cupy installation guide to explain how to get the version appropriate for your system.
| @@ -61,10 +61,15 @@ The code also requires a working version of glue, supporting `glue.ligolw.ligolw | |||
| The code uses cupy to access GPUs. If you don't have one, the code will still work. | |||
| If you do need one, make sure to install cupy **on a machine that supports GPUs ** | |||
|
|
|||
| You probably do not want to pip install `cupy` directly, but rather a pre-compiled version for your system's CUDA version. To find your system's version (assuming it's installed) run `nvidia-smi`, and look for "CUDA Version" in the top-right. For version X.Y, the appropriate cupy package will be called `cupy-cudaXY` (e.g., if you have CUDA 11.0, install `cupy-cuda110`) | |||
There was a problem hiding this comment.
This is strictly speaking not accurate. CUDA Version here refers to the maximal CUDA Toolkit version supported by the driver version reported on its left, but what you really need is the runtime library version installed locally, which can be decoupled/fall behind this version (say you have CUDA 10.2 installed, but you upgrade the driver to 11.1). Perhaps it's better to use nvcc --version if there is a local installation available to check.
FYI in some environments where there is no CUDA Toolkit installed locally, but has the driver installed, you can install CuPy from Conda-Forge:
conda install -c conda-forge cupy cudatoolkit=X.Y
X.Y is again up to the driver version. The cudatoolkit contains all CUDA runtime libraries (but no nvcc or the compiler toolchain).
Most users will not want to compile Cupy from source. I have expanded the Cupy installation guide to explain how to get the version appropriate for your system.