Open johnmauff opened 1 year ago
@johnmauff thanks for the heads up. This isn't a compiler wrapper setting. CUDA_HOME
is set by the cuda module, which I would think users would not load for CPU-only workflows.
Before nvhpc/22.7, CUDA_HOME
was used to tell the toolkit to use a different version of CUDA than the default. Evidently they have changed that.
I'll add the new variable to the module. Hopefully when NVHPC_CUDA_HOME
is set, the compiler will not complain about the old one - it will be necessary to keep CUDA_HOME
set in the cuda modules both for older NVHPC versions and the many libraries that have adopted using that variable (like PyTorch and TensorFlow).
I am noticing that even when I am compiling CPU-only code using nvfortran, I am getting the following deprecated warning message:
nvfortran-Warning-CUDA_HOME has been deprecated. Please, use NVHPC_CUDA_HOME instead.
It might be confusing to users why a CUDA environment variable is set for CPU-only compilation