I want to know why there are two different Compute Platform installation options for “CUDA 11.7” and “CUDA 11.8” on the PyTorch official website.
As “CUDA 11.7” is known to be compatible with “CUDA 11.8”, what is the reason for releasing these two different versions of PyTorch?
Your answer and guidance will be appreciated!
Our current workflow is to release one “stable” binary package with a CUDA runtime (+ their corresponding libs) which were tested for a period of time in the nightly release and add another “experimental” binary with a newer CUDA runtime, which users can start using and should report any issues.
Once a newer version is available and no critical issues were reported for the “experimental” release, we drop the old “stable”, replace it with the old “experimental”, and add the newly built binary.
Thanks sincerely for your explanation!