I try to install pytorch 1.2.0 with anaconda, but found that torch.cuda.is_available() always returns False. Turns out our server environment can only support CUDA 9.0, but I do not have administrator right to upgrade the nvidia driver.
Give that pytorch 1.2.0 introduced Bool type, and I have already updated my code for corresponding change of API, I do not want to undo this efforts.
Will pytorch 1.2.0 support CUDA 9.0 backend in the future?
You could build PyTorch from source as described here and use your local CUDA installation to build it.
The binaries currently ship with CUDA9.2 and CUDA10.0.