Compling source of PyTorch

[Predrag Punosevac]

Yichong Xu <yichongx at cs.cmu.edu> wrote:

> Hi Autonians,
> Has anyone got experience on installing PyTorch from source? I was able to install it but it said it was complied without cuDNN. Does anyone know if cuDNN is installed and where it is? Also it seems it does not work quite well with cuda 9.1????
> 

Ok I looked at this more carefully. cuDNN is a proprietary library which
requires NVidia account to be downloaded and must be installed manually.
That is not the problem. Bigger problem is that cuDNN has dual version
system. One number refers to the cuDNN library itself and the another is
refereed to CUDA version. I went carefully through GPU servers and
unfortunately

/usr/local/cuda 

simlinks to /usr/local/cuda-9.0 on some servers and on some servers 

/usr/local/cuda-9.1

All servers have at least partial installation of CUDA 9.1 but IIRC most
TensorFlow, Caffe, Teano users went back to 9.0 version. It is not
possible to install 9.0 version without at least partial 9.1
installation (using YUM and NVidia repos). 

Long story short this is not 5 minute job and we are risking major GPU
down time (just like the one from 2 weeks ago) if things are not done
properly. Properly would mean that one of PyTorch users who insists on
using cuDNN actually tries compiling PyTorch with CUDA-9.1 and cuDNN-7.1
then repeat the exercise with CUDA-9.0 and cuDNN-7.1, then repeat the
exercise with CUDA-9.1 and cuDNN-7.0 and finally repeat the exercise
with CUDA-9.0 and cuDNN-7.0. Only once we are 100% sure that one of
those 4 combinations works and doesn't break the machines for other
users we can go with this thing. Oh and don't forget that I an not
interested if this works on your favourity OS (mine is OpenBSD and this
definitly doesn't work). Testing needs to be done on Springdale 7.4
which we use in the lab.

In the mean time my bandwidth is limited in part due to the major
upgrade of the main file server and this high risk (many users could
be negatively affected) low reward (few people will benefit) operation 
is not my high priority.

Best,
Predrag


> Thanks in advance!
> 
> Thanks,
> Yichong
> 
> 
>