TeeratP
(Teerat Pipitcharulerd)
December 28, 2017, 8:06am
1
My company’s GPU cluster is running on CUDA 8.0 cuDNN 6, which I shared with the other teams which use tensorflow for their project.
I find that the default release is CUDA 8.0 cuDNN 7, so I wonder if I can get pytorch 0.3 CUDA 8.0 cuDNN 6?
Thank you in advance
smth
December 28, 2017, 11:10am
2
pytorch comes with it’s own CUDA and CuDNN files, so regardless of what’s on the cluster, the pytorch install will work.
If you so choose to want to compile pytorch with CuDNN6, you can compile from source.
1 Like
TeeratP
(Teerat Pipitcharulerd)
December 29, 2017, 2:54am
3
Hi smth
I got this error
RuntimeError: cuDNN version mismatch: PyTorch was compiled against 7003 but linked against 6021
So, I believe the problem is due to difference between cuDNN version required by pytorch and version installed on machine.
I check this link https://github.com/pytorch/pytorch#from-source and I tell me to install CUDA and cuDNN separately.
Can you give me some guidance on how to “compile pytorch with CuDNN6”
Thank you
smth
December 29, 2017, 9:58am
4
you must have $LD_LIBRARY_PATH environment variable set that has libcudnn v6. Unset it.