I’m trying to setup a GPU cluster on using tf-nightly-gpu-2.0-preview
After some hacks (will document if someone is interested) I got to the point where I see the following error:
python3: Relink
/lib/x86_64-linux-gnu/libudev.so.1' with
/lib/x86_64-linux-gnu/librt.so.1’ for IFUNC symbol `clock_gettime’
Segmentation fault (core dumped)
I’ve learned from other sources that installing cudaDNN seems to solve the issue (for whatever reason).
It also seems that tf-nightly-gpu-2.0-preview is linked to CUDA 10.0 (and not the latest 10.1)
I’ve successfully? installed cuda-repo-ubuntu1804_10.0.130-1_amd64.deb
I’ve also successfully? installed libcudnn7_7.5.0.56-1+cuda10.0_amd64.deb
But on the website I only find
libcudnn7-dev_7.5.0.56-1+cuda10.1_amd64.deb
and
libcudnn7-dev_7.4.2.24-1+cuda10.0_amd64.deb
Installing libcudnn7-dev_7.4.2.24-1+cuda10.0_amd64.deb gives me:
libcudnn7-dev depends on libcudnn7 (= 7.4.2.24-1+cuda10.0); however:
Version of libcudnn7 on system is 7.5.0.56-1+cuda10.0.
Installing libcudnn7-dev_7.5.0.56-1+cuda10.1_amd64.deb gives me:
libcudnn7-dev depends on libcudnn7 (= 7.5.0.56-1+cuda10.1); however:
Version of libcudnn7 on system is 7.5.0.56-1+cuda10.0.
Any ideas?