I have successfully installed Tensorflow v1.12 (cloned github repo), Bazel 0.15 with CuDNN 7.4.1 and CUDA toolkit 10.0 inside my docker container.
I have HeliOS 7 in the container with GCC 4.8.5.
The target system I wish to deploy my container in has NVIDIA GPU Tesla P40 and V100.
I have installed the Nvidia driver version 410.48 in the container and the same Linux kernel versions same as the target host system.
For building tensorflow, I run the following -
…/bazel/bin/bazel build --config=opt --config=cuda //tensorflow:libtensorflow_cc.so --per_file_copt=//tensorflow/.*.cu.cc@-D_FORCE_INLINES --local_resources 2048,.5,1.0
I added “–per_file_copt=//tensorflow/.*.cu.cc@-D_FORCE_INLINES” to resolve nvcc error, and
added “–local_resources 2048,.5,1.0” to pass through memory limitations on my local system.
I am building this docker container on my Windows 10.
I still see the nvcc error - like this -
“’cicc’ died due to signal 9(Kill signal)”
In the compatibility table at - https://www.tensorflow.org/install/source,
there is no mention of compatibility for CUDA 10.0.
Can someone please confirm on Tensorflow and Bazel versions that will be successful with CUDA 10 ?
Or any idea why my bazel build keeps failing ?