build against CUDA 10.0 fails - incorrect path to libcublas library

System information

• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Linux Ubuntu 18.04.3
• TensorFlow version and how it was installed (source or binary): binary (Anaconda package tensorflow-2.0.0-gpu_py37h768510d_0)
• TensorFlow-Addons version and how it was installed (source or binary): source (a1a1e4c)
• Python version: 3.7.6
• Is GPU used? (yes/no): yes

Describe the bug

Attempting to build against CUDA 10.0 (which is installed in /usr/local/cuda) by entering “10.0” when prompted for the desired CUDA version by configure.sh and modifying requirements.txt to contain tensorflow>=2.0.0 results in the following exception when bazel build build_pip_pkg is run:

ERROR: Analysis of target '//:build_pip_pkg' failed; build aborted: no such package '@local_config_cuda//cuda': Traceback (most recent call last):
        File "/home/lebedov/tensorflow-addons/build_deps/toolchains/gpu/cuda_configure.bzl", line 1047
                _create_local_cuda_repository(<1 more arguments>)
        File "/home/lebedov/tensorflow-addons/build_deps/toolchains/gpu/cuda_configure.bzl", line 869, in _create_local_cuda_repository
                _find_libs(repository_ctx, <1 more arguments>)
        File "/home/lebedov/tensorflow-addons/build_deps/toolchains/gpu/cuda_configure.bzl", line 568, in _find_libs
                _find_cuda_lib("cublas", <4 more arguments>)
        File "/home/lebedov/tensorflow-addons/build_deps/toolchains/gpu/cuda_configure.bzl", line 529, in _find_cuda_lib
                find_lib(repository_ctx, <2 more arguments>)
        File "/home/lebedov/tensorflow-addons/build_deps/toolchains/gpu/cuda_configure.bzl", line 507, in find_lib
                auto_configure_fail(<1 more arguments>)
        File "/home/lebedov/tensorflow-addons/build_deps/toolchains/gpu/cuda_configure.bzl", line 261, in auto_configure_fail
                fail(<1 more arguments>)

Cuda Configuration Error: No library found under: /usr/local/cuda/lib64/libcublas.so.10

This is because CUDA 10.0 contains the lib symlinks /usr/local/cuda/lib64/libcublas.so and /usr/local/cuda/lib64/libcublas.so.10.0, but not /usr/local/cuda/lib64/libcublas.so.10. I’m specifically interested in using addons with tensorflow 2.0.0 because that is the latest version of the GPU-enabled packages provided by Anaconda as of today.

Code to reproduce the issue

See above.

Other info / logs

[lebedov]

@seanpmorgan I was able to build 0.6.0 against CUDA 10.0 with bazel 2.0.0 successfully, although I had to tweak build_deps/toolchains/gpu/cuda_configure.bzl to suppress the following exception:

ERROR: Analysis of target '//:build_pip_pkg' failed; build aborted: no such package '@local_config_cuda//cuda': Traceback (most recent call last):
        File "/home/lebedov/tensorflow-addons/build_deps/toolchains/gpu/cuda_configure.bzl", line 1329
                _create_local_cuda_repository(<1 more arguments>)
        File "/home/lebedov/tensorflow-addons/build_deps/toolchains/gpu/cuda_configure.bzl", line 1231, in _create_local_cuda_repository
                _host_compiler_includes(<2 more arguments>)
        File "/home/lebedov/tensorflow-addons/build_deps/toolchains/gpu/cuda_configure.bzl", line 357, in _host_compiler_includes
                get_cxx_inc_directories(<2 more arguments>)
        File "/home/lebedov/tensorflow-addons/build_deps/toolchains/gpu/cuda_configure.bzl", line 336, in get_cxx_inc_directories
                inc not in includes_cpp_set
argument of type 'depset' is not iterable. in operator only works on lists, tuples, dicts and strings.

I’m specifically interested in using the seq2seq code - as the seq2seq unit tests in 0.6.0 seem to pass with my custom build, I’m hoping that whatever remaining tf 2.0 compatibility issues still exist will not be a concern.

[cecabert]

I add the exact same error while trying to build for older cuda version (i.e. 9.2). I solve the issue by replacing cuda_config.cuda_major_version to cuda_config.cuda_version on the lines 573, 580, 587, 594 in the build_deps/toolchains/gpu/cuda_configure.bzl file, because the cuda library are named <libname>.so.<Major>.<Minor> and not <libname>.so.<Major>.