TensorFlow 1.6.0 fix: use the absolute path for the C compiler when compiling with GPU support#1386
Conversation
|
Hmm, I think this reverts something that @akesandgren did as a part of the sprint during the EasyBuild User Meeting in Amsterdam (see commit history in #1287)... Why didn't this pop up earlier? Is this specific to TF 1.6.0? |
|
Also, what happens when |
|
It doesn't touch the things i added as far as i can see. I don't really see how doing this change would make a difference. What is the actual problem or error output from not doing this? |
|
@akesandgren this is the error I get. |
|
@boegel concerning compiler wrappers, it would indeed be good if someone tested that. |
|
There might be another way to solve this, but i have to look at that later today. |
| config_env_vars.update({ | ||
| 'CUDA_TOOLKIT_PATH': cuda_root, | ||
| 'GCC_HOST_COMPILER_PATH': which(os.getenv('CC')), | ||
| 'GCC_HOST_COMPILER_PATH': os.path.realpath(which(os.getenv('CC'))), |
There was a problem hiding this comment.
@smoors Please use the resolve_path function from filetools (which has error handling)
|
Doing a build of 1.6.0 with goolfc, will report back. |
|
I have a different change that at least solved the problem on our system. |
|
@akesandgren You should probably make that
|
|
@akesandgren ok, will test. |
|
the solution of @akesandgren seems to work and is probably safer wrt wrappers, so let's use it. |
|
Thanks @smoors and @akesandgren! |
This is needed for the
bazel buildcommand with the option--config=cudaInspiration for this fix was found here:
tensorflow/tensorflow#14380