Build issue on GCP about xla HOT 9 CLOSED

@farhaanbukhsh You can try either with export NO_CUDA=1 or with gcc 7 (instead of Clang).

from xla.

@asuhan I want cuda support should I try with gcc 7 ?

from xla.

@farhaanbukhsh You can try either with export NO_CUDA=1 or with gcc 7 (instead of Clang).

gcc7 doesn't work it fails the simple test of compiling the program, when I did a install

from xla.

I don't understand what failed -- compiling a simple C++ program with gcc or something else?

from xla.

This is what I got when I ran with export NO_CUDA=1 and COMPILE_PARALLEL=0 python setup.py install

from xla.

export CC=gcc CXX=g++-7 and COMPILE_PARALLEL=0 python setup.py install`

I realised i could be mistaken I should have used export CC=gcc-7 CXX=g++-7

from xla.

export CC=gcc-7 CXX=g++-7

from xla.

I suggest you using clang-7, as it is the one we used to work on and test internally.
But the blas error is likely due to a specific version of MKL liked by pytorch.
These are the versions a conda list shows me in my setup:

blas                      1.0                         mkl  
mkl                       2019.3                      199  
mkl-include               2019.1                      144  
mkl-service               1.1.2            py36he904b0f_5  
mkl_fft                   1.0.10           py36ha843d7b_0  
mkl_random                1.0.2            py36hd81dba3_0 

from xla.

In general if you want to build with both Cuda and XLA support, use gcc to build pytorch and clang7 to build torch_xla works well for me. If you only want XLA support clang7 should work for both by building pytorch with USE_CUDA=0.
As @dlibenzi said above, upgrading mkl version should fix the error shown in the last snapshot.
Closing issue as resolved, please feel free to reopen if you have follow up questions.

from xla.