[theano-users] Problem with installation of gpuarray and with try_blas in Mac OS 10.12 Sierra

[Orlando Bisacchi Coelho]

I am installing Theano in a Mac running Sierra. My python installation
comes from the last Anaconda release. Theano runs fine in the CPU but not
in the GPU.

During installation of gpuarray I came across some errors but using the
workaround python setup.py build_ext -I /usr/local/include -L /usr/local/
lib sorted it out.

When I try to run Theano in the GPU I get an OsX generated window that says:

try_blas_n_NNN cannot be opened because of a problem. [NNN is a
alphanumeric code; every Theano run generates an error message with a
different code.]
Check with the developer to make sure try_blas_n_NNN works wit this version
of MacOS. [Etc]

I did not install BLAS by myself. So I understand that the version of BLAS
in use is the one provided by Apple's Accelerate framework. And that, on
the other hand, python uses Intel's MLK library. But I can't figure out
where exactly try_blas comes from.

I've run that exhaustive Theano test. It generated the following errors.
The path error is trivial to solve of course. Theano is not happy about
gpuarray and I wonder if the problem is BLAS-related. But I have no clue on
how to solve it. Can anyone give me a hint? Many thanks in advance.

Orlando

ERROR (theano.sandbox.cuda): nvcc compiler not found on $PATH. Check your
nvcc installation and try again.


ERROR (theano.sandbox.cuda): nvcc compiler not found on $PATH. Check your
nvcc installation and try again.

[repeated several times]


ERROR (theano.gof.opt): EquilibriumOptimizer max'ed out by
'local_upcast_elemwise_constant_inputs'. You can safely raise the current
threshold of 5.000000 with the theano flag 'optdb.max_use_ratio'.


ERROR (theano.gof.opt): EquilibriumOptimizer max'ed out by
'local_upcast_elemwise_constant_inputs'. You can safely raise the current
threshold of 5.000000 with the theano flag 'optdb.max_use_ratio'.

[repeated several times]


ERROR: Failure: RuntimeError (('Wrong major API version for gpuarray:', 1,
'Make sure Theano and libgpuarray/pygpu are in sync.'))


RuntimeError: ('Wrong major API version for gpuarray:', 1, 'Make sure
Theano and libgpuarray/pygpu are in sync.')

[repeated several times]


ERROR: Failure: RuntimeError (('Wrong major API version for gpuarray:', 1,
'Make sure Theano and libgpuarray/pygpu are in sync.'))

[repeated several times]


RuntimeError: ('Wrong major API version for gpuarray:', 1, 'Make sure
Theano and libgpuarray/pygpu are in sync.')

[repeated several times]


ERROR: Only test when pygpu isn't


ERROR: Failure: RuntimeError (('Wrong major API version for gpuarray:', 1,
'Make sure Theano and libgpuarray/pygpu are in sync.'))


ERROR: Verify that the random numbers generated by GPUA_mrg_uniform,
serially,


RuntimeError: ('Wrong major API version for gpuarray:', 1, 'Make sure
Theano and libgpuarray/pygpu are in sync.')


ERROR: theano.sandbox.tests.test_rng_mrg.test_overflow_gpu_new_backend


RuntimeError: ('Wrong major API version for gpuarray:', 1, 'Make sure
Theano and libgpuarray/pygpu are in sync.')


----------------------------------------------------------------------

Ran 3002 tests in 36433.657s


FAILED (SKIP=89, errors=16)