Just a small comment, I didn't try that driver. We use 367.57 and 367.44 and they work well.
Can you try a newer 375 driver? Which is the exact driver that you use? Anyway, I don't think someone can double check here as holidays are close. If someone else have this problem or have this working, tell us. Was your Theano cache empty before using the 375 driver? Be sure to use the recommanded driver by nvidia for your GPU. I never understoold the driver version number from NVIDIA, maybe the 375 is the old one this test was mean to catch the bug. Fred On Thu, Dec 15, 2016 at 5:31 AM, V <[email protected]> wrote: > Hi, > > Why doesn't Theano work with new CUDA drivers? (newer than version 375) > > It fails at > python2.7/site-packages/theano/sandbox/cuda/tests/test_driver.py > > Doing the test: > > import numpy; >> import theano; >> import theano.sandbox.cuda as cuda >> mode_with_gpu = theano.compile.mode.get_default_mode().including('gpu') >> a = numpy.random.rand(10000).astype("float32") >> A = cuda.shared_constructor(a) >> f = theano.function(inputs=[], outputs=A.sum(), mode=mode_with_gpu, >> profile=False) >> print a.sum(); >> print A.sum(); >> print f(); >> > > With driver version < 375 the results match: > > 4979.78 > Sum{acc_dtype=float64}.0 > 4979.77539062 > > But with the new driver they don't: > > 5019.74 > Sum{acc_dtype=float64}.0 > 1402.4197998 <(402)%20419-7998> > > Is it a bug in the driver? > > Thanks! > > -- > > --- > You received this message because you are subscribed to the Google Groups > "theano-users" group. > To unsubscribe from this group and stop receiving emails from it, send an > email to [email protected]. > For more options, visit https://groups.google.com/d/optout. > -- --- You received this message because you are subscribed to the Google Groups "theano-users" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. For more options, visit https://groups.google.com/d/optout.
