Frédéric Bastien

The ideal would be to wrap the cuDNN implementation of that. But this would take more time then wrapping the code in Pylearn2. But as said, it is low in...

I already saw this. Which GPU are in that box? I saw some GPU get underclocked and that stay like that. I never understood what was the problem. On Tue,...

If you use device=cuda for the new back-end, you have this printed during the import: Using cuDNN version 7001 on context None Mapped name None to device cuda: Tesla P100-PCIE-12GB...

Why? I think it was fixed. Do you use Theano dev version? Le 6 oct. 2016 12:57, "Erfan Noury" [email protected] a écrit : > I think using the default implementation...

You can put a symbolic batch size. That should work. Where do you put the ndim value? I don't see it in that interface. On Thu, Oct 6, 2016 at...

For opt 2 is this the get output of the input layer? What about putting it there? Le ven. 30 juin 2017 08:44, Jan Schlüter a écrit : > One...

Closed by https://github.com/tensorflow/tensorflow/pull/58638

I guess nvidia/cuda:12.1.1-cudnn8-devel-ubuntu20.04 should work if you update the JAX/JAXlib version installed.

The original issue was about sharing one GPU by multiple job on SLURM. Is this your case? It was clear. If it isn't the case, can you open a new...

For the original issue, the issue is that SLURM prevent that by default. This is a normal SLURM behavior. Otherwise, other users could just allocate on your GPUs and use...