Caffe, Theano, Torch, TensorFlow on GPU10

[Predrag Punosevac]

I would prefer that we try with with CUDA 10.1. CUDA 9 branch is
considered retired by NVidia. If I am to break things which work now
and reboot machines I would like to get similar mileage as I did this
the last time around  (GPU 1-9 have uptime of nearly 300 days).

Predrag

On Tue, May 14, 2019 at 5:43 PM Abhay Gupta <abhayg at andrew.cmu.edu> wrote:
>
> Hi Predrag,
>
> Pytorch does work fine on other GPUs. Is it possible to have both cuda 9.1 and cuda 9.2 installed on the remaining GPUs? Currently, I think the GPUs (2-9) have only 9.1 and gpu1 has 8.0 and 9.2 also.
>
> On Tue, May 14, 2019 at 5:39 PM Predrag Punosevac <predragp at andrew.cmu.edu> wrote:
>>
>> Thanks for the report. Does PyTorch work on all other GPUs? This is an
>> opportunity for me to go and fix the things. Anybody else?
>>
>> Predrag
>>
>> On Tue, May 14, 2019 at 4:22 PM Emre Yolcu <eyolcu at andrew.cmu.edu> wrote:
>> >
>> > PyTorch is working fine on gpu10. Thank you very much Predrag.
>> >
>> >
>> >
>> > Emre
>> >
>> >
>> >
>> > From: Predrag Punosevac
>> > Sent: Monday, May 13, 2019 4:24 PM
>> > To: users at autonlab.org
>> > Subject: Caffe, Theano, Torch, TensorFlow on GPU10
>> >
>> >
>> >
>> > Dear Autonians,
>> >
>> >
>> >
>> > I have just upgraded NVidia driver on GPU10 to 418.74 as well as
>> >
>> > libcudnn7 to 7.5.1.10-1 which is compatible with CUDA 10.1 (the
>> >
>> > default on this server). I have also upgrade MATLAB to 2019a relase.
>> >
>> >
>> >
>> > nvidia-smi works as expected as well as MATLAB. gpuDevice() which was
>> >
>> > broken on the previous release of MATLAB now correctly shows four
>> >
>> > available GeForce GTX 1081 Ti cards.
>> >
>> >
>> >
>> > I would like to help the Auton Lab deep learning community and fix
>> >
>> > Caffe, Theano, Torch/PyTorch, TensorFlow or whatever are people using
>> >
>> > these days. I need you to log into the server and try to install your
>> >
>> > usual tool and test it.
>> >
>> >
>> >
>> > I would also appreciate feedback about other 8 gpu servers(GPU7 is
>> >
>> > reserved for the special project). I know at the very least that GPU3
>> >
>> > scratch directory is filled. I am aware of a dead GPU card on GPU4
>> >
>> > computing node (I have 1081 Ti on my desk which I can use instead of
>> >
>> > the dead Titan X on that server). GPU3 perhaps has another dead GPU
>> >
>> > card.
>> >
>> >
>> >
>> > Most Kind Regards,
>> >
>> > Predrag Punosevac
>> >
>> >
>
>
>
> --
> Regards,
> Abhay Gupta