[theano-users] How to select certain cudnn version (I have mulitple cudnn in my cuda_root folder) when I compile theano

[w***@gmail.com]

Hi,
I have multiple version of cudnn in my CUDA_ROOT folder, and I want to
compile theano with certain one. How can I do that? Currently, I have the
following error "Tesla K80 (CNMeM is enabled with initial size: 95.0% of
memory, cuDNN Mixed dnn version. The header is from one version, but we
link with a different version (5105, 4007))" It seems that I compile theano
with an older version of cudnn.

Thanks.

[J Zam]

Hi,

It seems that there are multiple versions of cudann loaded at runtime. You
can check which modules are loaded from the terminal by typing:

module list

You should see that there are multiple cudann modules loaded and theano is
seeing all of them, that's why you get the "cuDNN Mixed dnn version"
message. You could unload a version by typing:

module unload "your particular cudann module"

Unload all until you are left with one, but be careful too test it again so
that your theano version is compatible with the cudann version you are
loading.


Jose

Post by w***@gmail.com
Hi,
I have multiple version of cudnn in my CUDA_ROOT folder, and I want to
compile theano with certain one. How can I do that? Currently, I have the
following error "Tesla K80 (CNMeM is enabled with initial size: 95.0% of
memory, cuDNN Mixed dnn version. The header is from one version, but we
link with a different version (5105, 4007))" It seems that I compile theano
with an older version of cudnn.
Thanks.

[w***@gmail.com]

Thank you for your reply. So I should type "module list" under the terminal
of python or just on the command? It seems that module list is not a
command in Linux.

Post by J Zam
Hi,
It seems that there are multiple versions of cudann loaded at runtime. You
module list
You should see that there are multiple cudann modules loaded and theano is
seeing all of them, that's why you get the "cuDNN Mixed dnn version"
module unload "your particular cudann module"
Unload all until you are left with one, but be careful too test it again
so that your theano version is compatible with the cudann version you are
loading.
Jose

Post by w***@gmail.com
Hi,
I have multiple version of cudnn in my CUDA_ROOT folder, and I want
to compile theano with certain one. How can I do that? Currently, I have
the following error "Tesla K80 (CNMeM is enabled with initial size: 95.0%
of memory, cuDNN Mixed dnn version. The header is from one version, but we
link with a different version (5105, 4007))" It seems that I compile theano
with an older version of cudnn.
Thanks.

[J Zam]

Hi,

Yes, sorry "module list" might be specific to distributions, depending on
yours you might have to use modprobe as shown here:

http://www.thegeekstuff.com/2010/11/modprobe-command-examples/?utm_source=feedburner

Jose

Post by w***@gmail.com
Thank you for your reply. So I should type "module list" under the
terminal of python or just on the command? It seems that module list is
not a command in Linux.

Post by J Zam
Hi,
It seems that there are multiple versions of cudann loaded at runtime.
module list
You should see that there are multiple cudann modules loaded and theano
is seeing all of them, that's why you get the "cuDNN Mixed dnn version"
module unload "your particular cudann module"
Unload all until you are left with one, but be careful too test it again
so that your theano version is compatible with the cudann version you are
loading.
Jose

Post by w***@gmail.com
Hi,
I have multiple version of cudnn in my CUDA_ROOT folder, and I want
to compile theano with certain one. How can I do that? Currently, I have
the following error "Tesla K80 (CNMeM is enabled with initial size: 95.0%
of memory, cuDNN Mixed dnn version. The header is from one version, but we
link with a different version (5105, 4007))" It seems that I compile theano
with an older version of cudnn.
Thanks.

[z***@gmail.com]

Hi, I have the same question as you. How did you solve your problem at the
end? Looking forward to your reply!

Post by w***@gmail.com
Hi,
I have multiple version of cudnn in my CUDA_ROOT folder, and I want to
compile theano with certain one. How can I do that? Currently, I have the
following error "Tesla K80 (CNMeM is enabled with initial size: 95.0% of
memory, cuDNN Mixed dnn version. The header is from one version, but we
link with a different version (5105, 4007))" It seems that I compile theano
with an older version of cudnn.
Thanks.