[NLPL Task Force (A)] New accounts on Taito

[Joakim Nivre]

Thanks, Stephan. I have already forwarded your message to the “complainants”.

Joakim

From: Stephan Oepen <oe at ifi.uio.no>
Date: Friday, 1 February 2019 at 09:10
To: Joakim Nivre <joakim.nivre at lingfil.uu.se>
Cc: "infrastructure at nlpl.eu" <infrastructure at nlpl.eu>
Subject: Re: [NLPL Task Force (A)] New accounts on Taito

hi again, joakim,

please encourage your students to look into the latest PyTorch etc. versions on Abel.  they should not hesitate to make contact with ‘infrastructure’ for any missing pieces or technical problems!   there are also folks here in oslo who would like to start training custom ELMo networks.  if this software configuration plays out okay, it should be relatively straightforward to replicate these modules on Taito ...

cheers, oe


On Tue, 22 Jan 2019 at 16:52 Joakim Nivre <joakim.nivre at lingfil.uu.se<mailto:joakim.nivre at lingfil.uu.se>> wrote:
Apparently, one problem is that PyTorch is only available for Python 3.5 on Abel, while for example ELMo requires Python 3.6. However, this may be independent on the CPU/GPU issue and may not be any different on Taito. I will try to gather more information and keep you posted.

Joakim

On 2019-01-22, 11:45, "Stephan Oepen" <oe at ifi.uio.no<mailto:oe at ifi.uio.no>> wrote:

    dear joakim,

    thanks for the heads-up about incoming account requests!

    but what made you think that TensorFlow and PyTorch gpu support on
    Abel is lacking?  the NLPL installations of these frameworks
    (‘nlpl-tensorflow’ and ‘nlpl-pytorch’) should work transparently on
    both cpu and gpu nodes.  if there are issues with these modules,
    please encourage your students to contact us at
    ‘infrastructure at nlpl.eu<mailto:infrastructure at nlpl.eu>’.

    that being said, Taito has more and newer gpus (though also tends to
    have lots more gpu users :-), so having accounts on both sides is not
    a bad thing.  i believe the difference in gpu architecture (K20 on
    Abel vs. K80 and P100 on Taito) is not critical to what we typically
    do, but i recall tentative benchmarks of OpenNMT (PyTorch-based)
    achieving training throughputs around 50 percent higher on Taito than
    on Abel.

    if your students or others at uppsala come up with more detailed
    (contrastive) experience reports, i think we would be interested in
    that among the infrastructure task force too.

    best wishes, oe

    On Tue, Jan 22, 2019 at 9:14 AM Joakim Nivre <joakim.nivre at lingfil.uu.se<mailto:joakim.nivre at lingfil.uu.se>> wrote:
    >
    > Hi,
    >
    >
    >
    > We have two new research assistants, Johannes Gontrum and Elena Fano. They already have accounts on Abel (because they are also master’s students at Uppsala University), but they need access to Taito as well to be able to run Tensor Flow and/or PyTorch on GPUs (which I understand is still not possible on Abel).
    >
    >
    >
    > They will submit the form as instructed on http://wiki.nlpl.eu/index.php/Infrastructure/resources with the hope that Martin can approve their applications.
    >
    >
    >
    > Best,
    >
    > Joakim
    >
    >
    >
    >
    >
    >
    >
    >
    >
    > När du har kontakt med oss på Uppsala universitet med e-post så innebär det att vi behandlar dina personuppgifter. För att läsa mer om hur vi gör det kan du läsa här: http://www.uu.se/om-uu/dataskydd-personuppgifter/
    >
    > E-mailing Uppsala University means that we will process your personal data. For more information on how this is performed, please read here: http://www.uu.se/en/about-uu/data-protection-policy

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.nlpl.eu/archives/infrastructure/attachments/20190201/8ada5a42/attachment.htm>