<html xmlns:o="urn:schemas-microsoft-com:office:office" xmlns:w="urn:schemas-microsoft-com:office:word" xmlns:m="http://schemas.microsoft.com/office/2004/12/omml" xmlns="http://www.w3.org/TR/REC-html40"><head><meta http-equiv=Content-Type content="text/html; charset=utf-8"><meta name=Generator content="Microsoft Word 15 (filtered medium)"><style><!--
/* Font Definitions */
@font-face
{font-family:"Cambria Math";
panose-1:2 4 5 3 5 4 6 3 2 4;}
@font-face
{font-family:Calibri;
panose-1:2 15 5 2 2 2 4 3 2 4;}
/* Style Definitions */
p.MsoNormal, li.MsoNormal, div.MsoNormal
{margin:0cm;
margin-bottom:.0001pt;
font-size:11.0pt;
font-family:"Calibri",sans-serif;}
a:link, span.MsoHyperlink
{mso-style-priority:99;
color:blue;
text-decoration:underline;}
.MsoChpDefault
{mso-style-type:export-only;}
@page WordSection1
{size:612.0pt 792.0pt;
margin:72.0pt 72.0pt 72.0pt 72.0pt;}
div.WordSection1
{page:WordSection1;}
--></style></head><body lang=FI link=blue vlink="#954F72"><div class=WordSection1><p class=MsoNormal><span lang=EN-US>I think this is a good move. While you could probably fit enough examples per batch by using 2 or 3 times the cards, the half-precision performance of the V100s is a good deal better. Plus Puhti is a system we know we can run this on.<o:p></o:p></span></p><p class=MsoNormal><span lang=EN-US><o:p> </o:p></span></p><p class=MsoNormal><span lang=EN-US>-Antti</span></p><p class=MsoNormal><o:p> </o:p></p><div style='mso-element:para-border-div;border:none;border-top:solid #E1E1E1 1.0pt;padding:3.0pt 0cm 0cm 0cm'><p class=MsoNormal style='border:none;padding:0cm'><b>From: </b><a href="mailto:oe@ifi.uio.no">Stephan Oepen</a><br><b>Sent: </b>perjantai 31. tammikuuta 2020 19.53<br><b>To: </b><a href="mailto:sajvir@utu.fi">Antti Virtanen</a><br><b>Cc: </b><a href="mailto:andreku@ifi.uio.no">Andrei Kutuzov</a>; <a href="mailto:infrastructure@nlpl.eu">infrastructure</a>; <a href="mailto:outreach@nlpl.eu">outreach@nlpl.eu</a><br><b>Subject: </b>Re: rolling your own BERT (and maybe ELMo) on Saga</p></div><p class=MsoNormal><o:p> </o:p></p><p class=MsoNormal>belatedly, thanks for the link, antti!</p><p class=MsoNormal><o:p> </o:p></p><p class=MsoNormal>> Here's a (quick and dirty) repo for the code we used to train FinBERT: https://github.com/haamis/DeepLearningExamples_FinBERT/tree/master/TensorFlow/LanguageModeling/BERT_nonscaling. This one has the sbatch files used: https://github.com/haamis/BERT-pretraining</p><p class=MsoNormal><o:p> </o:p></p><p class=MsoNormal>the README claims that V100 gpus are required, whereas Saga only has</p><p class=MsoNormal>P100 cards. so i just gave up on the prospect of getting this to run</p><p class=MsoNormal>in norway and rather requested an allocation of billing units on Puhti</p><p class=MsoNormal>:-). i think for your presentation next week, you can just assume</p><p class=MsoNormal>that Saga is not a relevant target system for this work.</p><p class=MsoNormal><o:p> </o:p></p><p class=MsoNormal>looking forward to the tutorial! oe</p><p class=MsoNormal><o:p> </o:p></p></div></body></html>