IRC channel logs


back to list of logs

<civodul>hey zimoun!
<civodul>ACTION just sent an activity report email to the usual suspects
<zimoun>civodul: cool!
<zimoun>rekado: in case you missed this Mumi bug report #54204
<rekado>zimoun: I can no longer reproduce the main claim
<rekado>“mumi date:2022-02-01..2022-02-28” does in fact return the two issues mentioned
<rekado>we recently changed filtering
<rekado>previously we would get up to 500 or so matches from xapian and then filter them down
<rekado>that would sometimes lead to the intersection being empty
<rekado>now we do more on the xapian side of things because recent versions of the guile bindings gained some more features
<zimoun>Cool! I have seen many updates (guile-xapian and mumi) but haven’t had thechance to test. :-)
<rekado>our keras package is very old
<rekado>this is now a problem because this old version of keras (and also tensorflow) don’t work with h5py 3
<rekado>can’t upgrade keras because they switched to bazel
<rekado>can’t upgrade tensorflow because of bazel
<rekado>so… we gotta get bazel
<rekado>I really don’t want to work on it, but I think we haven’t got much of a choice
<civodul>any hope keras can be built in some other way?
<rekado>with Python (and any other language that forces us to propagate packages) it’s not easy
<rekado>we can introduce h5py 2.10 and make keras use that
<rekado>but anyone using the regular h5py will have a bad time
<rekado>I wish we had a sane way to do that sort of multi-version loading
<rekado>(I also hope Guile could do better in this regard.)
<rekado>forcefully rename h5py 2.10 to h5py2 and … somehow … rename all uses of “h5py” to “h5py2” in the keras closure…?
<rekado>that would be just keras and python-keras-applications; luckily nowhere else in the closure
<rekado>but I really think we need to tackle the bigger problem: bazel
<civodul>bah, doesn't sound like great news
<civodul>i wonder if the Bazel folks are still coming to the R-B summit
<toddgamblin>@civodul: I've been sent by Kenneth to pester you to respond to his question about making your HPC devroom talk a lightning talk
<toddgamblin>sorry for crashing #guix-hpc :)
<civodul>hey, hi toddgamblin! :-)
<civodul>i think i replied some time ago?
<civodul>that was before holidays though
<civodul>right now i'm processing my personal email backlog, so it could be i missed something too :-)
<civodul>lemme see
<toddgamblin>it could also be that he set the reply-to to the devroom reflector again -- there were issues with that email address
<toddgamblin>if you responded, maybe respond again directly to kenneth... I'll also just tell him you already responded
<toddgamblin>also it'll be nice to see you all again at fosdem!