<@tflink:fedora.im>
16:30:15
!startmeeting ai-ml-sig
<@meetbot:fedora.im>
16:30:16
Meeting started at 2025-05-08 16:30:15 UTC
<@meetbot:fedora.im>
16:30:16
The Meeting name is 'ai-ml-sig'
<@tflink:fedora.im>
16:30:32
who all is here for the ai-ml-sig meeting?
<@tflink:fedora.im>
16:30:34
!hello
<@zodbot:fedora.im>
16:30:35
Tim Flink (tflink)
<@mystro256:fedora.im>
16:32:34
!hello
<@zodbot:fedora.im>
16:32:35
None (mystro256)
<@trix:fedora.im>
16:34:34
!hello
<@zodbot:fedora.im>
16:34:35
Tom Rix (trix)
<@trix:fedora.im>
16:34:50
sorry for being late, give me all the action items
<@tflink:fedora.im>
16:35:33
not going to argue about that one :)
<@tflink:fedora.im>
16:35:45
anyhow, lets get this party started
<@tflink:fedora.im>
16:35:54
!topic ROCm 6.4 update
<@trix:fedora.im>
16:36:45
Thanks Jeremy Newton for getting the update out, i _think_ we are in pretty good shape.
<@tflink:fedora.im>
16:37:25
yeah, things seem to be working from the bits of testing I've done
<@trix:fedora.im>
16:37:26
its all built, i am taking care of rebuilding and testing pytorch bits now
<@tflink:fedora.im>
16:38:21
which also seem to be working for at least 2 of the 3 ISAs that it's built for
<@mystro256:fedora.im>
16:38:59
Team effort, no worries
<@trix:fedora.im>
16:39:06
get more folks to tests their boards would be helpful, we added a lot of isa's and i just have a couple of them.
<@trix:fedora.im>
16:39:43
the ones i have a gfx1100, gfx1201 and gfx1151
<@tflink:fedora.im>
16:40:00
note that pytorch 2.7 in rawhide is only going to work with gfx1100, gfx1101 and gfx90a at the moment
<@trix:fedora.im>
16:40:21
oh sorry, i don't have gfx1100 anymore.. or its on the shelf,
<@trix:fedora.im>
16:40:48
is that a hipblaslt issue ?
<@tflink:fedora.im>
16:40:52
yeah
<@trix:fedora.im>
16:41:12
there should be a runtime check, that disable hipblaslt
<@tflink:fedora.im>
16:41:31
I assume that it won't work for other ISAs. pytorch was blowing up on gfx1100 until I rebuilt it with gfx1100;gfx1101 support
<@trix:fedora.im>
16:41:49
sorry, been busy.. it seems to work on the 1201 and 1151, but maybe not ?
<@tflink:fedora.im>
16:42:24
I only have 1100 and 1101, can't speak to the others :)
<@trix:fedora.im>
16:42:53
it would be helpful if fedora had some qa :(...
<@trix:fedora.im>
16:43:04
i am still just manually doing things.
<@tflink:fedora.im>
16:43:32
on that note, I have a few more of the test subpackages working
<@trix:fedora.im>
16:43:41
sweeeeet!
<@tflink:fedora.im>
16:43:47
https://copr.fedorainfracloud.org/coprs/tflink/rocm-next-test/packages/
<@tflink:fedora.im>
16:43:55
it's not much yet but I'm working on them as I have time
<@trix:fedora.im>
16:44:24
good enough that maybe we turn them on in the default build ?
<@tflink:fedora.im>
16:44:39
no, they can't build in koji - several of them require network
<@trix:fedora.im>
16:45:12
booo.. ok i know rocsparse is a dog about downloading a lot of stuff.
<@trix:fedora.im>
16:45:44
turn them on in the default copr ? RH ?
<@tflink:fedora.im>
16:45:48
yeah, that downloads a ton of test files but I don't think it's the only one
<@tflink:fedora.im>
16:46:21
at some point, maybe. until they're more done, I figure it makes sense just to keep them in the copr I've been working on
<@tflink:fedora.im>
16:46:34
until more of them are done, anyways
<@trix:fedora.im>
16:46:43
cool. maybe we point folks at it somewhere ?
<@tflink:fedora.im>
16:47:13
as in writing docs?
<@trix:fedora.im>
16:48:00
oh jeeze not that!! maybe just a cross link of the copr's we have, 'look here ... ' for the test copr
<@trix:fedora.im>
16:48:38
just a thought, its more important imo to have the tests than document the tests.
<@tflink:fedora.im>
16:49:33
either way is fine with me. I'm not terribly optimistic that many folks will run the tests, especially before I get the running/setup scripts done
<@tflink:fedora.im>
16:50:59
!info ROCm 6.4 is in rawhide
<@tflink:fedora.im>
16:51:08
!info pytorch 2.7 is in rawhide
<@trix:fedora.im>
16:51:43
2.7 is there, and i am refreshing all of the other torch* packages atm, currently fighting with triton.
<@trix:fedora.im>
16:52:38
these should get enough attention next week to have them in good shape, if there are no big issues.
<@trix:fedora.im>
16:53:22
i still have no way to test aarch64. so that i won't be testing.
<@tflink:fedora.im>
16:53:23
!info other torch packages (triton et. al) are still in progress, should be finished soon
<@trix:fedora.im>
16:54:53
open floor ?
<@tflink:fedora.im>
16:55:11
is there anything else on the updates?
<@trix:fedora.im>
16:55:46
just i will be away end of may. so i am hustling to get these in good shape.
<@tflink:fedora.im>
16:56:30
did we ever set up the rocm packages to have the rocm-packagers-sig as the default asignee instead of just the owner of each package?
<@tflink:fedora.im>
16:56:36
for bugzilla
<@trix:fedora.im>
16:57:16
that would be good, as i don't want things orphaned because i took a vacation.
<@tflink:fedora.im>
16:57:25
yeah, that's what brought it to mind
<@tflink:fedora.im>
16:57:48
we can sync up on that after the meeting since in practice, you'll have to do most of that work
<@trix:fedora.im>
16:58:04
all the actions are mine weeeeee
<@trix:fedora.im>
16:58:16
yes, sounds good.
<@man2dev:fedora.im>
16:58:19
😂😂
<@tflink:fedora.im>
16:58:31
on the bright side, it should be pretty simple - just a mind numbing process of clicking the same thing on a bunch of different pages
<@tflink:fedora.im>
16:58:37
anyhow, we can move on to
<@tflink:fedora.im>
16:58:42
!topic open floor
<@tflink:fedora.im>
16:58:59
are there any other topics that folks want to bring up?
<@tflink:fedora.im>
17:05:14
ok. thanks for coming, everyone
<@trix:fedora.im>
17:05:23
o///
<@tflink:fedora.im>
17:05:26
!endmeeting