<@tflink:fedora.im>
16:31:21
!startmeeting fedora-ai-ml-sig
<@nirik:matrix.scrye.com>
16:33:19
switch reboots, it's not online, but it should catch up when it gets reconnected
<@trix:fedora.im>
16:33:21
dang! now i can't bail.
<@tflink:fedora.im>
16:34:42
ok, thanks for the update
<@tflink:fedora.im>
16:34:45
!hi
<@trix:fedora.im>
16:35:30
i do need to go at top of hour.
<@tflink:fedora.im>
16:35:32
the question of whether we continue or not is still valid, though
<@trix:fedora.im>
16:36:42
how about i give an update ?
<@trix:fedora.im>
16:37:34
llvm18 changes/fixes are in rawhide, waiting on them in F41.
<@tflink:fedora.im>
16:38:34
sounds good
<@trix:fedora.im>
16:38:36
new 6.2 options in the compiler should make removing the rocm split possible, so i am trying that out.. changes are already up, building needs to happen.
<@tflink:fedora.im>
16:38:41
!topic packaging updates
<@trix:fedora.im>
16:39:09
so only 'default' gpu set remains, and its installed in the usual place.
<@tflink:fedora.im>
16:39:23
!info new options with 6.2 should make it possible to get rid of the packaging split in rocm (sub-packages and separate libraries for each shader family)
<@tflink:fedora.im>
16:39:58
that would be a great improvement in build time for both ROCm and the bits which use rocm
<@tflink:fedora.im>
16:40:15
have the llvm18 compat package issues with ROCm been resolved?
<@trix:fedora.im>
16:40:27
yes. indeed. until we add more gpu's because who doesn't want more gpus ??!?
<@trix:fedora.im>
16:40:58
the buildtime issue i believe are resolved in rawhide.. i do not know about runtime.
<@tflink:fedora.im>
16:41:02
yeah, deciding what ROCm is built for is going to be a constant source of fun, I think
<@trix:fedora.im>
16:41:22
still buildtime issues in F41
<@trix:fedora.im>
16:41:36
and there is a runtime issue in F40 .. with llvm17
<@tflink:fedora.im>
16:41:56
is that the blender issue?
<@trix:fedora.im>
16:42:01
yes
<@tflink:fedora.im>
16:42:24
!info the llvm18 compat package build issues in rawhide have been fixed
<@tflink:fedora.im>
16:42:37
!info llvm18 compat package build issues remain in F41
<@trix:fedora.im>
16:43:55
ai-wise, i am looking at rocm-ing llama-cpp and them ollama or whatever llama-cpp can plug into.
<@trix:fedora.im>
16:44:37
if someone else wants to do ollama that would be ok.
<@trix:fedora.im>
16:45:16
otherwise in my opinion is has risen up to 'valuable' package.
<@trix:fedora.im>
16:45:25
and i'll take a crack.
<@trix:fedora.im>
16:46:19
apu target 1035 for 680M and 1103 for 780M are now part of the default set.
<@tflink:fedora.im>
16:46:35
there's also the push for ramalama, not sure what their plans are WRT accelerators
<@trix:fedora.im>
16:47:24
building of the set is happening now.
<@otaylor:fedora.im>
16:47:38
ramalama ships its own image, so it just cares about what devices, etc, are exposed.
<@tflink:fedora.im>
16:48:00
so there are no plans to package ramalama in Fedora?
<@otaylor:fedora.im>
16:48:54
tflink: Not an expert here, but having just talked to the maintainer, I believe their is work to package the ramalama CLI in Fedora, but that's hardcoded to use a UBI9 based image
<@otaylor:fedora.im>
16:49:10
tflink: Not an expert here, but having just talked to the maintainer, I believe there is ongoing work to package the ramalama CLI in Fedora, but that's hardcoded to use a UBI9 based image
<@tflink:fedora.im>
16:49:25
good to know, thanks
<@tflink:fedora.im>
16:50:23
the set of ROCm packages? for F41?
<@trix:fedora.im>
16:51:28
The big changes are only for F42. F41 is waiting on llvm18, no system changes to do, because doing things big late is stooopid.
<@tflink:fedora.im>
16:52:20
ok, I wasn't sure which set you were referring to. is there a rebuild of F42 rocm packages going on?
<@trix:fedora.im>
16:52:55
yes. all the mathlib+ai libs need to be rebuilt to use the new default set of gpus.
<@tflink:fedora.im>
16:53:10
ah, ok. thanks for clarifying
<@tflink:fedora.im>
16:53:33
!info many F42 rocm packages are being rebuilt to add support for recent AMD APUs
<@tflink:fedora.im>
16:53:49
is there anything else you wanted to cover on this topic?
<@trix:fedora.im>
16:53:58
nope. carry on
<@tflink:fedora.im>
16:54:15
!topic Fedora ai-ml containers
<@tflink:fedora.im>
16:54:28
I know you listed it as rocm containers but I think it's also a general question
<@trix:fedora.im>
16:54:36
yes.
<@trix:fedora.im>
16:55:10
for things we can't full build in fedora like vLLM, how far can we get with containers ?
<@tflink:fedora.im>
16:55:37
ah, I misunderstood your question
<@trix:fedora.im>
16:56:13
my part would be to do the rocm bits of container things., i am just really clueless on containers.
<@tflink:fedora.im>
16:56:26
davdunc covered the fedora process for containers which is mostly constrained by Fedora packaging guidelines AFAIK - the contents of those containers would need to be Fedora packages
<@tflink:fedora.im>
16:56:40
what is the issue with vLLM?
<@trix:fedora.im>
16:57:09
wouldn't it be nice to have it running somewhere a fedora person could use it ?
<@tflink:fedora.im>
16:57:26
I mean what is the reason that vLLM can't be packaged for Fedora
<@tflink:fedora.im>
16:57:51
I assume that there is something that is causing it to be non-packagable other than -ENOTIME
<@trix:fedora.im>
16:57:54
oh. it has a very deep python depency tree. it could be done.
<@tflink:fedora.im>
16:58:00
ah, gotcha
<@trix:fedora.im>
16:58:04
just not by me.
<@tflink:fedora.im>
16:59:22
we could figure something out to make container images available but I honestly don't know if we could say they were "official" images or not. Conan Kudo or davdunc might know so we should probably ask them
<@trix:fedora.im>
17:00:09
ok.. well i need to drop!
<@tflink:fedora.im>
17:00:28
ok, we can continue the conversation later offline
<@tflink:fedora.im>
17:00:39
thanks for the updates and participating
<@tflink:fedora.im>
17:01:00
since there aren't really other people here, I'm going to end the meeting
<@tflink:fedora.im>
17:01:04
thanks for coming, everyone
<@tflink:fedora.im>
17:01:13
!endmeeting