2024-01-18 14:03:32 <@tflink:fedora.im> !startmeeting pytorch
2024-01-18 14:03:33 <@meetbot:fedora.im> Meeting started at 2024-01-18 14:03:32 UTC
2024-01-18 14:03:33 <@meetbot:fedora.im> The Meeting name is 'pytorch'
2024-01-18 14:03:45 <@tflink:fedora.im> !topic welcome
2024-01-18 14:03:53 <@martinsk0801:matrix.org> Thanks!
2024-01-18 14:04:31 <@martinsk0801:matrix.org> Yeah, i saw it in the chat. It is cool.
2024-01-18 14:04:57 <@martinsk0801:matrix.org> And thanks for the meeting notes that captured
2024-01-18 14:05:39 <@tflink:fedora.im> !chair Teng Ma
2024-01-18 14:05:50 <@tflink:fedora.im> maybe not, the bot didn't react
2024-01-18 14:05:57 <@martinsk0801:matrix.org> Hi devashar , welcome to the discussion
2024-01-18 14:06:16 <@martinsk0801:matrix.org> I think that is okay.
2024-01-18 14:06:28 <@devashar:matrix.org> Hi Teng. Thank you
2024-01-18 14:07:14 <@martinsk0801:matrix.org> Check the time here. It is 9:07 ET. I think we can get started.
2024-01-18 14:07:46 <@trix:fedora.im> a topic i want to discuss is to go over the desired in-pytorch features and see if folks want to add/do some
2024-01-18 14:07:47 <@tflink:fedora.im> sounds good, we have at least a few people and it sounds like kaitlyn is not going to make it today
2024-01-18 14:08:25 <@trix:fedora.im> i am going to be busy with rocm for a while, what else do we want ?
2024-01-18 14:08:49 <@tflink:fedora.im> are we jumping right into that topic?
2024-01-18 14:09:16 <@martinsk0801:matrix.org> Yeah, Kaitlyn wont be able to join. We can start with that
2024-01-18 14:09:34 <@tflink:fedora.im> !topic desired pytorch features
2024-01-18 14:09:38 <@trix:fedora.im> i proposed this in the ai/ml channel yesterday.. hopefully folks have their lists :)
2024-01-18 14:09:55 <@martinsk0801:matrix.org> The only two other topics that I want to cover is - the legal discussion regarding including AI models. 
2024-01-18 14:10:10 <@tflink:fedora.im> !info pytorch has been packaged for fedora but right now, it has the bare-minimum of features
2024-01-18 14:10:36 <@martinsk0801:matrix.org> And get your opinion on if we should start packaging while waiting on the legal opinion to happen
2024-01-18 14:10:36 <@tflink:fedora.im> other than rocm, I don't have much personally
2024-01-18 14:11:25 <@tflink:fedora.im> is there a list of features which have yet to be enabled?
2024-01-18 14:12:03 <@trix:fedora.im> in the nitty-gritty, its the cmake options
2024-01-18 14:12:41 <@trix:fedora.im> the ones i am working on now in priority order are rocm, caffe, openmp. 
2024-01-18 14:13:33 <@trix:fedora.im> some that come to mind are
2024-01-18 14:13:48 <@trix:fedora.im> ditributed, cuda, oneapi, opencl
2024-01-18 14:14:10 <@tflink:fedora.im> !info currently, the features that will be enabled soon are: rocm, caffe and openmp
2024-01-18 14:15:02 <@tflink:fedora.im> there have already been requests for cuda but that'll be a legal question for the scope of what we're talkinga bout
2024-01-18 14:15:44 <@trix:fedora.im> my thoughts are we add the logic to the spec file with bcond logic to allow someone else to build it
2024-01-18 14:17:00 <@tflink:fedora.im> did we ever answer the question of whether it would be possible to include other packages dynamically, allow someone to do the CUDA bits in rpmfusion without having to redo all of torch?
2024-01-18 14:17:14 <@trix:fedora.im> i don't mind testing the logic if we had some instructions on where to get the cuda bits from documented on the sig page
2024-01-18 14:17:39 <@trix:fedora.im> yes this would allow rpmfusion to just use it
2024-01-18 14:18:18 <@martinsk0801:matrix.org> Do we want to also provide some guidance to the packagers for each feature on things like license/version compatibility or dependencies etc?
2024-01-18 14:18:57 <@trix:fedora.im> yes.  the thing we need it the buildrequire version > && version < info
2024-01-18 14:19:17 <@trix:fedora.im> this will be a pita to collect and is exactly what folks doing it themselves need
2024-01-18 14:20:37 <@trix:fedora.im> i'm not going to be the guy to do this, its real work ;)
2024-01-18 14:21:25 <@tflink:fedora.im> yeah, I'm not terribly interested in that work either
2024-01-18 14:22:14 <@martinsk0801:matrix.org> Can we talk about this work? Maybe we can see any volunteers from SIG and make it dived and conquered with a bit more clear instructions?
2024-01-18 14:22:18 <@tflink:fedora.im> the only things that come to mind other than what you mentioned is vulkan unless opencl is doing what I think that does
2024-01-18 14:22:50 <@tflink:fedora.im> I'd rather see the volunteers who understand the level of work that would entail first, honestly
2024-01-18 14:22:56 <@martinsk0801:matrix.org> If you all feel this work is adding value
2024-01-18 14:23:14 <@martinsk0801:matrix.org> okay that makes sense
2024-01-18 14:24:22 <@tflink:fedora.im> !info if people are interested in adding cuda support in Fedora's pytorch package(s), they are encouraged to reach out on discourse or matrix
2024-01-18 14:24:24 <@trix:fedora.im> yes we want cuda volunteer but then have to be committed, its not something you do once
2024-01-18 14:24:40 <@trix:fedora.im> similar for other features
2024-01-18 14:25:11 <@tflink:fedora.im> !info adding cuda support will not be trivial and it is unlikely to be a one-time effort
2024-01-18 14:25:22 <@tflink:fedora.im> anything else on this or shall we move on?
2024-01-18 14:25:39 <@trix:fedora.im> any cuda takers here ?
2024-01-18 14:26:22 <@trix:fedora.im> if not we move on
2024-01-18 14:26:40 <@martinsk0801:matrix.org> Yeah lets move on.
2024-01-18 14:26:45 <@tflink:fedora.im> I don't think so. neither of the two people who might take that on (in my mind, at least) are here but I know one of them has a conflict
2024-01-18 14:27:28 <@tflink:fedora.im> I don't know what the next topic is, feel free to change it
2024-01-18 14:28:26 <@martinsk0801:matrix.org> !topic legal discussion
2024-01-18 14:28:55 <@tflink:fedora.im> I'm not sure what you're looking to cover but remember that this is a public forum and anything said will be archived publicly
2024-01-18 14:29:17 <@martinsk0801:matrix.org> Noted.
2024-01-18 14:29:26 <@martinsk0801:matrix.org> Just quick questions- very general
2024-01-18 14:29:42 <@trix:fedora.im> i am covering legal internally
2024-01-18 14:29:49 <@trix:fedora.im> we can move on
2024-01-18 14:29:50 <@martinsk0801:matrix.org> Do you think we need to wait on discussion on AI models to proceed with any packaging working you currently have?
2024-01-18 14:30:03 <@martinsk0801:matrix.org> We want to know if there is a blocker here
2024-01-18 14:30:20 <@tflink:fedora.im> we already have some models packaged but talking to legal is on my TODO list
2024-01-18 14:30:38 <@martinsk0801:matrix.org> And second, besides cuda and AI models what are the other topics we would like to cover in the upcoming meetings with legal
2024-01-18 14:30:43 <@martinsk0801:matrix.org> Awesome
2024-01-18 14:30:47 <@tflink:fedora.im> I don't believe that we have any weights packaged, though and AFAIK, that's more of the legal issue
2024-01-18 14:30:57 <@trix:fedora.im> I have made a public statement on 'NO' 
2024-01-18 14:31:25 <@trix:fedora.im> this effects only the couple of torch* packages so far
2024-01-18 14:32:05 <@tflink:fedora.im> and if we ever want to package the huggingface stuff but I don't think anyone has even started on that
2024-01-18 14:32:43 <@tflink:fedora.im> llama-cpp and it's ilk could be affected but that seems unlikely to me but IANAL
2024-01-18 14:33:10 <@trix:fedora.im> llama-cpp models are delivered sperately
2024-01-18 14:33:20 <@trix:fedora.im> they are the llama models
2024-01-18 14:33:34 <@tflink:fedora.im> yeah, I know the differences and I'm not convinced they're that different. again, I stress that IANAL
2024-01-18 14:33:38 <@trix:fedora.im> i had to sign a bunch of things to get the 'free' models
2024-01-18 14:33:42 <@martinsk0801:matrix.org> Bear with me for my limited knowledge here, I would like to ask some clarification here - the 'NO' statement means we are not going to include parameters of the trained models?
2024-01-18 14:34:01 <@tflink:fedora.im> parameters?
2024-01-18 14:34:25 <@trix:fedora.im> it is the pretrained models that concern me.
2024-01-18 14:34:26 <@martinsk0801:matrix.org> the pre-trained models 
2024-01-18 14:34:52 <@tflink:fedora.im> if we're getting into specfiics, can we be clear on  models vs. weights?
2024-01-18 14:34:56 <@trix:fedora.im> trained for things with an untrackable provencence
2024-01-18 14:35:55 <@tflink:fedora.im> AFAIK, there are no legal issues with the models which is more of the interface. the legal issue, AFAIK, is with the weights that come out of the training which could involve copyrighted and other potentially problematic material
2024-01-18 14:35:56 <@trix:fedora.im> trained from things
2024-01-18 14:36:27 <@trix:fedora.im> google 'New York Times sues OpenAI'
2024-01-18 14:36:27 <@martinsk0801:matrix.org> Thanks for the clarification! 
2024-01-18 14:38:19 <@tflink:fedora.im> so, we have already packaged some of the models. however, we have not packaged any of the pretrained weights. the closest that we've gotten to that is torchvision which can automatically download pretrained weights for a few models
2024-01-18 14:39:27 <@martinsk0801:matrix.org> Got it. 
2024-01-18 14:39:50 <@martinsk0801:matrix.org> !info we have already packaged some of the models. however, we have not packaged any of the pretrained weights. the closest that we've gotten to that is torchvision which can automatically download pretrained weights for a few models
2024-01-18 14:40:23 <@tflink:fedora.im> wait, I misspoke
2024-01-18 14:40:30 <@tflink:fedora.im> !undo
2024-01-18 14:41:09 <@tflink:fedora.im> !info we have already packaged some of the models but not many and none of the models that can automatically download pretrained weights
2024-01-18 14:41:13 <@martinsk0801:matrix.org> sorry i captured it too quickly
2024-01-18 14:41:41 <@tflink:fedora.im> no, I manged to forget the whole discussion around pretrained stuff that we had earlier - my bad
2024-01-18 14:42:26 <@tflink:fedora.im> the models which we have or will likely have soon are in llama-cpp
2024-01-18 14:43:42 <@trix:fedora.im> https://bugzilla.redhat.com/show_bug.cgi?id=2255828
2024-01-18 14:43:59 <@trix:fedora.im> once we have that in, i will be doing the rocm port
2024-01-18 14:44:16 <@martinsk0801:matrix.org> Thanks for adding the context here, that is very helpful.
2024-01-18 14:44:37 <@martinsk0801:matrix.org> We are running over time. Do we have any other quick topic before wrapping up please?
2024-01-18 14:44:40 <@trix:fedora.im> its a pretty light weight package 
2024-01-18 14:45:56 <@martinsk0801:matrix.org> If not, i need to drop off. And please continue the discussion if you guys want to stay.
2024-01-18 14:46:10 <@martinsk0801:matrix.org> Thanks again!
2024-01-18 14:46:17 <@tflink:fedora.im> it doesn't sound like we have anything else
2024-01-18 14:46:29 <@tflink:fedora.im> !topic open floor
2024-01-18 14:46:39 <@tflink:fedora.im> if anyone has more topics, feel free to bring them up
2024-01-18 14:47:42 <@trix:fedora.im> thanks for everyone's help getting pytorch in F40!
2024-01-18 14:48:38 <@tflink:fedora.im> thanks for coming, everyone
2024-01-18 14:49:28 <@tflink:fedora.im> !info next meeting is scheduled for 2024-02-01 at 14:00 UTC as a video meeting
2024-01-18 14:49:36 <@tflink:fedora.im> !endmeeting