I recommend to drop a number of AI/ML packages from openSUSE Leap.
Some of these packages have proven to be high maintenance:
Tensorflow2 is tedious to keep building even on Factory due to its tight dependency requirements - especially to bazel. This also limits the ability to upgrade the package on Leap. Maintaining old versions would require to backport security fixes with - little benefit to the user: AI/ML users expect the availability of the latest version. For Leap, as for any other OS version, this is provided best thru directly build application containers.
Tensorflow2
bazel
Tensorflow(1) has meanwhile been superseded by Tensorflow2, users will have migrated to the successor version by now.
Tensorflow(1)
Other AI/ML packages: Most of the following packages are maintained in the science:machinelearning devel project, however, not even their maintainers have shown interest to submit updates to Leap or have requested such updates to be done. It is therefore gathered that uptake of these packages in Leap is low - or non-existing: 1. ComputeLibrary 2. azure-storage-cpp 3. onednn 4. python-Keras-Applications 5. python-Keras-Preprocessing 6. python-onnx 7. python-onnxconverter-common 8. python-opt-einsum 9. python-torch The current maintainers are no longer able to maintain these packages for openSUSE Leap. Despite being a community project, so far, no one has stepped up to 'share the pain'. The uptake of these packages in Leap seems to be low. These packages should therefore be dropped from Leap going forward. Most of them will live on happily in openSUSE:Tumbleweed and/or their respective devel projects in OBS.
science:machinelearning
ComputeLibrary
azure-storage-cpp
onednn
python-Keras-Applications
python-Keras-Preprocessing
python-onnx
python-onnxconverter-common
python-opt-einsum
python-torch
Just for reference these are all community maintained packages (Backports). My recommendation is to ask on factory for volunteers.
I did also ask PM for his opinion as we'll essentially stop providing these to SLE customers via package hub.
Blocking from further progress until we hear back from PM and then factory@
Metadata Update from @lkocman: - Issue set to the milestone: 15.4
Metadata Update from @lkocman: - Issue assigned to lkocman
Metadata Update from @lkocman: - Issue tagged with: SLE-Accept-Pending
Metadata Update from @lkocman: - Issue tagged with: Leap-Maintainer-Missing
Just for reference these are all community maintained packages (Backports). My recommendation is to ask on factory for volunteers. I did also ask PM for his opinion as we'll essentially stop providing these to SLE customers via package hub.
The purpose of this exercise is to see if there are volunteers who'd be willing to help out on these packages. The PM who has requested this has been notified already.
Lubos to send a looking for volunteers email today. If we do not find volunteers until Beta then packages will be dropped.
Removing sle-accept-pending tag as I did forward the call for maintainer to devel@suse.de
Unless somebody shows interest until 18th of February (code-submission deadline for Beta) packages will be removed from 15.4.
I'll send one last reminder on February 1st (~ 2 weeks prior submission deadline)
Metadata Update from @lkocman: - Issue untagged with: SLE-Accept-Pending
Hi,
somehow I ended up as (co-)maintainer of some of the Python based packages in science:machinelearning, because I put in some efforts to keep them building in Factory. Given that these packages are meant to be run in upstream provided containers, it is even arguable whether they should be kept in Tumbleweed/Factory. For example, right now, tensorflow2 builds in Factory, but is untested and likely does not fully work with the Tumbleweed's NumPy version [1]. So it is questionable how many Tumbleweed users actually use the distro package.
[1] https://github.com/tensorflow/tensorflow/pull/48935.
Hi, somehow I ended up as (co-)maintainer of some of the Python based packages in science:machinelearning, because I put in some efforts to keep them building in Factory. Given that these packages are meant to be run in upstream provided containers, it is even arguable whether they should be kept in Tumbleweed/Factory. For example, right now, tensorflow2 builds in Factory, but is untested and likely does not fully work with the Tumbleweed's NumPy version [1]. So it is questionable how many Tumbleweed users actually use the distro package. [1] https://github.com/tensorflow/tensorflow/pull/48935.
Indeed, this was also my take on this. I haven't pushed for this more strongly since there was always someone willing to step in and help. I figured that one of those who did was actually using the stuff. I may have been wrong, though. In any case, I'd like to add more people here to solicit opinions, however, I don't seem to be able to do that. I'd like to add more people to this
I need ComputeLibrary and tensorflow2-lite-devel for armnn.
tensorflow2-lite-devel
armnn
Some updates were just ignored for Backports SP3: https://build.opensuse.org/project/show/openSUSE:Backports:SLE-15-SP3:Staging:adi:1
I am going to disable the build of non-lite versions of Tensorflow2 in Backports:SLE-15-SP4. Lite version is required by armnn.
One more recruitment email sent to factory :-)
Login to comment on this ticket.