keras is broken –> package bazel?

  • Open
  • quality assurance status badge
Details
3 participants
  • Ricardo Wurmus
  • Sharlatan Hellseher
  • Simon Tournier
Owner
unassigned
Submitted by
Ricardo Wurmus
Severity
normal
Merged with
R
R
Ricardo Wurmus wrote on 6 Jan 2023 22:16
(address . bug-guix@gnu.org)
87h6x32wdz.fsf@elephly.net
We have an old version of keras. This is now broken. It is not
compatible with python-h5py > 2.10. Our version of Tensorflow also
won’t work with h5py 3.

We can’t upgrade keras to anything above 2.2.5 because any later version
is built with Bazel. We cannot build Bazel, because it needs a big pile
of jars that we haven’t packaged yet.

Tensorflow > 1.9 also requires Bazel.

We should work towards packaging Bazel. Debian has a Bazel bootstrap
package, so I’m hopeful that it’s possible to package with some effort.

--
Ricardo
S
S
Simon Tournier wrote on 9 Jan 2023 13:00
control message for bug #60608
(address . control@debbugs.gnu.org)
874jszzzmw.fsf@gmail.com
merge 60608 60545
quit
R
R
Ricardo Wurmus wrote on 21 Oct 2023 17:18
keras is broken –> package bazel?
(address . 60608@debbugs.gnu.org)
87il70j95c.fsf@elephly.net
Bazel is now available in the guix-science channel. We may need to move
keras there too and build it with bazel.

--
Ricardo
S
S
Sharlatan Hellseher wrote on 19 Dec 2024 00:46
keras is broken –> package bazel?
(address . 60608@debbugs.gnu.org)(name . Ricardo Wurmus)(address . rekado@elephly.net)
8734ikpnaq.fsf@gmail.com
Hi,

I'm in a limbo to upgrade https://github.com/spf13/afero,which requires
go-google-golang-org-api, which relays on auto-generated code in
have in Guix (yet?)

Any plan to have Bazel in main Guix git repository or it's not
compatibly license wise?

--
Oleg
-----BEGIN PGP SIGNATURE-----
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=l3ll
-----END PGP SIGNATURE-----

R
R
Ricardo Wurmus wrote on 19 Dec 2024 12:19
(name . Sharlatan Hellseher)(address . sharlatanus@gmail.com)(address . 60608@debbugs.gnu.org)
87o717gbt8.fsf@elephly.net
Sharlatan Hellseher <sharlatanus@gmail.com> writes:

Toggle quote (3 lines)
> Any plan to have Bazel in main Guix git repository or it's not
> compatibly license wise?

Bazel is in guix-science, with a bazel-build-system as well. It's
compatible when looking at just the licenses, but there are a number of
annoying wrinkles:

- Bazel itself bundles lots of binaries (Java libraries); I have made no
serious attempt at unbundling them. Packaging Java libraries gives me
ulcers.

- the bazel-build-system works in ways that are really quite mismatched
with how Guix works. It performs a two-stage build: the first step is
to run Bazel to download "sources" in a fixed output derivation; the
second step (at a later time) is to unpack these hash-identified
sources and run Bazel again without the download.

- Bazel makes it *very* hard or even impossible to replace parts of the
build dependencies with packages from Guix. I recently added
python-ray to guix-science, which insists on using its own GNU Make
(among many others). This means that the source blob is usually
incredibly large.

- Bazel does not care about whether something is source or binary. The
bazel-build-system assumes that the first step is to download sources
only. Bazel does not see it this way and is perfectly happy to fetch
or build binaries at this stage. It has happened a number of times
that the hash of the fixed output derivation changed after an update
of seemingly unrelated libraries, because the fixed output wasn't
fixed after all.

I'd rather move Keras to guix-science than to contaminate Guix proper
with the abomination that is the bazel-build-system (and its relative
Bazel, which I will attempt not to call an abomination in public).

--
Ricardo
?
Your comment

Commenting via the web interface is currently disabled.

To comment on this conversation send an email to 60608@debbugs.gnu.org

To respond to this issue using the mumi CLI, first switch to it
mumi current 60608
Then, you may apply the latest patchset in this issue (with sign off)
mumi am -- -s
Or, compose a reply to this issue
mumi compose
Or, send patches to this issue
mumi send-email *.patch