IRC channel logs


back to list of logs

<rekado>now moving on to tensorflow 2.14
<civodul>wo0t! brave & fearless!
<civodul>i’m looking at
<civodul>i can’t see where it’s “cheating”
<rekado>it’s here:
<civodul>oh, got it
<rekado>lots of bundled jars
<civodul>and where’s source code injected?
<rekado>and also bundled third-party source code
<rekado>bazel is special because it includes a bunch of zip files for the source code
<rekado>the gross source code fetching via bazel happens in downstream packages such as jaxlib
<rekado>e.g. here:
<rekado>every bazel-using package will need something like that
<rekado>I’ll eventually turn this into a bazel-build-system once I have finished building tensorflow
<rekado>(then I’ll know what features should be generic and which are just for jaxlib)
<rekado>later I’ll also have to build all the CUDA-enabled variants of these packages for guix-science-nonfree. Not looking forward to that.
<civodul>i see
<civodul>python-jaxlib is a bit intimidating
<rekado>I find it annoying that I can’t get tensorflow to use the already built XLA from jax.
<rekado>it insist on building everything from scratch.
<rekado>same for that custom llvm it bundles
<rekado>[10,778 / 11,791]
<civodul>we’re going to have to double the capacity of :-)
<rekado>this is like these windows progress bars
<rekado>[11,279 / 12,136]
<rekado>and that’s just libtensorflow without any of the Python bits
<rekado>(i.e. it’s essentially useless)
<rekado>tensorflow is more difficult to package.
<rekado>it brings its own binary of Python and insists on using pip install as part of the configuration process.
<rekado>ACTION shudders
<rekado>it looks like nixpkgs doesn’t actually use their own python packages in the tensorflow package.