IRC channel logs

<rekado>today it’s my turn to ask a FAQ:

<rekado>does anyone have a more recent tensorflow package?

<rekado>I’ll try packaging tensorflow-lite now

<civodul>my turn to say "good luck" ;-)

<civodul>(speaking of which, i'm still not done with pytorch, but i've never been this close)

<rekado>I hope you can push pytorch over the finish line

<rekado>one of these days I’ll also have to see if I can get keras to work. We have a python-keras package, and I’ll need *something* for CRAN packages that want to talk to keras.

<rekado>interesting: https://github.com/google/ruy/blob/master/cmake/bazel_to_cmake.py

<civodul>oh! looks like something that could save us

<civodul>looks like folks at Google feel the need to have an escape hatch

<rekado>the bazel stuff for tensorflow is rather complex. I fear this converter may not be smart enough to deal with it.

<rekado>I think their need for cmake is due to the desire to support building on Windows.

<PurpleSym>Do we actually have a proper GPU runtime like ROCm for Guix somewhere? (I presume everyone is running NVIDIA/CUDA anyway…?)

<civodul>PurpleSym: we don't, but we were discussing it the other day, and it seems like it'd be welcome

<civodul>rekado: i suppose Bazel supports Windows, but my guess is that Bazel still lacks momentum outside Google and so they may prefer CMake for their "open source products"

<rekado># tensorflow-lite uses find_package for this package, so override the system

<rekado># installation and build from source instead.

<rekado>why??

<rekado>ugh…

<civodul>uh

<civodul>in the same vein, take a look at https://github.com/pytorch/pytorch/tree/master/third_party

<rekado>oh my…

<civodul>there are direct and indirect deps, and also forks of the same thing (like XNNPACK/QNNPACK)

IRC channel logs

2021-08-02.log