> Are the people who depend on it seriously incapable of taking up the load?
Probably. First of all most are domain experts not at all systems programmers. That’s the main value of such libraries.
And even for those who work close to the iron: plenty depend on, say, the Linux kernel without being able to write patches or even diagnose problems. And few, if any kernel developers have submitted patches to, say, GCC.
There seems to be serious talk about reimplementing the Torchscript core in Julia (i.e., reimplement the C++ part in Julia), something that makes sense to me given the excellent support there already is in Julia for autodiff. Cf. e.g., https://dev-discuss.pytorch.org/t/where-we-are-headed-and-wh...
If these suggestions are credible and don't come from the Facebook team, that might sidestep your concerns.
While I would love for that to happen. Quite recently (e.g. [1], also discussed on HN, I believe) the Julia community seemed content with more niche applications, quite deliberately choosing not to challenge the biggies (TensorFlow, Torch, etc.) on deep learning, because the amount of mainainers needed to be competitive was just not there.
Is there reason to expect this to change? Maintaining something like Torch and keeping it competitive in terms of speed is a HUGE amount of work for systems development, writing insane numbers of GPU kernels, etc. After having read it, I wouldn't quite yet call the tone of the discussion you linked "serious talk"...
Thanks