> No single piece of software has wasted more of my time than Nvidia's drivers
I see you haven't used printers much :)
I see you haven't used printers much :)
I swear to god if I ever become wealthy the printer industry is what I intended to completely destroy. Not in it for profit. Not positive sum whatever startup thinking. It will be Zero Sum.
Edit: Lasers are fine. That will be left alone.
Not he hero we deserve, but the hero we need!
It's just everywhere else that they are a problem.
"Here, run this piece of code. No don't bother trying to read, build or understand it yourself, we know you won't be able to do it, so we put it in a little mystery box over here that you can run on top of your existing operating system. Now please don't ask any questions and go away."
I can read and understand tf just fine and I’ve had to work around bugs (like the broken image resampling) by doing just that.
Treating tf as a blank box has nothing to do with it. The problem is that nothing none of the ABIs involved here (tf, cudnn, distro libc, nvidia driver, etc) are stable. The best you can hope for is to use the same driver for several containers with matched (tf, libc, cudnn) and that’s worked out well for me in the last couple of years.
But even if you compile it yourself, what exactly are you gaining? Extra work for something you won't look into anyway? You'd think open source stuff would be more peer reviewed, but as a maintainer of a few open company repos it's surprising how much nobody actually cares and will just run whatever.
Most of the time it works fine because the average person isn't a malicious actor, but every so often something gets compromised and it's always discovered by pure chance.
So long as the driver doesn’t change, you can even run multiple versions simultaneously and nothing will break.