[ot][crazy][hobby] karl’s recent projects

[Karl Semich]

i don't do much but coped with small software projects in python this week.

i found a very powerful language model architecture called sgconv that
is one of the leading world models atm. the current state of the model
it required 20GB of ram per layer, and would not run on my old system,
so i poked at optimizing the ram usage. i removed temporaries from the
kernel allocation and broke the core convolution phase into tiles so
the ram usage could be controlled. the work is at
https://github.com/ctlllll/SGConv/pull/5 . i am daydreaming of next
adding implicit dealiasing to the convolution, which further increases
speed and reduces memory requirements, and i think with some tuning it
could be done in such a way that the kernel construction can be
removed entirely. i would likely add pushes to that pull request if i
did that.

i found also a parametric generator for lasercut boxes and other
parts, including gear stages. this was really exciting to me because i
have spent a lot of personal energy trying to begin something similar,
and fought hard with my dissociated parts, and mostly lost the work. i
thought i would use the hears for the rotary encoder project, and
poking at adding bolts to them to clip them in some way to shafts of
the cheap motors i get which don’t have a D cut. i found the the code
has existing code for bolts that is disabled and the issues for
enabling them was disrupted by a question and answer session with a
visitor. i added imperial measurements to the bold code to be easy to
use in my country, and a command line option to specify the bolt
standard, and enabled the code. this work is at
https://github.com/florianfesti/boxes/pull/468 . it likely has bugs
since i have not tested it, which requires a lasercutter, but there is
a reasonable chance that it works well enough to use. in my xloem
github user i forked the repository and there is a branch with the
gear bolt work. it’s not the best way to attach a gear to a shaft, and
i’m curious what the best way would be, but i wanted to try it.

both of these projects were quite enjoyable to me, and i’m presently
holding plans to continue them, but they are recreational and not what
i intended to dedicate much time to. still, i do really appreciate the
experiences of, despite my issues, meeting goals I find hard and used
to be skilled it. i look for those successes and want to nurture them.