Ok, I'm back.
After LOTS of testing, EVERY PyTorch version with EVERY Pi OS version...
The problem simply lies in all the software problems we face when changing OS versions, language versions, updating software, upgrading, etc...
Stuff gets broken, or was depreciated, or a security issue was found which pretty much deems everything useless.
In this particular case, Newer version of Python on Trixie will not install PyTorch 2.5.1,
which has a good memory footprint to run on an 8GB Pi5, compared to other versions on the Pi5 (before Trixie) using the same codebase.
But PyTorch 2.5.1 has a security issue, Critical vulnerability in PyTorch 2.5.1 (CVE-2025-32434), so we can not use that version.
If we go with 2.7.1 or higher of PyTorch like PIP wants to do, PyTorch's memory footprint grows. Which on an 8GB Pi5 will not work out, without some/alot of re-arranging how the normal workflow goes using the Diffusers library which is what is using the PyTorch library to begin with.
So we try PyTorch 2.6.0, no security issues to date, so far fairly stable, memory footprint grown a tad I think from 2.5.1. It did not grow to anywhere near 2.7.1 and higher was. Could not even get one image out before it would run out of RAM.
I'm sure later versions adds more features to the library, but that means bigger, more RAM used, ie... bloat.
Because they added nothing useful for me.
But 2.6.0 is still questionable. Which is why I decided to try DietPi.
Small, simple install setup, just like Pi OS, but did not have Python and Git plus a few other things, but a sudo apt install... fixes that.
So far worked ok, so far.
One thing that upsets me is that I have been using this, and beating on it, for over 7 months, no problems on both an 8GB Pi5 and a 16GB Pi5.
Had only one problem in the beginning, and bumped up the swap space a bit, problem solved.
And now the brick wall.
So, as I mentioned, I tried them ALL, with ALL Raspberry Pi OS's, plus DietPi.
Turned off most everything not needed to reclaim the precious RAM which IS the problem.
I even stripped out everything SDXL since not needed on an 8GB Pi5. Still RAM usage was terrible.
It's mainly the way diffusers/pytorch libraries handle memory management within the libraries themselves.
Also possible kernel or zram thing on Trixie. Lots of new, new, new on the release notes, and less, fix, fix.
But I'll be trying it on the future Trixie's too.
So far it's narrowed down to this:
using PyTorch 2.6.0, with default Python version, or latest Python version in case of DietPi.
Trixie 64bit Lite: No - uses 500-600MB of RAM when idle (Too much of a difference to bother with, plus lots of zram errors, errors, i mean hassle, switch to old school swap, runs out of memory, but runs faster, before it breaks)
Bookworm 64bit Lite: Yes - uses 177-195MB of RAM when idle (Was developed and tested on 2024-11-19 Bookworm Lite, Note: not all
Bookworms performed as well)
DietPi: Yes - uses 98-110MB of RAM when idle (Worked the same as Bookworm, yet is Trixie based according to DietPi)
Once I am done and finalize my decision in a day or two, and update everyone.
It may take a week to get everything finished and uploaded. I will be rebuilding the requirements file and re-tweaking some code (I still got a few more ideas, and takes hours to test one thing because you have to try all the features in the app that utilize the 'one change' in code).
In the meantime I made the repo at Github 'private' to keep anyone from downloading something that does not work correctly.
Thanks for everyone who has feed back to me and/or tried it. As well as your patience.
After LOTS of testing, EVERY PyTorch version with EVERY Pi OS version...
The problem simply lies in all the software problems we face when changing OS versions, language versions, updating software, upgrading, etc...
Stuff gets broken, or was depreciated, or a security issue was found which pretty much deems everything useless.
In this particular case, Newer version of Python on Trixie will not install PyTorch 2.5.1,
which has a good memory footprint to run on an 8GB Pi5, compared to other versions on the Pi5 (before Trixie) using the same codebase.
But PyTorch 2.5.1 has a security issue, Critical vulnerability in PyTorch 2.5.1 (CVE-2025-32434), so we can not use that version.
If we go with 2.7.1 or higher of PyTorch like PIP wants to do, PyTorch's memory footprint grows. Which on an 8GB Pi5 will not work out, without some/alot of re-arranging how the normal workflow goes using the Diffusers library which is what is using the PyTorch library to begin with.
So we try PyTorch 2.6.0, no security issues to date, so far fairly stable, memory footprint grown a tad I think from 2.5.1. It did not grow to anywhere near 2.7.1 and higher was. Could not even get one image out before it would run out of RAM.
I'm sure later versions adds more features to the library, but that means bigger, more RAM used, ie... bloat.
Because they added nothing useful for me.
But 2.6.0 is still questionable. Which is why I decided to try DietPi.
Small, simple install setup, just like Pi OS, but did not have Python and Git plus a few other things, but a sudo apt install... fixes that.
So far worked ok, so far.
One thing that upsets me is that I have been using this, and beating on it, for over 7 months, no problems on both an 8GB Pi5 and a 16GB Pi5.
Had only one problem in the beginning, and bumped up the swap space a bit, problem solved.
And now the brick wall.
So, as I mentioned, I tried them ALL, with ALL Raspberry Pi OS's, plus DietPi.
Turned off most everything not needed to reclaim the precious RAM which IS the problem.
I even stripped out everything SDXL since not needed on an 8GB Pi5. Still RAM usage was terrible.
It's mainly the way diffusers/pytorch libraries handle memory management within the libraries themselves.
Also possible kernel or zram thing on Trixie. Lots of new, new, new on the release notes, and less, fix, fix.
But I'll be trying it on the future Trixie's too.
So far it's narrowed down to this:
using PyTorch 2.6.0, with default Python version, or latest Python version in case of DietPi.
Trixie 64bit Lite: No - uses 500-600MB of RAM when idle (Too much of a difference to bother with, plus lots of zram errors, errors, i mean hassle, switch to old school swap, runs out of memory, but runs faster, before it breaks)
Bookworm 64bit Lite: Yes - uses 177-195MB of RAM when idle (Was developed and tested on 2024-11-19 Bookworm Lite, Note: not all
Bookworms performed as well)
DietPi: Yes - uses 98-110MB of RAM when idle (Worked the same as Bookworm, yet is Trixie based according to DietPi)
Once I am done and finalize my decision in a day or two, and update everyone.
It may take a week to get everything finished and uploaded. I will be rebuilding the requirements file and re-tweaking some code (I still got a few more ideas, and takes hours to test one thing because you have to try all the features in the app that utilize the 'one change' in code).
In the meantime I made the repo at Github 'private' to keep anyone from downloading something that does not work correctly.
Thanks for everyone who has feed back to me and/or tried it. As well as your patience.
Statistics: Posted by rock-stevens — Wed Oct 29, 2025 2:22 pm