Hi,
Yes, there will be a public release for r35 as well.
Thanks
Just wanted to say thank you. This got me working again and thanks to Ashwin for the original work, you guys are amazing
1 Like
ggilman, the script worked flawlessly for me as well and resolved the CUDA allocation error/memory fragmentation issues. Thank you for sharing it.
Is there still no official guidance? I’m uncomfortable running community scripts that have such a deep effect.
2 Likes
They released 36.5.. supposed to be fixing this.. I am still trying it now but thought I update this thread
Update: it worked!
2 Likes
Thanks for the info. I follow the instructions. Looks like I can run the bigger models again, yay! By bigger models I mean up to 8.0B.
BTW I followed the instruction on the update link, but could not run: sudo apt install --fix-broken -o Dpkg: :=“–force-overwrite”
Received the following error: E: Option Dpkg:: Configuration item specification must have an =.
I’m not sure it matters cause everything seems to running ok, will monitor and report back with any issues.
try this:
sudo apt install --fix-broken -o Dpkg::Options::=“–force-overwrite”
Some NVIDIA repository entries were duplicated or incomplete, so APT could not detect the newer version. I disabled the old and duplicate repository file and kept one clean repository configuration. I updated the repository entries so that all required NVIDIA repositories (jetson/common, jetson/t234, jetson/ffmpeg) point to r36.5.
After running sudo apt update, I verified that Jetson Linux r36.5 was available as the upgrade candidate.
I then ran sudo apt dist-upgrade and accepted the updated NVIDIA configuration files.
I waited about four months for this update, so it is great that it finally works. Thanks!
Now even larger LLaMA language models, like llama3.2:3B, gemma3:4b run correctly, which is very nice.
The upgrade completed successfully and the system is now running Jetson Linux r36.5.0 (6.2.2+b24).