-
-
Notifications
You must be signed in to change notification settings - Fork 593
using PA on firefox with ollama / gemma3 causes severe memory leak #516
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Comments
Hey, sorry about the issue 😐. Does this happen with other models or just with Gemma 3? |
Also, does this issue occur when using the Chrome or Edge version of Page Assist ? |
Hi! This only seems to happen with gemma3 but i've also noticed similar behavior when enabling web search on other models to a lesser extent. I also tried PA in supermium 126 (chromium fork) on a windows VM and it doesn't seem to happen over there, its quite bad on firefox 136.01 on windows though. have not tested linux yet. |
@lands39 Latest Ollama? |
yes |
did some testing with gemma and page assist on linux, everything seems fine so perhaps this is an issue specifically with windows and firefox |
i've run into another issue on firefox: when deleting old chats, it would always give me a confirmation prompt, but also had a box that siad "don't prompt me about this again". i checked the box thinking it would not ask to confirm anymore, and when i tried to delete chats, chats would not delete. i uninstalled the plugin, reinstalled it and now new chats don't appear under history unless i restart firefox. deleting works again though. |
okay, restarted ollama, cleared cache and that fixed the above, memory leak does still happen though |
Same issue with firefox. it uses all my base ram, But doesn't really touch my GPU's VRAM at all, even with the GPU layers set to 1 |
is there a way i could provide logs or more in-depth information on this? would love to see this get fixed. currently as a workaround i'm running ollama on my pc and the browser plugin on another system. |
Hey @lands39 , both Chrome and Firefox implementations are the same, but I don't know why it's causing a memory leak when using the Gemma model on Firefox. If you're okay with it, could you try running it on Firefox Nightly to see if there's an issue? I found a few reports on Bugzilla about memory leaks with extensions. For example: https://bugzilla.mozilla.org/show_bug.cgi?id=1932080 |
sure, i can try this in the morning. i have a windows VM with nightly already installed so i'll try on there too and see what it does |
doesn't appear to happen when connecting to a remote ollama instance on nightly from my windows 11 VM, and testing locally with nightly doesn't seem to do it. i also tried copying over my firefox profile and i can't trigger the issue either, but when installing the page assist plugin firefox / explorer? did kinda freak out for a second. |
i updated my normal firefox install to 136.0.4 and it seems to be behaving as expected, when i ask gemma3 to analyze an image firefox does spike by 1 - 2 gigs after the prompt finishes generating, but this quickly drops back to normal. i'll do further testing and see what happens |
Thanks for the response. From what I understand in the Gemma 3 technical report, Gemma 3 uses a vision encoder called SigLIP when processing input. I think Ollama does not load it into memory if we are not using images, which is why there is a spike. Maybe I’m wrong |
The latest version is currently under review at the add-on store. Once it's approved, the issue will hopefully be resolved. |
did the new version come out? it seems to still have issues, interestingly when i cleared my history firefox had a bad memory leak. |
when using page assist to talk to gemma3 running on ollama, a severe memory leak occurs and i've observed firefox using upwards of 17 gigs of RAM, rendering my system completely unusable as ollama is filling up the rest of my RAM. this does not happen when interacting with other models. system has 32 gigs of RAM.
The text was updated successfully, but these errors were encountered: