

Nothing concerning for me here personally. However, if you have an issue with Firefox’s ‘QoL’ features dependent on Mozilla infrastructure, you could consider Librewolf instead as an alternative.
As for the reverse lookup requests, those are targeting local multicast addresses on your internal network, except the last one which is for the Fastly CDN service
I have to admit I’m not entirely convinced these requests are coming from a compiler…
Is it possible for you to virtualize an non-networked system with your GPU passed through? That seems like the best option IMO. Next best thing would be to set up an airgapped machine just for this, but not everyone has a 2nd machine.
Personally when I was trying out local LLMs I used a virtual machine, mainly due to the known code execution vulnerability related to Tensorflow model data being saved in python’s Pickle format. I believe the recommended save method changed twice since those days though.
With the Firefox stuff I’m assuming you’ve also checked their enterprise config options too (which are admittedly difficult to find and piece together online), so not too sure what else you could do except aside from continuing to block them. Librewolf however provides documentation on how to disable the extra requests sent from their browser in their FAQ, and covers some preferences on the subsequent page linked at the bottom