Sixty-twelfth?

I’d be interested in setting up the highest quality models to run locally, and I don’t have the budget for a GPU with anywhere near enough VRAM, but my main server PC has a 7900x and I could afford to upgrade its RAM - is it possible, and if so how difficult, to get this stuff running on CPU? Inference speed isn’t a sticking point as long as it’s not unusably slow, but I do have access to an OpenAI subscription so there just wouldn’t be much point with lower quality models except as a toy.

Bevy, cause I’m a sucker for Rust

This is a use-after-free, which should be impossible in safe Rust due to the borrow checker. The only way for this to happen would be incorrect unsafe code (still possible, but dramatically reduced code surface to worry about) or a compiler bug. To allocate heap space in safe Rust, you have to use types provided by the language like Box, Rc, Vec, etc. To free that space (in Rust terminology, dropping it by using drop() or letting it go out of scope) you must be the owner of it and there may be current borrows (i.e. no references may exist). Once the variable is droped, the variable is dead so accessing it is a compiler error, and the compiler/std handles freeing the memory.

There’s some extra semantics to some of that but that’s pretty much it. These kind of memory bugs are basically Rust’s raison d’etre - it’s been carefully designed to make most memory bugs impossible without using unsafe. If you’d like more information I’d be happy to provide!

That’s the point. Malicious compliance.

Even as an (older) zoomer in the US, this was never a thing for me. No one cared what phone you used. If you had an Android you wouldn’t be in iMessage group chats but no one judged you for it.

The GPU I used is actually a 1080, with a (rapidly declining in usefulness) Intel 4690k. But I suppose laptop vs desktop can certainly make all the difference. What I really want is GPU virtualization, which I’ve heard AMD supports, but I’m not about to buy a new GPU when what I’ve got works fine.

My experience with single GPU passthrough on Proxmox to a media VM was pretty positive, especially for it being an old Nvidia card. Even as someone doing it for the first time, it just took about 10 minutes to figure out the passthrough itself and another ~15 to figure out some driver issues. And it’s worked perfectly since then. All in all much better than what I’d expected.