Hacker News Reader

Comment by davidy123 | Hacker News Reader

Oh funny, I just, inspired by Stable Diffusion, ordered a 3090 ti at what I thought was a good price. Well, I have 30 days to see what happens… There could be an interesting ripple effect if GPUs become more widely available in general. «Simple division yields an estimate of 13.982 million GPUs mining Ethereum on April 16 2021.» (from https://linustechtips.com/topic/1327701-honest-question-how-...)

View on HN

Off-topic but have you got any recommendations on stuff to read to set up a GPU at home to run stable diffusion?

If you have a GPU that has more than 4 GB VRAM, you should be able to easily run it:

https://github.com/basujindal/stable-diffusion

I didn't manage to run it on mine, as it segfaults, but I've heard of people running it fine on an old 1060 (which I have too).

Anyone tried on an intel Mac? I'm ok with Bootcamp.

I don't believe there's anything OS specific for it, just most guides are for Windows for simplicity.

However it does require a nvidia gpu and Macs are generally AMD or integrated gpus.

I believe there is some workaround for AMD as well, but I believe it will be a pain to set up on my machine. I have Radeon Pro 580, on Windows I even use the Adrenaline custom drivers…

Getting VR to work was tiresome, so now I don't want to even start before I see someone else succeeded.

Are you running the scripts in the optimizedSD folder instead of the scripts folder? For me a 512x512 image uses 3.7GB of VRAM. You can reduce the resolution to use less (to a point, the model needs a fixed amount).

I am, yes. I don't think it's the memory because it segfaults immediately, even if I try 64x64... It's very odd, I'm running Ubuntu 22.04, I'm not sure if the CUDA version is somehow wrong.

Oh damn that's weird. Any chance the segfault is on the CPU side? When I try to generate too large of an image I get an allocation error rather than a segfault.

Also crazy idea but could it be related to open source / proprietary driver differences? I haven't done nvidia on Linux in a long time so I don't know if that's in a more reasonable place yet.

Hmm, thanks for the info, it might be a CPU error! There's zero information in the error, it just says "error SIGSEGV" or something similar. I don't think it's an open source driver issue, I'm using nVidia's proprietary one...

It's so odd, it just dies outright. Maybe it is a CPU error, as you say, I think I'll investigate there.

I found this reddit post that has a guide on how to set it up. I recommend reading through the comments as well:

https://reddit.com/r/MachineLearning/comments/wvr23n/d_how_t...

https://www.youtube.com/watch?v=z99WBrs1D3g

Reply on news.ycombinator.com