I just got Oobabooga running for the first time with Llama-2, and have Automatic1111, and ComfyUI running for images. I am curious about ML too but I don’t know where this start with that one yet.
For the uninitiated, all of these tools are running offline open source (or mostly) models.
WizardLM 30B at 4 bits with the GGML version on Oobabooga runs almost as fast as Llama2 7B on just the GPU. I set it up with 10 threads on the CPU and ~20 layers on the GPU. That leaves plenty of room for a 4096 context with a batch size of 2048. I can even run a 2GB Stable Diffusion model at the same time with my 3080’s 16GBV.
Have you tried any of the larger models? I just ordered 64GB of ram. I also got kobold mostly working. I hope to use it to try Falcon 40. I really want to try a 70B model at 2-4 bit and see how its accuracy is.
I just tried it a few hours ago. Indeed, it is quite good. I knew it when a NSFW prompt test on an uncensored model generated a stable diffusion picture of a robot skeleton and a snarky reply. Like, yay we finally have a bight spot with this one.
You are not logged in. However you can subscribe from another Fediverse account, for example Lemmy or Mastodon. To do this, paste the following into the search field of your instance: !technology@beehaw.org
A nice place to discuss rumors, happenings, innovations, and challenges in the technology sphere. We also welcome discussions on the intersections of technology and society. If it’s technological news or discussion of technology, it probably belongs here.
Remember the overriding ethos on Beehaw: Be(e) Nice. Each user you encounter here is a person, and should be treated with kindness (even if they’re wrong, or use a Linux distro you don’t like). Personal attacks will not be tolerated.
Check out Wizard 30B Uncensored. IMO it’s about as good as NerfedGPT 4… except free and private.
What hardware does it take to run a 30B?
I’m running it in GPT4All (CPU-based) with 64GB of RAM, and it runs pretty well. I’m not sure what you’d need if you were running it on GPU instead.
WizardLM 30B at 4 bits with the GGML version on Oobabooga runs almost as fast as Llama2 7B on just the GPU. I set it up with 10 threads on the CPU and ~20 layers on the GPU. That leaves plenty of room for a 4096 context with a batch size of 2048. I can even run a 2GB Stable Diffusion model at the same time with my 3080’s 16GBV.
Have you tried any of the larger models? I just ordered 64GB of ram. I also got kobold mostly working. I hope to use it to try Falcon 40. I really want to try a 70B model at 2-4 bit and see how its accuracy is.
I just tried it a few hours ago. Indeed, it is quite good. I knew it when a NSFW prompt test on an uncensored model generated a stable diffusion picture of a robot skeleton and a snarky reply. Like, yay we finally have a bight spot with this one.