search

johnnymcmike / Gravital

A Discord AI Chatbot that uses GPT-2 and aitextgen for fast, believable responses that you can train on your own discord server's message history
MIT License
34 stars 5 forks source link

Is it possible to train this without a nvidia gpu / with an integrated gpu? #5

Closed soupslurpr closed 2 years ago

soupslurpr commented 2 years ago

Hi is that possible because when I tried it said CUDA is not installed

johnnymcmike commented 2 years ago

It is technically possible but not supported by this project. Training it on your own hardware when you aren't running an NVIDIA gpu on linux isn't something I've figured out how to do with this library yet, and even then it'd be really slow. Don't worry though, you can still get it trained really well (better than even leaving your nvidialess computer on training for a week probably) using Google Colab. I think I outlined this in the tutorial but if you have questions lmk here

soupslurpr commented 2 years ago

It is technically possible but not supported by this project. Training it on your own hardware when you aren't running an NVIDIA gpu on linux isn't something I've figured out how to do with this library yet, and even then it'd be really slow. Don't worry though, you can still get it trained really well (better than even leaving your nvidialess computer on training for a week probably) using Google Colab. I think I outlined this in the tutorial but if you have questions lmk here

Thanks, but I am afraid of the privacy of google colab since the server I want to do this in is private

johnnymcmike commented 2 years ago

That's understandable. If it's any comfort, I don't think colab is a data harvesting thing. It has more to do with getting researchers and hobbyists into their platform and using their open source frameworks like tensorflow so that they'll eventually buy google cloud compute, and that business model has worked out really well for them so they don't really have a lot of reason to snoot through random hobby devs' projects looking for souls to steal.

The worst that would happen is that your downloaded training data would be in google drive briefly, but that's about it. I actually think there's a way to "cache" your data with this lib, which sort of compresses/encrypts it and makes it faster for training, and then only upload that cached unreadable version. Could help you with that if it would make you feel better

johnnymcmike commented 2 years ago

And worst-case, I could also train it for you on my 3060 but i'll be out this weekend

soupslurpr commented 2 years ago

That's understandable. If it's any comfort, I don't think colab is a data harvesting thing. It has more to do with getting researchers and hobbyists into their platform and using their open source frameworks like tensorflow so that they'll eventually buy google cloud compute, and that business model has worked out really well for them so they don't really have a lot of reason to snoot through random hobby devs' projects looking for souls to steal.

The worst that would happen is that your downloaded training data would be in google drive briefly, but that's about it. I actually think there's a way to "cache" your data with this lib, which sort of compresses/encrypts it and makes it faster for training, and then only upload that cached unreadable version. Could help you with that if it would make you feel better

Hmm yes, your probably right, I think I'm going to try on google colab . thanks