Show HN: GPT-2 implemented using graphics shaders
github.comBack in the old days, people used to do general-purpose GPU programming by using shaders like GLSL. This is what inspired NVIDIA (and other companies) to eventually create CUDA (and friends). This is an implementation of GPT-2 using WebGL and shaders. Enjoy!
As the guy who did GPT2 in Excel, very cool and kudos!!
Curious why you chose WebGL over WebGPU? Just to show it can be done?
(Also see my other comment about fetching weights from huggingface)
This was a final project for a graphics class where we used WebGL a lot. Also I was just more familiar with OpenGL and haven't looked that much into webGPU
Someone needs to implement Excel using graphics shaders now.
Probably because WebGPU support is still rather iffy.
> Curious why you chose WebGL over WebGPU? Just to show it can be done?
For a WebGPU implementation, one can use transformers.js directly (or many other libraries actually), maybe WebGL is more original.
[1]: https://huggingface.co/docs/transformers.js/index
Transformers.js wraps the ONNX runtime which is rather versatile (WASM, WebGL, WebGPU, and WebNN). It's not the backend that makes it novel.
ianand, I immediately thought of you when I saw this post. Miss you friend.
I'm reminded of this VRChat world, which runs Qwen2-0.5B in a shader: https://vrchat.com/home/launch?worldId=wrld_e919c1f4-8399-43...
Super cool project! How much of the inference time is actually GPU-bound vs CPU-bound in the browser? Curious if WebGPU would offer a big perf boost over WebGL2 here.
Would absolutely read a 'making of' post on this project!
Would be really cool to have it hosted on a website if it can be a static page so we could try it out
Working on getting it up on Github pages, will let you know. Not sure if there will be an issue with hosting the weights, I'll look into it
I haven't looked at your implementation, but if you can read ONNX weights... https://huggingface.co/Xenova/gpt2
You may be able to fetch the weights directly from Hugging Face. I'd try that first.
Checkout https://github.com/jseeio/gpt2-tfjs fetches the weights for GPT2 from huggingface on the fly.
Nice, that takes some serious grinding to pull off in shaders. Makes me kinda wonder - does pushing old tools like WebGL actually force deeper understanding of how these models work compared to just grabbing a new library?
Incredible. Was going to suggest tossing up a github page with it, but looks like one is currently in-progress? (https://nathan-barry.github.io/gpt2-webgl/) but doesn't seem to load for me currently.
Looks like it's trying to load the typescript directly / hosting the source files: `<script type="module" src="src/main.ts"></script>` - instead of hosting the dist folder.
Yeah, right now the weights aren't even pushed to the repo so there's no way for the github page to load them in. I've looked a tiny bit into it and might end up making a fix for it, having it load the weights by fetching them from somewhere else on load, but probably not today.
There is a link to the zipped up folder with the weights here: https://github.com/nathan-barry/gpt2-webgl/releases/download...
If anyone wants to make a PR which fetches the file from there, unzips it, then loads in the weights from there that would be greatly appreciated.
Now do it with the fixed function pipeline so I can run it on my 3dfx Voodoo card.
This is pretty rad. I've built and run the project on my MacBook and can't seem to get the visualizations to show as they appear in the GitHub demonstration. I'm running Firefox but I think that should support WebGL features you're using?
Happy to file an issue on the repository as well if you'd prefer that.
Yeah it'd be great if you could file an issue and show what you see! I was using Safari but I'm gonna check other browsers to test if it works on them
Great! I've filed an issue.