Pre-trained models run in WebAssembly via the project site.
Pufferlib is insane. You can train neural networks to play games out of the box if you have a CUDA GPU. Like breakout, Atari games, continuous action space problems. You can go to the website right now and they have neural nets running in wasm
Many users praise Pufferlib for simplifying RL training on Atari games via CUDA by treating env throughput as first-class and easing the learning curve, while some dismiss it over GPU requirements and skepticism about hardware scaling.
No Digg Deeper questions have been answered for this story yet.
This week I'm going to do some research using pufferlib for fun. My intention isn't to push the needle on algorithms, I still have much to learn. I'm mostly trying to share The Good News. Pufferlib has been indispensible for me, for learning more about RL, and properties of NNs
Pufferlib is insane. You can train neural networks to play games out of the box if you have a CUDA GPU. Like breakout, Atari games, continuous action space problems. You can go to the website right now and they have neural nets running in wasm
@yacineMTB + 5.0 will be a short update - coming soon!
Pufferlib is insane. You can train neural networks to play games out of the box if you have a CUDA GPU. Like breakout, Atari games, continuous action space problems. You can go to the website right now and they have neural nets running in wasm
@yacineMTB wasm + neural nets feels illegal on paper but its such a natural evolution
did u test the built in atari envs yet?
@yacineMTB what kind of GPU do you need to do meaningful stuff in a reasonable amount of time?
what does your setup look like for what you're doing with the drone stuff?
also, what's the break even on your own rig vs just renting H100 hours in the cloud?
@zackslab @yacineMTB If you can get an H100… about $4/hr. So it’ll take you 500h or about 20 days to spend $2000, or enough for a RTX 3090.
@zackslab @yacineMTB An H100 is like $40k before you get all the infra to run it. But just assuming $40k, it’ll take 416 days to break even on the card itself.
It just depends on your situation, if using the card is fundamental to your revenue stream or you just need it to crank out a model.
@ntbrown01 @yacineMTB yeah but i assume a single 3090 is hardly enough to do the same work you’d have done on those first 500 hrs of an H100?
how many 3090s would it take to have the equivalent training capabilities as one H100?
@yacineMTB Thanks for puffer pilling me Been messing with neural mmo and customized ocean envs
@yacineMTB I also wanna explore the wasm llm stuff -- just implemented the onnx / llama.cpp pipelines in jai for text-embedding -- wasm next?
@yacineMTB One day if I get bored I might try and train one to play Minecraft and see if it can do better than stuff like Baritone
@yacineMTB I stopped reading after CUDA GPU
@ntbrown01 @yacineMTB yeah I’m realizing there is a lot more nuance to this depending on what you’re doing.
@yacineMTB good tools like this make the rl learning curve actually manageable
@yacineMTB it really is the coolest thing to come out for training in recent memory
@yacineMTB but its paid no?
@yacineMTB I keep telling everyone the future is 100% PHP WASM and they think that I'm joking...
@yacineMTB If you could document and annotate your steps to share would be very nice, I tried once and felt stuck at a very early stage and gave up :)
@yacineMTB please write an article on it!
@yacineMTB moore's law ended and now we just rent bigger gpus and pretend it's progress
@yacineMTB i have to check this out