91–100 of 334 posts
Andrej doesn't need to do this. He's done it because he evidently loves it, and wants to share his hard-earned knowledge with the rest of the world. He may be a product of the ivory tower, but he's been in the trenches . He knows firsthand how f-ing hard it is to ship a product. And here he is, sharing useful personal code with everyone. This github repo now has collected ~4K stars and its antecessor (minGPT) has col…
Andrej doesn't need to do this. He's done it because he evidently loves it, and wants to share his hard-earned knowledge with the rest of the world. He may be a product of the ivory tower, but he's been in the trenches . He knows firsthand how f-ing hard it is to ship a product. And here he is, sharing useful personal code with everyone. This github repo now has collected ~4K stars and its antecessor (minGPT) has col…
Edit: the OP has updated their wording to make it clear they meant any kind of viewing or usage. I don’t think any of us would disagree more people use code than star repos. Original comment left below with original quote, since this has gotten a number of replies that would stop making sense with a larger edit. > Normally, the number of people who clone or copy code from a repo is one to two orders of magnitude larg…
Of course not everyone is the same, but I'd be surprised if overall clones were less than an order of magnitude more than forks or stars, and find two or even three orders of magnitude believable depending on the target group of the repo.
This is really good, and I was really excited by it but then I read: > running on a single 8XA100 40GB node in 38 hours of training This is a $40-80k machine. Not a diss, but I would love to see an advance that would allow anyone with a high end computer to be able to improve on this model. Before that happens this whole field is going to be owned by big corporations.
Earlier quoted context omitted.
Edit: the OP has updated their wording to make it clear they meant any kind of viewing or usage. I don’t think any of us would disagree more people use code than star repos. Original comment left below with original quote, since this has gotten a number of replies that would stop making sense with a larger edit. > Normally, the number of people who clone or copy code from a repo is one to two orders of magnitude larg…
If I want to use a repository, my first step is to either download a released binary or clone the repository. Forking is much further down the line for me, when I've used the code, encountered a problem, fixed it, and decided to polish the fix up to make a PR. I star something when I either have used it and like it, or when I think I want to use it in the future and want to bookmark it (though the former more often t…
My estimate of "one to two orders of magnitude" is based on anecdotal evidence. I edited my comment to reflect as much.
This is really good, and I was really excited by it but then I read: > running on a single 8XA100 40GB node in 38 hours of training This is a $40-80k machine. Not a diss, but I would love to see an advance that would allow anyone with a high end computer to be able to improve on this model. Before that happens this whole field is going to be owned by big corporations.
Edit: or about $340 if you get the 8xA100 instance from lambdalabs, in the realm of normal hobby spending
This is really good, and I was really excited by it but then I read: > running on a single 8XA100 40GB node in 38 hours of training This is a $40-80k machine. Not a diss, but I would love to see an advance that would allow anyone with a high end computer to be able to improve on this model. Before that happens this whole field is going to be owned by big corporations.
You can rent on AWS and other cloud providers.
Are there any possible technologal or scientific leaps on the horizon that would reduce training time by an order of magnitude or more? GPT-3 took 355 years to train with incredibly expensive hardware, which means small players have no chance to push the state of the art
Yes, see DeepMind RETRO: > In our experiments on the Pile, a standard language modeling benchmark, a 7.5 billion parameter RETRO model outperforms the 175 billion parameter Jurassic-1 on 10 out of 16 datasets and outperforms the 280B Gopher on 9 out of 16 datasets. https://www.deepmind.com/blog/improving-language-models-by-r... Though, there hasn't been much follow-up research on it (or DeepMind is not publishing it)…
RETRO did get press, but it was not the first retrieval model, and in fact was not SOTA when it got published; FiD was, which later evolved into Atlas[0], published a few months ago.
This is really good, and I was really excited by it but then I read: > running on a single 8XA100 40GB node in 38 hours of training This is a $40-80k machine. Not a diss, but I would love to see an advance that would allow anyone with a high end computer to be able to improve on this model. Before that happens this whole field is going to be owned by big corporations.
> This creates a much smaller Transformer (4 layers, 4 heads, 64 embedding size), runs only on CPU, does not torch.compile the model (torch seems to give an error if you try), only evaluates for one iteration so you can see the training loop at work immediately, and also makes sure the context length is much smaller (e.g. 64 tokens), and the batch size is reduced to 8. On my MacBook Air (M1) this takes about 400ms per iteration. The network is still pretty expensive because the current vocabulary is hard-coded to be the GPT-2 BPE encodings of vocab_size=50257. So the embeddings table and the last layer are still massive. In the future I may modify the code to support simple character-level encoding, in which case this would fly. (The required changes would actually be pretty minimal, TODO)
This is really good, and I was really excited by it but then I read: > running on a single 8XA100 40GB node in 38 hours of training This is a $40-80k machine. Not a diss, but I would love to see an advance that would allow anyone with a high end computer to be able to improve on this model. Before that happens this whole field is going to be owned by big corporations.
You can rent on AWS and other cloud providers.
Edit: i just noticed lambda lab. It seems they ask $8.8 per hour for an instance of this caliber. That puts the total training cost around $334. I wonder how come it is that much cheaper.
This is really good, and I was really excited by it but then I read: > running on a single 8XA100 40GB node in 38 hours of training This is a $40-80k machine. Not a diss, but I would love to see an advance that would allow anyone with a high end computer to be able to improve on this model. Before that happens this whole field is going to be owned by big corporations.