moyix

13 ms·

121.
by moyix 4y ago
Bernard Berenson is only very briefly mentioned in the article, but he's an interesting figure in his own right. He is credited with introducing and popularizing the idea of authenticating paintings by the artist's characteristic
122.
by moyix 4y ago
The Authors Guild v Google decision about Google Books seems relevant: > In late 2013, after the class action status was challenged, the District Court granted summary judgement in favor of Google, dismissing the lawsuit and affirming th
123.
by moyix 4y ago
Nope. The way it works is FasterTransformer splits the model across the two GPUs and runs both halves in parallel. It periodically has to sync the results from each half, so it will go faster if you have a high-bandwidth link between the GP
124.
by moyix 4y ago
Yes, you can do 2x4090s as well. NVLink is not required (though it will make things a bit faster).
125.
by moyix 4y ago
I don't have any to test on unfortunately! But it's a wiki; if you get some of the models running on those (it should be as easy as running ./setup.sh) please add a line saying that it works!
126.
by moyix 4y ago
Yep, I definitely agree that the 6B and below models are worse than Copilot. The 16B ones are pretty good! But IMO still undertrained compared to Copilot, and of course much less accessible (though see elsewhere in this comment section; I t
127.
by moyix 4y ago
I will inform Andrei and Katya that they have a fan :) Although they already have a pretty high opinion of themselves!
128.
by moyix 4y ago
I believe right now the VSCode extension just passes along the entire file up to your cursor [1] rather than trying to figure out how much will fit into the context limit – it's definitely still very early stages :) It would be pretty
129.
by moyix 4y ago
Well, on the FauxPilot side I did do the initial implementation personally (over the course of a few weeks this summer) – but of course I'm building on the work the Salesforce team put into training the CodeGen models, and the work NVI
130.
by moyix 4y ago
It shouldn't require retraining, nope. I believe for INT4 there is a small adapter layer added that needs to be trained, but it is small and wouldn't require much data or computation to do so. Will know more once we actually start
131.
by moyix 4y ago
The smaller models run on smaller GPUs too! :) You can see how much VRAM is required for various models in the Documentation: https://github.com/moyix/fauxpilot/blob/main/documentation/s... And we h
132.
The VSCode GitLab extension now supports getting code completions from FauxPilot(twitter.com)
190 points by moyix 4y ago|65 comments
133.
by moyix 4y ago
Salesforce CodeGen (particularly the 16B-multi and 16B-mono models) is pretty good already and can be used with FauxPilot [1] to get an open Copilot-like experience with local compute :) I am also very excited about the upcoming BigCode pro
134.
by moyix 4y ago
You can download it here: https://github.com/rom1504/img2dataset/blob/main/dataset_exa... You probably would want to stop after getting the metadata, unless you have 240TB available for the images :) Mor
135.
by moyix 4y ago
Agreed (especially since I just got a third A6000 this week), but by the time they come out the faster training/inference time might be enticing to some. I would love to see an 80GB workstation card though...
136.
by moyix 4y ago
Several groups already have. Facebook's OPT-175B is available to basically anyone with a .edu address (models up to 66B are freely available) and Bloom-176B is 100% open: https://github.com/facebookresearch/metaseq
137.
by moyix 4y ago
The A6000s have become very popular for ML engineering, especially in the NLP space since those models are enormous. 48GB VRAM and 2x faster training time is tempting for those applications!
138.
by moyix 4y ago
Yep, it's targeted at the same kind of people who bought the A6000s (me).
139.
by moyix 4y ago
I just grabbed an A6000 for $3500 on eBay, so you can probably get a pretty decent deal on those now. They're pricey but IMO it's a great deal if you really need the VRAM (e.g. for training LLMs).
140.
Ffast 2 Furious(github.com)
1 points by moyix 4y ago|0 comments
141.
by moyix 4y ago
Just to be clear, cryptographic hashes and other outputs of cryptographic primitives are designed to be uniform random. If you find a detectable bias from uniform on the outputs of /dev/urandom then you should consider the underly
142.
by moyix 4y ago
I suspect the "ESP guy" is actually meant to refer to Daryl Bem? https://slate.com/health-and-science/2017/06/daryl-bem-prove... https://slatestarcodex.com/2014/04/28/
143.
by moyix 4y ago
"This is depressing for project creators. Early on, they have the triumphant experience of building exactly what they want, and solving their nagging problem. And the result is that they’re cleaning up after a bunch of requests from pe
144.
by moyix 4y ago
Thanks for this excellent idea! I implemented it in this histogram plotting demo: https://github.com/moyix/2_ffast_2_furious
145.
by moyix 4y ago
Ah you're right! I was misled by the `v4_float = {0x0, 0x0, 0x0, 0x0}` portion.
146.
by moyix 4y ago
I don't think this is correct. The return value of nextafterf is a float, and the ABI requires those to be returned in xmm0. So although nextafterf is using integers internally, the flush to zero happens inside of nextafterf when it
147.
by moyix 4y ago
Updates welcome to make it less contrived! :) It's clearly very confused code, but it does work under normal FP behavior.
148.
by moyix 4y ago
Not quite. nextafterf() is trying to return the smallest float after 0.0 in the direction of infinity (infinity is still valid even when the FPU is configured to flush denormal numbers to 0; infinity isn’t a denormal number). But since the
149.
by moyix 4y ago
All right fine, I give in. In current amd64 Debian unstable (main, contrib, and non-free), 48 of the 69,112 packages include a shared library built with -ffast-math, a total of 485 .so files. Actually it wasn't so bad :) All the curren
150.
by moyix 4y ago
Just a small joke :)

More ›