bekantan
Joined 6 years ago
Joined 6 years ago
However, I don't think that underlying array is resized every time `add` is called. I'd expect that resize will happen less than 30 times for 1M adds (capacity grows geometrically with a=10 and r=1.5)
It was so cool to see them diving around.
Interestingly, when I showed the clips to some of my senior family members, they didn't seem interested at all. I think they couldn't comprehend what was going on, even after I explained.
Their (several independent trials) reaction was similar to showing them some AI-generated image of something which clearly can't exist. It was so absurd that it was just filtered out with a comment "yeah, yeah - nice car".
That was my experience as well - 3-bit version is pretty good.
I also tried 2-bit version, which was disappointing.
However, there is a new 2-bit approach in the works[1] (merged yesterday) which performs surprisingly well for Mixtral 8x7B Instruct with 2.10 bits per weight (12.3 GB model size).
> Operated by the governing Ustaše regime, Europe's only Nazi collaborationist regime that operated its own extermination camps
> It quickly grew into the third largest concentration camp in Europe
> Unlike German Nazi-run camps, Jasenovac lacked the infrastructure for mass murder on an industrial scale, such as gas chambers. Instead, it "specialized in one-on-one violence of a particularly brutal kind", and prisoners were primarily murdered with the use of knives, hammers, and axes, or shot
> Ustaše regime having murdered somewhere near 100,000 people in Jasenovac between 1941 and 1945
--
Flower Monument on Spomenik Database: https://www.spomenikdatabase.org/jasenovec
Then I asked it to tell more about the specific item and it gave me a decent answer (model: pplx-7b-online).
Pretty good experience so far.
What would be x86 alternative in that price range (if any)? Xeons with HBM are more expensive IIRC