ted_dunning
Joined 12 years ago
Joined 12 years ago
Or because you have a history of being in that position and feel like it is important to pass the knowledge on?
That might be changing (slightly) because of the Tamago work on NOOS.
The OP article here is nice because it allows selective enforcement of allocation free code. If you have code that has some discipline about heap allocation such as having all small allocations then you are OK on the fragmentation front.
| ... cybertruck of sports cars ...
ECMWF ASML Pumas JPMorgan Chase
and many other companies.
I use it routinely, but it isn't a company standard where I work.
This is similar to the way that $\int_{-\infty}^\infty exp(-x^2) dx = \sqrt \pi$. The answer is a constant that doesn't give you any information about the function erf(x).
https://arxiv.org/abs/2603.00042
It is true that block-headed quantization of everything doesn't work. But, as I am sure will read, if you remove the spiky parts of the parameter values, the residue can be dramatically quantized and compressed while retaining performance.
This is a multi-modal sort of compression where you use different techniques for different phenomena. Simply compression all of the weights, each in isolation, ignores the benefits of compressing them collectively.
Biodiesel or biomethane looks at first glance to escape this, but it has serious other costs.