Viewing profile — SekstiNi

> I wish it was possible to download that 36GB books3.tar in the wild these days. There... is a torrent.

There is one trained on 600B tokens from SlimPajama [1], but that's fairly tiny compared to other recent releases (ex. stablelm-3b [2] trained on 4T tokens). > low quality data (th…

I've seen this sentiment echoed a lot, and I find it somewhat perplexing as my own twitter experience has been mostly pleasant. Do people simply not bother curating their follows?

> Apple Silicon was great because Apple had invested huge amounts of resources for a decade on smartphone processors first, not because they had some kind of geniuses on the projec…

There's actually a good reason. In short, a large portion of the apple silicon team ended up leaving a few years ago to could start a new company named Nuvia. Their goal was to pro…

Surprised people don't know about this, as it has been common knowledge in the SD community [1] since october last year. Strictly speaking you don't even need cuda 11.8+ to get the…

They took down the blogpost, but from what I remember the model is composite and consists of a text encoder as well as 3 "stages": 1. (11B) T5-XXL text encoder [1] 2. (4.3B) Stage …

> Interesting, though apparently the OPT175B model is 350GB: Only in FP16. In the paper they use int4 quantization to reduce it to a quarter of that. In addition to the model weigh…

That would indeed be linear time, but he doesn't make that suggestion as far as I can see.

I had this happen to my XPS 15 when I was a student, and within 2 years of purchase the battery had degraded by 80%. If I recall correctly it would last about 20 minutes before abr…

https://pytorch.org/get-started/pytorch-2.0/#requirements

Here's a writeup exploring this very topic, and showing how Go's generics will currently give worse performance than just using an interface. https://planetscale.com/blog/generics-…

Just to be sure, you did compile the Rust program using the --release flag?

Depends on how much of the screen is being refreshed, among other factors, but for writing the rM2 is apparently down to 24ms. [1] [1] https://www.youtube.com/watch?v=c5XUTnPQ5i4

The Firestorm cores are capable of issuing two 128bit stores per cycle [1], giving a bandwidth of 102.4 GBps. This matches the experiments by Andrei with mixed reads/writes [2], an…

Unfortunately, but expectedly, it cannot reason about it [1]. Even if it could I'm not sure it would be allowed to elide the allocation(s). [1] https://godbolt.org/z/63xEcfaEE

This is the way. Additionally I went into about:config and set browser.proton.enabled to false, which fixed most of the issues I had with the new design.

From the article: > Hitting nearly 1TFlops in the browser (50% of peak) is extremely empowering and it's exciting to see such technology available.

Looking at the die shot [1], they have plenty of space for cores and cache if they remove the GPU. Surely it's not that simple, but I believe they should be able scale to at least …

Interesting, I ran it on my laptop (i7-7700HQ) with the following results: - 2-wise+ is 19% slower than 2-wise - 3-wise is 48% slower than 2-wise - 3-wise is 25% slower than 2-wise…

> There's not enough info to figure out what's going on. If you only look at the article this is true. However, the source code is freely available: https://github.com/lemire/Code-…