SekstiNi

7 ms·

1.
by SekstiNi 3y ago
> I wish it was possible to download that 36GB books3.tar in the wild these days. There... is a torrent.
2.
by SekstiNi 3y ago
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 (the Pile only) The Pile is pretty good quality wise. I
3.
by SekstiNi 3y ago
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?
4.
by SekstiNi 3y ago
> 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 project. Given that they were confident enough to leave and st
5.
by SekstiNi 3y ago
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 produce high performance chips for the enterprise/ser
6.
HuggingFace support and evals for Mistral's new 7B model(twitter.com)
1 points by SekstiNi 3y ago|0 comments
7.
by SekstiNi 3y ago
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 speedup; it's sufficient to use cuDNN 8.6+,
8.
by SekstiNi 3y ago
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 1 UNet 3. (1.3B) Stage 2 upscaler (64x64 -> 256
9.
by SekstiNi 4y ago
> 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 weights, there's also a KV cache that takes up considerab
10.
by SekstiNi 4y ago
That would indeed be linear time, but he doesn't make that suggestion as far as I can see.
11.
by SekstiNi 4y ago
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 abruptly turning off, making it completely useless as a laptop.
12.
by SekstiNi 4y ago
https://pytorch.org/get-started/pytorch-2.0/#requirements
13.
by SekstiNi 4y ago
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-can-make-your-go-code-...
14.
by SekstiNi 4y ago
Just to be sure, you did compile the Rust program using the --release flag?
15.
by SekstiNi 4y ago
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
16.
by SekstiNi 5y ago
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], and I have personally verified that it is attainable for
17.
by SekstiNi 5y ago
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
18.
by SekstiNi 5y ago
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.
19.
by SekstiNi 6y ago
From the article: > Hitting nearly 1TFlops in the browser (50% of peak) is extremely empowering and it's exciting to see such technology available.
20.
by SekstiNi 6y ago
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 8+4 cores without large interconnect changes, and at th
21.
by SekstiNi 6y ago
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+ However, as mentioned in the post the numbers can vary a l
22.
by SekstiNi 6y ago
> 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-used-on-Daniel-Lemire-s-bl