bytes and nuggets @pytorch drisspg.github.io/nuggets/
- Tweets 657
- Following 298
- Followers 1,668
- Likes 640
drisspg.github.io/nuggets/KD…
wrote up a new one; lil bit product pitch, lil bit click bait
2
18
133
16,606
Corpo appreciation post; I spend obscene amounts of money on tokens. And all of those tokens go towards helping make OSS software better/faster/stronger. I think that is neat
1
6
360
I am more and more convinced that products like modal will be the future; it is very possible with today's models to build entirely locally and just fast submit to big gpu pools. python dsl hav made this quite convenient to even for quick iteration cycles
10
498
It gets worse github.com/pytorch/pytorch/p…
it turns out
"Consider setting torch.set_float32_matmul_precision('high') for better performance."
was worth paying attention to
7
71
26,552
Be me: "Ohh wow this looks cool I wonder what their code is like"
github.com/OpenVDN/vdn-minim…
"Create the environment. We recommend PyTorch 2.13 (torch.__version__ = 2.13.0+cu129) and installing FlashAttention 4, since our code requires FlexAttention's Flash backend."
Nice!
Open-source video generation is now faster than playback without compromising quality.
Introducing Video Delta Net (VDN): hybrid attention for live text-to-video with near-lossless quality.
VDN accelerates Minimax-H3 by 75 - 90 x, generating 14 seconds of 768p video in 11 seconds on 8× NVIDIA B200 GPUs.
Checkpoints + training/inference code + Technical Blog ⬇️
(1/6)
6
3
69
19,191
Here I wrote a thing: meta-pytorch.org/attention-g…
as always I'm human and it probs has some mistakes :) enjoy nonetheless
7
9
149
12,832
Some future AI will look back at the stratification of GitHub code base and glean so much.
"Ohh that's the opus 4.5 Cambrian explosion of try; except".
"That's the 5.6 sol period; you can tell by the insane input validation"
What a glorious future it will be
1
20
1,102
I am yet again about to try and prompt a 3rd GitHub stack pr framework; I want the style of ghstack (commits are the atomic unit) but it to renders in native GH stacking. Anyone else got this built out already?
3
4
1,119