@drisspg profile picture
driss guessous @drisspg
Joined December 2023
  • Tweets 497
  • Following 254
  • Followers 1,522
78 Photos and videos
Agreed
Aaron Gokaslan
@SkyLi0n
Jun 10
Inevitable: nobody wants to allow the use of an LLM that purposefully inserts bugs.
1
447
Holy chart crime
Cursor
@cursor_ai
Jun 9
Claude Fable 5 is now available in Cursor. It sets a new state of the art on CursorBench at 72.9%, 8 points above the previous best.
44
27
1,420
234,915
Shower thought: what’s the most valuable bit string per bit? On one end you have a windows key, the other end maybe 16tb of mythos weights? No cheating and saying some random string in library of babel. It’s probs a bitcoin wallet though - sad
4
7
1,317
Where my async proxy kings at; github.com/pytorch/pytorch/i… cc @gaunernst I feel like this one is probably easy for you

[vllm] [2.12 regression][B200] test_batch_invariance: nondeterministic outputs 3/5 trials with...

Summary Under torch 2.12.0 triton 3.7.0, vLLM's test_v1_generation_is_deterministic_across_batch_sizes_with_needle[FLASH_ATTN] fails on B200 because outputs diverge across batch sizes: Failed...

github.com
1
2
33
2,403
Ohh you use worktrees, nice! I just use different b200 nodes
46
3,141
Ahhh yes 'uv' one of humanity's greatest scientific achievements
1
18
1,690
I am trying to make ideogram usable on my spark; Problem 1. github.com/ideogram-oss/ideo… Problem 2. Bitsandbytes is unbelievable slow

Speed up quantized transformer loading by drisspg · Pull Request #9 · ideogram-oss/ideogram4

Summary I was going through the example and on my spark it takes ~2 minutes to get to first image. Using the run_inference.py entry point. Profiling showed 0.0s start 0.5s state dict fi...

github.com
3
2
5
1,923
Okay that's enough codex for now-> at least now I can finally look at the generation step
1
431
"The purpose of abstracting is not to be vague, but to create a new semantic level in which one can be absolutely precise." This really is a nice quote
Modal
@modal
Jun 1
Reinforcement learning has exploded on Modal, and we've been cooking. Here's a review of lessons learned helping teams train at scale, the patterns we kept seeing, and an open-source library to get started with RL on Modal quickly.
0:05
1
14
1,770
Lol it took 7 hours for it to find GemmUniversal and do some hyper parameter tuning. Can't you feel the AGI!!!
3
2
57
4,465
To hell with big TMA long live ld/st
3
4
38
3,440
driss guessous retweeted
Han Guo
@HanGuo97
May 21
LLM training is built on fast MatMuls. But many surrounding ops still run as memory-bound kernels. CODA reparameterizes them to hide in the matmul’s shadow, fused into its epilogue before results leave the chip. Bonus: LLMs can write fast CODA kernels too (approaching SoLs).
15
103
685
197,791
Omni looks really cool, everything else is so mehh
4
639
driss guessous retweeted
Degen CPA
@DrewVento
May 19
BREAKING: Victor Wembanyama has joined Anthropic.
58
319
7,089
359,246
slop begets slop
2
334
If you couldn’t be bothered to write it why in the world would I read it
1
4
66
5,875
Yooooooooo so like what did OAI do with all those sora face scans
9
1,074
the death of rigor happens one prompt at a time
7
431
fun problem; given an integer function f and a domain D, which functions preserve ordered contiguity for every consecutive subspan of D?; winner gets 4x faster FlexFlash Attention (for some mask mods)
2
1
13
2,242
Im on a quest to inject Pi into everything; github.com/drisspg/pi-review is this weekends project. Github review is not keeping up with the new asymmetric world. PRs are basically inverted NP -> trivial to produce hard to verify. It is the age of personal software!

GitHub - drisspg/pi-review

Contribute to drisspg/pi-review development by creating an account on GitHub.

github.com
9
983
Load more