HackerTrans
トップ新着トレンドコメント過去質問紹介求人

gkapur

no profile record

投稿

Surfacing a 60% performance bug in cuBLAS

kernelspace.substack.com
10 ポイント·投稿者 gkapur·2 か月前·0 コメント

A Principled ML Compiler Stack in 5k Lines of Python

kernelspace.substack.com
1 ポイント·投稿者 gkapur·2 か月前·0 コメント

Agentic systems redraw the Pareto frontier on ARC-AGI

poetiq.ai
7 ポイント·投稿者 gkapur·8 か月前·1 コメント

コメント

gkapur
·2 か月前·議論
On the limitation side:

Do you think this would scale to larger transformer models with more parameters per layer?

How would this work with MOE models or sparse models?
gkapur
·9 か月前·議論
Adding to the prior comments as my intuition matched yours, there’s a nice Reddit thread that gives some context into how it can be faster even if you require exact matches: https://www.reddit.com/r/LocalLLaMA/s/ARxHLqRjdM

The TLDR/key (from my understanding) is that verifying N tokens can be faster than generating N tokens.