HackerTrans
TopNewTrendsCommentsPastAskShowJobs

gkapur

no profile record

Submissions

Surfacing a 60% performance bug in cuBLAS

kernelspace.substack.com
10 points·by gkapur·2 месяца назад·0 comments

A Principled ML Compiler Stack in 5k Lines of Python

kernelspace.substack.com
1 points·by gkapur·2 месяца назад·0 comments

Agentic systems redraw the Pareto frontier on ARC-AGI

poetiq.ai
7 points·by gkapur·8 месяцев назад·1 comments

comments

gkapur
·2 месяца назад·discuss
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 месяцев назад·discuss
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.