Pinned post
אין לי מספיק אישיות לציוץ נעוץ
Ariel Bereslavsky
@ArielBAMath
אבא של סנדי (מאמי שלי כלבה חמודה) וכותב מאמר בזמני הפנוי
752 Following 276 Followers 2.8K Posts
Latest posts 5
Scaling Expert Parallelism across nodes? Compute isn't your bottleneck anymore. The network is.
I wrote a breakdown on why wide-EP serving for MoE models is fundamentally network-bound.
getimpala.ai/blog/wide-ep-s…
I wrote a breakdown on why wide-EP serving for MoE models is fundamentally network-bound.
getimpala.ai/blog/wide-ep-s…
www.getimpala.ai Why Wide-EP Serving Becomes Network-Bound | Impala AI Blog A first-principles communication model for multi-node MoE serving, and why per-rank payloads decide when the wire becomes the bottleneck. I couldn't find a Python implementation of the Jonckheere-Terpstra test for my current paper, so I decided to build one myself.
Just released it as an open-source package. If you’re working with ordered alternatives in statistics, check it out! 📊
github.com/ScarWar/jonckh…
Just released it as an open-source package. If you’re working with ordered alternatives in statistics, check it out! 📊
github.com/ScarWar/jonckh…
github.com GitHub - ScarWar/jonckheere-test Contribute to ScarWar/jonckheere-test development by creating an account on GitHub. What an excellent paper!
🚨 New Paper Alert! 🚨View quoted post on X
Characterizing equivariant layers has traditionally relied on parameter-sharing, a complex and often tedious process where finding explicit schemes is often tricky. We’ve taken a simpler approach with promising results. (1/8) 🧵👇
Amazing work
🚨New paper on SLM evaluation🚨
We present SALMon🍣 which is a suite of benchmarks for evaluating how much Speech Language Models model acoustic elements like sentiment or background noise.
Project: pages.cs.huji.ac.il/adiyoss-lab/sa… 🧵👇🏻View quoted post on X
