Tootfinder

An Efficient, Reliable and Observable Collective Communication Library in Large-scale GPU Training Clusters
Ziteng Chen (Infrawaves), Xiaohe Hu (Infrawaves), Menghao Zhang (Beihang University), Yanmin Jia (Infrawaves), Yan Zhang (Infrawaves), Mingjun Zhang (Infrawaves), Da Liu (Infrawaves), Fangzheng Jiao (Beihang University), Jun Chen (Infrawaves), He Liu (Infrawaves), Aohan Zeng (Tsinghua University), Shuaixing Duan (Zhipu AI), Ruya Gu (Infrawaves), Yang Jing (Infrawaves), Bowen Han …

An Efficient, Reliable and Observable Collective Communication Library in Large-scale GPU Training Clusters
Large-scale LLM training requires collective communication libraries to exchange data among distributed GPUs. As a company dedicated to building and operating large-scale GPU training clusters, we encounter several challenges when using NCCL in production, including 1) limited efficiency with costly and cumbersome P2P communication, 2) poor tolerance to frequent RNIC port failures, and 3) insufficient observability of transient collective communication anomalies. To address these issues, we pro…

Affinity being bought by Canva was disappoiting but what really bugs me is the campaign they are building right now. Its slogan ""Creative Freedom is coming" sounds more like a threat for me. This smells of the idea of cloud service being sold as freedom as if it was the 2010s.

In our last #ISE2025 lecture last week, we were discussing what makes a node "important" in a knowledge graph. A simple heuristics can be borrowed from graph theory or communication theory: Degree Centrality
Interestingly, in Wikidata In-degree centrality states Jane Austen to be to most "important" female author, while Out-degree centrality claims J.K. Rowling as being…

"Mexico adopts protections for Atlantic sharks"
#Mexico #Sharks #Animals

Mexico adopts protections for Atlantic sharks
Mexico recently adopted national regulations protecting several threatened shark species in the Atlantic from being caught or retained as bycatch. Shark conservationists welcome the protections but say they are long overdue, coming years after the country’s commitments to a multilateral fishery regulator. Mexican fisheries catch a significant number of various shark species in the Atlantic […]

While I do maintain that "it's coming from the LAN" is not a good #security boundary, there are services where it is practical (eg. media center volume control), but also fault prone (oups my phone just switched to LTE for power saving – a generally justified thing).
Before I start formalizing how "a device can retain permissions it gets from being local for a few days&quo…

#WWE are coming back to Europe for another tour in the new year ahead of the #RoyalRumble
including:
* Monday Night RAW TV Taping on Monday 19th January from the SSE Arena in #Belfast

#photography #feet #red

I'm especially interested in a few things if anyone is willing to help...
1) missing words and simple typos - My ADHD brain skips words coming out and fulls them in when reading, so it's easy for me to make mistakes and hard for me to catch them.
2) questions - I tend to work from a lot of assumed knowledge, collected from all over, and I'm really trying to make my work more accessible. I assume I'm talking about a bunch of stuff most folks don't know, but I don't know which of them come from some rabbit hole I went down and which are more common knowledge.
3) challenges - I get bored when having things explained to me so I have a tendency to keep explanations light... Which can mean I leave out a bunch of critical context or logical steps.

In our last #ISE2025 lecture last week, we were discussing what makes a node "important" in a knowledge graph. A simple heuristics can be borrowed from graph theory or communication theory: Degree Centrality
Interestingly, in Wikidata In-degree centrality states Jane Austen to be to most "important" female author, while Out-degree centrality claims J.K. Rowling as being…

FlowMoE: A Scalable Pipeline Scheduling Framework for Distributed Mixture-of-Experts Training
Yunqi Gao, Bing Hu, Mahdi Boloursaz Mashhadi, A-Long Jin, Yanfeng Zhang, Pei Xiao, Rahim Tafazolli, Merouane Debbah
https://arxiv.org/abs/2510.00207

FlowMoE: A Scalable Pipeline Scheduling Framework for Distributed Mixture-of-Experts Training
The parameter size of modern large language models (LLMs) can be scaled up via the sparsely-activated Mixture-of-Experts (MoE) technique to avoid excessive increase of the computational costs. To further improve training efficiency, pipelining computation and communication has become a promising solution for distributed MoE training. However, existing work primarily focuses on scheduling tasks within the MoE layer, such as expert computing and all-to-all (A2A) communication, while neglecting ot…