Tom St. John

Technical lead for MTIA training performance at Meta AI; Ph.D. from University of Delaware; Brazilian jiu-jitsu and boxing practitioner; cat owner

[Tom St. John @tstjohn_hpcml]

@chipro Congrats!

[Tom St. John @tstjohn_hpcml]

@aronchick This sounds like an excerpt from Spinal Tap.

https://t.co/9X59gtGVc2

[Demis Hassabis @demishassabis]

The Gemini era is here. Thrilled to launch Gemini 1.0, our most capable & general AI model. Built to be natively multimodal, it can understand many types of info. Efficient & flexible, it comes in 3 sizes each best-in-class & optimized for different uses https://t.co/VUu1277bC2

[MLCommons @MLCommons]

Introducing the AlgoPerf: Training Algorithms Benchmark! Compete for a share of the $50,000 prize pool by submitting more effective and efficient neural network training algorithms. Learn more https://t.co/csfXkUCKbN

#Algorithms #MachineLearning #Competition

[MLCommons @MLCommons]

SC23 attendees join MLCommons BOF sessions to add your voice to the @MLCommons community. Wed, 11/15, 5:15pm in Rm 601-603 MLPerf: A Benchmark for Machine Learning, or in Rm 702 join the conversation around the Future of Benchmarks in Supercomputing. https://t.co/xjuOrQqXjW #SC23

[Google AI @GoogleAI]

Today on the blog, learn how we’re supporting a new effort by the non-profit MLCommons Association that aims to bring together expert researchers across academia and industry to develop standard AI safety benchmarks that everyone can use and understand. ↓ https://t.co/stExJ386UD

[Tom St. John @tstjohn_hpcml]

@ShriramKMurthi Based on my experience, the professor doesn't contribute much of anything to the papers written by their students.

[Sharon Zhou ✈️ ICML @realSharonZhou]

Excited to announce a HUGE secret with @LisaSu: @LaminiAI has been building LLMs on @AMD GPUs *in production* for over the past year!

We’ve made running LLMs on AMD super easy and a highly competitive option through our LLM Superstation, available now at ~10x lower cost than cloud. 👉🏻 https://t.co/wqx9d8mPOK

Our enterprise customers have already built *thousands* of private LLMs on @LaminiAI LLM Superstations, e.g. @iFit leading at-home fitness with millions of users and @AMD itself:

🚀 Easy & fast: “It was simple to iterate and deploy with a few lines of code and amazingly fast with the AMD Instinct™ hardware.”

⭐️ LLMs are the new IP: “Using a public LLM wasn’t enough: we needed something that we could easily and quickly personalize to our customers’ data and constantly improve on new data, while keeping all of our data private.”

⚙️ Any infrastructure: “We’ve deployed Lamini in our internal Kubernetes cluster with AMD Instinct GPUs, and are using finetuning to create models trained on [our data].”

We had a cameo quote from Joe Spisak at @MetaAI who leads the Llama efforts said: “…Llama 2 is becoming the foundation of some of the most innovative companies.”

🫱🏼‍🫲🏾 Join Fortune 500 enterprises, and get your own private LLM Superstations—hosted, VPC, or on-premise (just 2 questions): https://t.co/ECVOsux9ay

More (technical) details here👉🏻 https://t.co/wqx9d8mPOK

[Tom St. John @tstjohn_hpcml]

@HPC_Guru @Tesla @nvidia @tomshardware So much for their claims that they can train their networks with Dojo

[Tom St. John @tstjohn_hpcml]

If you're planning to attend @hotchipsorg, come check out our ML inference tutorial on Sunday. We've got a great line-up of speakers from @NVIDIAAI, @berkeley_ai, @Qualcomm, @MetaAI, and Moffett AI.

#HotChips35
https://t.co/pMtdPbfOu8

[Tom St. John @tstjohn_hpcml]

@jabsNtriangles Lethal Weapon showed us what happens when someone knows jiu-jitsu (they hired Rorion Gracie to serve as a technical advisor)

https://t.co/jUHJZnCheb

[Tom St. John @tstjohn_hpcml]

@krismicinski My lab was able to do that for experimental architectures we were using (and there was a lot of red tape), but there were exclusive contracts in place keeping us from doing that for any of our clusters.

[Tom St. John @tstjohn_hpcml]

@HPC_Guru @AmpereComputing @AMD @ServeTheHome Time will tell, but I think Ampere cutting their system level cache in half to accommodate the higher core count (compared to Altra) is going to hurt their performance in the long run.

[Abdulrahman Mahmoud @ARHmahmoud]

Call for applications for the inaugural Machine Learning and Systems Rising Stars 2023 workshop!

Website: https://t.co/kNbw8OmXMu

[Tom St. John @tstjohn_hpcml]

@karpathy Congratulations

[Tom St. John @tstjohn_hpcml]

@Meng2Fu Congratulations!

[Horace He @cHHillee]

Let's talk about a detail that occurs during PyTorch 2.0's codegen - tiling.

In many cases, tiling is needed to generate efficient kernels. Even for something as basic as torch.add(A, B), you might need tiling to be efficient! But what is tiling? And when is it needed?

(1/13)

[Vivek Natarajan @vivnat]

Delighted to share our new @GoogleHealth @GoogleAI @Deepmind paper at the intersection of LLMs + health.

Our LLMs building on Flan-PaLM reach SOTA on multiple medical question answering datasets including 67.6% on MedQA USMLE (+17% over prior work).

https://t.co/jZZuFDrxGw

[Xavier Bresson @xbresson]

Our paper "Benchmarking Graph Neural Networks" has been accepted for publication at Journal of Machine Learning Research @JmlrOrg!
https://t.co/OXq9uJwt9u

(after rejection from NeurIPS, ICLR and ICML :)