Efficient Systems for Foundation Models

Workshop at the International Conference on Machine Learning (ICML) 2024.

Banner Code it, run it, crash it–restart it.

➡️ ES-FoMO is back for ICML 2024! Find us in room Lehar 2, and check-out the schedule below, and the accepted papers on OpenReview.

🔥 the gist

  • what? A workshop to bring together interdisciplinary experts working on the emerging research questions and challenges associated with foundation model training and inference.
  • when & where?
  • questions? Contact us at esfomo.workshop@gmail.com.
  • looking for the 2023 edition?


📆 the plan

All times CET, UTC+2. Full schedule to be confirmed.

  Topic Speaker
  🎛️ Session I: Quantization, Pruning, and Sparsity  
9:00am Talk: Efficient Quantization Methods and Marlin, a Fast 4-Bit Inference Kernel Elias Frantar
(IST Austria)
9:30am Oral: Prompt-prompted Adaptive Structured Pruning for Efficient LLM Generation Harry Dong
(CMU)
9:45am Oral: Lottery Ticket Adaptation: Mitigating Destructive Inference in LLMs Ashwinee Panda
(Princeton)
10:00am Coffee break  
  🦾 Session II: Emerging Architectures  
10:15am Oral: Simple Linear Attention Language Models Balance the Recall-Throughput Tradeoff Simran Arora
(Stanford)
10:30am Oral: xLSTM: Extended Long Short-Term Memory Maximilian Beck
(JKU)
10:45am Talk: On the Tradeoffs of State-Space Models Albert Gu
(CMU)
11:15am Talk: Scaling Mixture-of-Experts: Lessons from DBRX Vitaliy Chiley
(Databricks)
11:45am Oral: Characterising Prompt Compression Methods for Long Context Inference Siddharth Jha
(UCB)
noon Lunch break  
1:00pm 🧑‍🎓 Poster Session  
2:15pm 🏅 Best Paper and Best Poster Awards  
  🔥 Session III: Hardware  
2:30pm Talk: Scaling Intelligence Azalia Mirhoseini
(Stanford / Google DeepMind)
3:00pm Off-the-record: Frontier Clusters for Frontier Models: Scaling to 100,000 GPUs and Beyond Dylan Patel
(SemiAnalysis)
3:30pm Coffee break  
3:45pm 💬 Panel: Data and Architecture Trends Across Industry and Open Communities  
  Deepak Narayanan (NVIDIA), Dylan Patel (SemiAnalysis), Dirk Groeneveld (AI2), Hailey Schoelkopf (EleutherAI)  
4:30pm 💾 Session IV: Data  
  Open Tooling for Large Data Pipelines Vaishaal Shankar
(Apple)
6:00pm 🎉 Post-workshop happy hour  


🦾 the pitch

As models increase in size and training budget, they not only systematically improve in upstream quality, but also exhibit novel emergent capabilities. This increase in scale raises proportionate difficulties for practitioners: foundation model training and inference lie at a unique interdisciplinary crossroad, combining open problems in algorithms, system design, and software engineering.

Machine learning practitioners are key stakeholders here: on the one hand, researchers may contribute algorithmic insights and novel methods to improving training and inference of large models; on the other hand, novel research findings may be best demonstrated at scale—which may require training models as efficiently as possible to make the best use of available resources.

The goal of this workshop is to bring together interdisciplinary experts working on the emerging research questions and challenges associated with foundation model training and inference. We welcome submissions around training and inference systems/algorithms for foundation models, focusing on scaling-up or on reducing compute, time, memory, bandwidth, and energy requirements. Notably, we encourage submissions concerning the entire spectrum of foundation models: from BERT-sized Transformers, to large models with 100B+ parameters. Topics include but are not limited to (see our 📝 call for papers for details):

  • Training and inference systems, either distributed at large scale or in resource-constrained scenarios;
  • Algorithms for improved training and inference efficiency;
  • Systems for foundation models, such as novel programming languages or compilers.

This is the second installment of ES-FoMo; we are bringing further focus in our sessions and talks on three trends observed in 2023:

  • The emergence of novel architectures, popularized by Mamba (state-space models) and Mixtral (mixture-of-experts);
  • Efficient open implementations, such as gpt-fast and vLLM;
  • Open questions on novel hardware and data tooling.


🧑‍🏫 the speakers


💬 the panelists (& moderators)


😎 the organizers