BEGIN:VCALENDAR
VERSION:2.0
PRODID:icalendar-ruby
CALSCALE:GREGORIAN
X-WR-CALNAME:Hardware-Aware Efficient Primitives for Machine Learning
X-WR-TIMEZONE:Eastern Time (US & Canada)
BEGIN:VEVENT
DTSTAMP:20260912T232452Z
UID:tag:localist.com\,2008:EventInstance_45642348927981
DTSTART:20240328T200000Z
DTEND:20240328T210000Z
DESCRIPTION:Efficiency is increasingly tied to quality to machine learning\
 , with more efficient training algorithms leading to more powerful models.
  However\, today's most popular machine learning models are built on asymp
 totically inefficient primitives. For example\, attention in Transformers 
 scales quadratically in the input size\, while MLPs scale quadratically in
  model dimension. In this talk\, I discuss my work on improving the effici
 ency of the core primitives in machine learning\, with an emphasis on hard
 ware-aware algorithms and long-context applications. First\, I focus on re
 placing attention with gated state space models (SSMs) and convolutions\, 
 which scale sub-quadratically in context length. I describe the H3 (Hungry
  Hungry Hippos) architecture\, a gated SSM architecture that matches Trans
 formers in quality up to 3B parameters and achieves 2.4x faster inference.
  Second\, I focus on developing hardware-aware algorithms for SSMs and con
 volutions. I describe FlashFFTConv\, a fast algorithm for computing SSMs a
 nd convolutions on GPU by optimizing the Fast Fourier Transform (FFT). Fla
 shFFTConv yields up to 7x speedup and 5x memory savings\, even over vendor
  solutions from Nvidia. Third\, I will briefly touch on how these same tec
 hniques can also be used to develop sub-quadratic scaling in the model dim
 ension. I will describe Monarch Mixer\, which uses a generalization of the
  FFT to achieve sub-quadratic scaling in both sequence length and model di
 mension. Throughout the talk\, I will give examples of how these ideas are
  beginning to take hold\, with gated SSMs and their variants now leading t
 o state-of-the-art performance in long-context language models\, embedding
  models\, and DNA foundation models.
GEO:42.363197;-71.127278
LOCATION:Science and Engineering Complex (SEC)\,  SEC LL2.224
SUMMARY:Hardware-Aware Efficient Primitives for Machine Learning
URL;VALUE=URI:https://events.seas.harvard.edu/event/hardware-aware_efficien
 t_primitives_for_machine_learning
CATEGORIES:Colloquia / Seminar / Lecture
END:VEVENT
END:VCALENDAR