Explore a groundbreaking conference talk on the development of Jamba, a novel Foundation Model based on a hybrid Transformer-Mamba mixture-of-experts (MoE) architecture. Delve into the decision-making process behind creating this innovative hybrid structure, and gain insights into its layered composition of SSM, Transformer, and MoE components. Learn how this flexible architecture enables resource- and objective-specific configurations, offering unprecedented throughput and the largest context window of 256K in its size class, while fitting 140K on a single GPU. Discover how Jamba introduces a paradigm shift in large language model development, presented by AI21 Labs CTO Barak Lenz. Access additional resources on LLM and MLOps, and connect with Databricks through various social media platforms for further exploration of cutting-edge AI technologies.
Overview
Syllabus
Pioneering a Hybrid SSM Transformer Architecture
Taught by
Databricks
Related Courses
-
Create a Large Language Model from Scratch with Python – Tutorial
-
Transformer Models and BERT Model
-
Understanding Snowflake Arctic 480B - A Mixture of Experts LLM Architecture
-
Stanford Seminar - Mixture of Experts Paradigm and the Switch Transformer
-
Mamba: Linear-Time Sequence Modeling with Selective State Spaces
-
Mixture of Experts (MoE) in Large Language Models - A Simple Guide
