Machine Learning frameworks for the gen AI era

Introduction to the concept of generative AI and its significance in the field of machine learning. Discusses the evolution of AI technology and the advancements in generative AI models.

Tracks the progression of machine learning frameworks, highlighting the impact of key releases like Transformer and Gemini on innovative models and AI technology. Emphasizes growth, adaptation, and innovation in the field.

Explores the challenges related to training and serving massive generative AI models, focusing on scalability, efficiency, and the increasing requirements in the industry. Introduces Caris and Jax as solutions to these challenges.

Details about Caris, a multi-framework deep learning API that provides access to powerful pre-trained models for generative AI. Discusses the features of Caris, including its support for computer vision and language models.

Explores the features and capabilities of Caris, such as pre-trained models, model adaptations, and memory management techniques like low-rank adaptation and model parallelism. Demonstrates fine-tuning a model using Caris.

Demonstrates the process of fine-tuning a large language model like Gemma using Caris, focusing on memory optimization techniques, model parallelism, and the reduction of trainable parameters. Shows the training progress and the successful fine-tuning of the model.

Introduces Jax as a powerful framework for large-scale training and model development, particularly used at Google for building models like Gemini. Discusses the efficiency and scalability benefits of Jax in developing models that are significantly larger than lightweight models like Gemma.

Explores the benefits of using Jax, including high performance, compiler-oriented design, and efficient operations on GPUs and TPUs. Highlights the modularity and ecosystem of Jax, enabling easy development of machine learning components and models.

Examines the efficiency and scalability aspects of Jax, focusing on its JIT transformation for performance optimization, parallelism for scaling to thousands of accelerators, and techniques like distributed data parallelism and fully-sharded data parallelism for efficient model training.