Keras 3 Distributed Training: Scaling Models with JAX using DataParallel, and ModelParallel

Опубликовано 4 марта 2026, 17:00

Training large deep learning models doesn't have to be complex. In this video, Yufeng Guo walks you through the Keras 3 Distribution API, showing you how it leverages JAX for efficient data and model parallelism. Whether you're scaling across multiple GPUs or a cluster of TPUs, Keras 3 has you covered.

Resources:
Distributed training with Keras 3 → goo.gle/4u8nGo9
Multi-device distribution → goo.gle/46CFOMX
LayoutMap API → goo.gle/3NfJXjd
Gemma get_layout_map → goo.gle/4smwNzM

Chapters:
0:00 - Intro
0:17 - The Keras 3 Distribution API
0:51 - The Global Programming Model (SPMD Expansion)
1:26 - Using the JAX Backend for Scalability
1:55 - Creating a Device Mesh & Tensor Layout
2:46 - Implementing Data Parallelism
3:45 - Understanding Model Parallelism
4:27 - Sharding with LayoutMap
5:43 - Tuning Your Device Mesh for Performance
6:14 - Conclusion & Next Steps

Subscribe to Google for Developers → goo.gle/developers

Speaker: Yufeng Guo
Products Mentioned: Google AI