Here's an overview of the features we intend to work on in the near future.

Core Keras

Performance and Optimization

• Introduce comprehensive support for model quantization, including:
  • Post-training quantization techniques likeGPTQ andAWQ.
  • Quantization-Aware Training (QAT) int8 support.

Scale and Distribution

• Distributed Training
  • Comprehensive guides formulti-host TPU andmulti-host GPU training.
  • Official performance benchmarks
  • ABackup and Restore callback to handle preemptions gracefully during long training runs.

Integrations and Ecosystem

• Add support for exporting models to theODML LiteRT format, simplifying deployment on edge and mobile devices.
• IntegrateQwix, a new JAX-based library for quantization.
• [Contributions Welcome] IntegratePyGrain for creating efficient, large-scale data loading and preprocessing pipelines.

Guides and Tutorials

• Deployment Guides: End-to-end tutorials on deploying Keras models toVertex AI, and on-device viaLiteRT.
• Guide on efficient inference using KerasHub models withvLLM.
• AI Agents and RAG: Advanced examples of buildingAI agents with function calling and creatingRetrieval-Augmented Generation (RAG) pipelines.
• Training Techniques: Guides onmodel distillation, handling training preemptions on TPUs, and best practices forimage augmentation (e.g., CutMix and MixUp).
• Others: Orbax checkpointing, FLUX model guide/example, etc.

KerasHub

See the roadmaphere.

KerasRS

See the roadmaphere.