Keras 3 backend transitions and gradient computation parity

Cross-Backend Optimization Consistency

Keras 3 enables a unified API that allows models to be executed across JAX, PyTorch, and TensorFlow by utilizing a backend-agnostic operations layer via keras.ops. While this architecture ensures high-level code portability, the underlying execution of gradient computations and the management of optimizer states rely on backend-specific primitives.

When migrating a model's weights from one backend to another, there is uncertainty regarding the mathematical parity of the resulting gradients. Because each backend employs different automatic differentiation engines and XLA compilation strategies, the precision and behavior of the optimization state may diverge during the transition.

• Does Keras 3 guarantee identical gradient values across different backends for the same weight set and input?
• How is the optimizer state handled when switching backends to ensure training continuity?