jax-config

To avoid recompilation when passing dataclass configs into jitted JAX functions, register the dataclass as a JAX pytree and wrap dynamic float fields as 0-d arrays. This separates static fields for tracing from dynamic fields for tunable hyperparameters.

Integrating nested dataclass configs with Tyro CLI requires using field(default_factory=...) for nested structures and exposing a tyro.cli entry point. This setup provides clean command-line options for your JAX training loop while maintaining pytree registration.

JAX recompiles your model because unregistered dataclass fields are treated as static, invalidating the trace. Registering the config as a pytree and wrapping dynamic floats as JAX arrays allows hyperparameter changes without triggering recompilation.

Yes, Equinox works seamlessly with pytree-registered dataclass configs. You can combine these registered configs directly with Equinox modules in large ML pipelines to maintain stable JAX tracing across complex model architectures.

You need to register a dataclass as a JAX pytree when building models or training loops that pass configuration dataclasses into jitted functions. This registration enables deterministic behavior and ensures dynamic fields can be traced correctly without recompilation.