torchax

Allows existing PyTorch code to run on TPUs by simply enabling torchax and setting device='jax', as shown in the running example where a standard nn.Module executes on JAX hardware.

Supports bidirectional calls between PyTorch and JAX, enabling use of JAX libraries like optax and jax.grad with PyTorch models for hybrid ML pipelines.

Provides JittableModule helper to apply jax.jit to PyTorch models, compiling them for better performance, as demonstrated in the functional model example.

Offers torchax.save_checkpoint and load_checkpoint to handle states mixing PyTorch weights and JAX optimizer states, simplifying hybrid workflow persistence.