torchax
torchax is a PyTorch frontend for JAX. It gives JAX the ability to author JAX programs using familiar PyTorch syntax. It also provides JAX-Pytorch interoperabil
What is torchax?
torchax is a PyTorch frontend for JAX. It gives JAX the ability to author JAX programs using familiar PyTorch syntax. It also provides JAX-Pytorch interoperability, meaning, one can mix JAX & Pytorch syntax together when authoring ML programs, and run it in every hardware JAX can run.
Overview of torchax
What it is
torchax is a PyTorch frontend for JAX. It gives JAX the ability to author JAX programs using familiar PyTorch syntax. It also provides JAX-PyTorch interoperability, meaning one can mix JAX and PyTorch syntax together when authoring ML programs and run it on every hardware JAX can run.
Key features
- Run PyTorch code on TPUs with minimal code changes.
- Call JAX functions from PyTorch, passing in
jax.Arrays. - Call PyTorch functions from JAX, passing in
torch.Tensors. - Use JAX features like
jax.grad,optax, andGSPMDto train PyTorch models. - Use a PyTorch model as a feature extractor with a JAX model.
- Save and load training state using
torchax.save_checkpointandtorchax.load_checkpoint.
Use cases
- Running PyTorch programs on Google Cloud TPUs.
- Mixing JAX and PyTorch code in the same ML program.
- Using JAX compilation (
jax.jit) to accelerate PyTorch models. - Training PyTorch models with JAX optimizers like
optax. - Using a PyTorch model as a feature extractor within a JAX pipeline.
Getting started
First, install the CPU version of PyTorch: bash # On Linux pip install torch --index-url https://download.pytorch.org/whl/cpu # On Mac pip install torch Next, install JAX for your desired accelerator: bash # On Google Cloud TPU pip install -U jax[tpu] # On GPU machines pip install -U jax[cuda12] # On Linux CPU machines or Macs pip install -U jax Finally, install torchax: bash pip install torchax # Or from source pip install git+https://github.com/google/torchax
License and pricing
The license is Apache-2.0. The project is open source.
Based on the project's public README and repository metadata.