torch2jax

Run PyTorch in JAX. 🤝

Mix-and-match PyTorch and JAX code with seamless, end-to-end autodiff, use JAX classics like jit, grad, and vmap on PyTorch code, and run PyTorch models on TPUs.

torch2jax uses abstract interpretation (aka tracing) to move JAX values through PyTorch code. As a result, you get a JAX-native computation graph that follows exactly your PyTorch code, down to the last epsilon.

from torch2jax import j2t, t2j

vit = torchvision.models.vit_b_16().eval()
batch = torch.randn(1, 3, 224, 224)
vit(batch)
# => [-5.3352e-01, ..., 2.0390e-01]

jax_vit = t2j(vit)
jax_batch = t2j(batch)
params = {k: t2j(v) for k, v in vit.named_parameters()}
jit(jax_vit)(jax_batch, state_dict=params)
# => [-5.3125e-01, ..., 2.0735e-01]

torch2jax even works with in-place PyTorch operations:

def f(x):
    x.add_(1)
    x.mul_(2)
    return x

f(torch.tensor([3]))                # => torch.Tensor([8])

jax_f = t2j(f)
jax_f(jnp.array([3]))               # => jnp.array([8])
vmap(jax_f)(jnp.array([1, 2, 3]))   # => jnp.array([[4], [6], [8]])
grad(jax_f)(jnp.array([2.0]))       # => jnp.array([2.0])

torch2jax offers a simple API with two functions:

1. j2t: Convert a JAX jax.numpy.ndarray to a torch.Tensor.
2. t2j: Convert a PyTorch function, torch.nn.Module, or torch.Tensor to their JAX equivalent.

Internally, the core of torch2jax is Torchish, a class that mimics torch.Tensor via __torch_function__. A Torchish object is backed by a JAX jax.numpy.ndarray, and proxies PyTorch operations onto the underlying jax.numpy.ndarray. As a result, you get a JAX-native computation graph that exactly follows your PyTorch code.

Installation

PyPI

pip install torch2jax

Nix flake

torch2jax is available as a Nix flake.

$ nix shell github:samuela/torch2jax
(shell) $ python -c "from torch2jax import j2t, t2j"

FAQ

Help! I've encountered a PyTorch operation that isn't implemented yet.

torch2jax is an implementation of the PyTorch standard library written in JAX. If you come across an operation that isn't implemented yet, please file an issue and/or PR!

Adding new PyTorch operations is straightforward. Check the source for functions decorated with @implements to get started.

My PyTorch model includes dropout or some other random operation. How does this work with torch2jax?

Pass a jax.random.PRNGKey to the converted function:

t2j(lambda: torch.randn(3))(rng=jax.random.PRNGKey(123))
# => [-0.56996626, -0.6440589 ,  0.28660855]

t2j(lambda: torch.randn(3))(rng=jax.random.PRNGKey(456))
# => [-1.3227656, -1.4896724, -2.5057693]

After conversion, random state will be handled entirely in JAX. torch.manual_seed and its ilk will have no effect on the converted function.

If you only care about running a model and not training it, you can call .eval() on it to avoid the randomness issue altogether, at least for most common random operations like dropout:

rn18 = torchvision.models.resnet18().eval()
t2j(rn18)(t2j(torch.randn(1, 3, 224, 224)))     # Look ma, no `rng` kwarg!

Note

Non-deterministic behavior is, well, non-deterministic. You will not see the same results with the same random seed when switching between PyTorch and JAX. However, the sampling process will be equivalent.

My PyTorch model includes batch norm or some other torch.nn.Module that mutates buffers. How does this work with torch2jax?

Some PyTorch modules like torch.nn.BatchNorm1d mutate internal state in the form of buffers.

torch2jax supports this with the optional return_state_dict argument:

rn18 = torchvision.models.resnet18()
batch = torch.randn(1, 3, 224, 224)

before_state_dict = {k: t2j(v) for k, v in rn18.state_dict().items()}
out, after_state_dict = t2j(rn18)(t2j(batch), state_dict=before_state_dict, return_state_dict=True)

As with randomness, if you only care about running a model and not training it, you can call .eval() on it to avoid buffer issues altogether in most cases.

Also, don't forget to avoid taking gradients w.r.t. buffers. For example,

rn18 = torchvision.models.resnet18().eval()
loss = lambda x: torch.sum(x ** 2)

batch = torch.randn(1, 3, 224, 224)
loss(rn18(batch)).backward()

parameters = {k: t2j(v) for k, v in rn18.named_parameters()}
buffers = {k: t2j(v) for k, v in rn18.named_buffers()}

jax_rn18 = t2j(rn18)
grad(lambda params, x: loss(jax_rn18(x, state_dict={**params, **buffers})))(parameters, t2j(batch))

I'm seeing slightly different numerical results between PyTorch and JAX. Is it a bug?

Floating point arithmetic is hard. There are a number of sources of divergence preventing bit-for-bit equivalence:

1. torch2jax guarantees equivalence with PyTorch standard library functions in the mathematical sense, but not necessarily in their operational execution. This can lead to slight differences in results. For example, the multi-head attention implementations calculate the same mathematical function, but may vary in execution details such as the order of operations, the use of fused kernels, and so forth.
2. The JAX/XLA and PyTorch compilers apply different optimizations and should be expected to rewrite computation graphs in exciting and unpredictable ways, potentially invoking different CUDA kernels.
3. CUDA kernels can be non-deterministic, for example as a result of floating point addition being non-associative.

Also bear in mind that floating point errors compound, so larger models will experience increased divergence.

What about going the other way around? Running JAX code in PyTorch?

Check out jax2torch.

Contributing

PyTorch has a non-trivial API surface to cover. Contributions are welcome!

Run the test suite with pytest running in nix develop. Format the codebase with ruff check --fix . && ruff format .. Build the package with nix build.

CI is handled by GitHub Actions. When modifying the CI configuration, it can be handy to test locally before pushing. This can be achieved with act. Run act within nix develop to run the CI locally.