Skip to main content

Cross-backend parity

omnibias guarantees that a given (activation, order) pair is float64-ULP- equal across PyTorch, JAX, and Keras 3. This page explains why, and how to check it yourself.

Why it holds​

There is exactly one definition of the polynomial coefficients, in pure-Python omnibias-core. Each backend imports it and applies it with its own tensor type. There is no per-backend reimplementation that could drift.

Verify it yourself​

import numpy as np
import torch, jax.numpy as jnp
from omnibias.torch import get_activation as torch_act
from omnibias.jax import get_activation as jax_act

z = np.linspace(-3, 3, 101).astype(np.float64)

for n in range(6):
    t = torch_act("tanh").fastpath(torch.tensor(z, dtype=torch.float64), n).numpy()
    j = np.asarray(jax_act("tanh").fastpath(jnp.asarray(z), n))
    assert np.max(np.abs(t - j)) <= 1e-15, f"order {n} mismatch"
print("bit-parity OK across torch/jax")
Use float64 for parity checks

Parity is a float64 guarantee. In float32, both backends round to 32-bit but intermediate orderings can differ at the last bit. Always check parity in float64.

What the release suite checks​

Each release runs a dedicated cross-backend suite asserting ULP-equality on every (activation, order) pair. The headline: 73/73 and 251/251 cross-backend tests pass per release.

If you see a mismatch​

  1. Confirm dtype. Are both sides float64? (See the tip above.)
  2. Confirm the same activation name and that neither side registered a custom override.
  3. Confirm order support. A NotImplementedError on one backend means the order is not implemented there — that is a support gap, not a parity bug.
  4. File it. A genuine float64 mismatch is a bug; report it with a minimal reproduction.

Why this enables certified numerics​

Backend-portable certified numerics is only possible because the bits agree: a certificate produced from the JAX path is meaningful for the PyTorch path, because both descend from the same coefficients. See the Certified register.

See also​

Last updated on