Latest Results
🐛 fix(unxt): support jax 0.12 (`erfc_p` accuracy, `polynomial_p`) (#1005)
* 🐛 fix(unxt): support jax 0.12 (`erfc_p` accuracy, `polynomial_p`)
The weekly newest-deps ecosystem job (jax 0.12.0, quax 0.4.6) failed twice:
- `lax.erfc` now binds `erfc_p` with an `accuracy` param, which the
`erfc_p` rule's `(x)` signature rejected. It now takes `/, **kw` and
passes them through, like `exp_p` / `log_p`.
- `jnp.polyval` (and so `jax.scipy.special.spence`) now lowers to the new
`polynomial_p` primitive instead of `mul` / `add`, so quax materialised
the quantity. Add `polynomial_p` rules (guarded on JAX 0.12.0): the result
takes the constant term's unit, and coefficient `k` is converted to that
unit over `x`'s unit to the `k`.
Require quax>=0.4.6, the first release that supports jax 0.12 (its own
floor is still jax>=0.7.2), and re-lock only quax (0.4.4 -> 0.4.6).
* ✅ test(unxt): doctests tolerate jax 0.12's more accurate special functions
jax 0.12 moved `igamma`, `igammac`, `gammaln` and `betainc` closer to their
true values (e.g. `betainc(2, 3, 0.5)` is now exactly 0.6875), which broke
the last float32 digits of 12 doctests on the newest-deps job. Truncate the
expected values to the digits both old and new jax agree on.
* 📝 docs(unxt): reword the polynomial_p uncovered-case comment
* ✅ test(unxt): xfail `jnp.hypot` on dimensionful quantities on jax 0.12
jax 0.12 rewrote `jnp.hypot` to rescale by powers of two, guarded by
comparisons of `max(|x1|, |x2|)` against raw constants, which unxt refuses
for dimensionful quantities. Mark `test_hypot` as a strict xfail on
JAX >= 0.12 until it is fixed in quax, and drop the two `select_n` doctest
examples that only used `jnp.hypot` to illustrate its old lowering.
* 🐛 fix(unxt): polynomial dtype promotion and static result; add one_minus_square_p
From an independent audit of this PR:
- `_polynomial` cast unit-converted (float) coefficients back to an int `x`'s
dtype, so e.g. `polyval(Q([1, 1], ""), Q(50, "%"))` gave 1 instead of 1.5.
Dtype-promote the stripped values first, as `add_p` does.
- A `StaticQuantity` constant term made the result a `StaticQuantity`, which
raises under `jit`. The result is computed, so return a `Quantity`.
- JAX 0.11.2 added `one_minus_square_p`, used by the JVPs of `tanh`, `asin`,
`acos` and `atanh`; differentiating them inside `quax.quaxify` refused to
materialise the quantity. Register a dimensionless rule.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> Latest Branches
0%
-11%
nstarman:claude/unit-systems-plum-d66f71 -19%
nstarman:quantity-parametric-rename © 2026 CodSpeed Technology