News

JAX 0.12.1 moves selected Pallas APIs to `jax.pallas`

JAX 0.12.1 introduces a public import path for selected Pallas APIs, adds gradient logging and new numerical operations, and updates accelerator support.

D
Nov 19, 2026 · 2 min read

JAX released version 0.12.1 on November 18, moving selected Pallas APIs into the public jax.pallas namespace and adding developer-facing features for numerical computing, differentiation and accelerators. Developers can now import core Pallas interfaces outside the experimental namespace, while two paths remain experimental.

The new public imports are jax.pallas.tpu, jax.pallas.tpu_sc, jax.pallas.mosaic_gpu and jax.pallas.fuser. Existing jax.experimental.pallas imports remain valid in 0.12.1, but triton and ops stay under jax.experimental.pallas. JAX’s compatibility policy says experimental APIs may change between releases without warning, although the project follows a three-month deprecation period in most cases.

Pallas, as JAX describes it, is an extension for writing custom GPU and TPU kernels when XLA’s automatic optimizations are insufficient. Developers can use it to write fusions, control memory movement and pipelining, and access low-level hardware features.

For differentiation, the release adds a keyword-only with_logs option to jax.grad and jax.value_and_grad. This allows backward-pass logs to be returned with gradients without a direct jax.vjp call. It also introduces jax.avals, along with Mesh.axes and AbstractMesh.axes, as shorthand for constructing ShapedArray abstract types with optional sharding annotations. On Python 3.14 and later, jax.debug.print accepts PEP 750 template strings.

The 0.12.1 changelog also lists a broad set of numerical additions. jax.lax.fma provides a single-rounding fused multiply-add where supported, while jax.lax.fmuladd falls back when fusion is unavailable. Floating-point evaluation in jax.lax.polynomial now uses fmuladd. New sinpi, cospi and tanpi operations work directly on pi-scaled inputs, and inverse counterparts return results divided by pi. JAX says the direct operations improve accuracy, including exact results at integers and half-integers for the forward functions.

Other additions include exp2m1 and log2p1 for calculations near zero, inverse regularized incomplete gamma functions, integer-order ordinary and modified Bessel functions, and ndtri_exp for inverse-normal calculations in log-probability space. New modes for jax.random.normal and jax.random.truncated_normal use more random bits to extend tail reach or improve tail accuracy. JAX’s compatibility policy says exact numerical values are not guaranteed to remain identical across releases, accelerator platforms or execution with and without jax.jit.

Explicit sharding support now covers optional out_sharding parameters for scatter, gather, take, reduction and quantile-family operations. top_k, approx_max_k and approx_min_k gain sharded reduction-axis support. On AMD hardware, new ROCm pip extras can install TheRock ROCm wheels from AMD’s package index without a system ROCm installation; jax[rocm7-local] remains available for pre-installed ROCm.

Version 0.12.1 also changes several defaults and compatibility boundaries. Simplified jaxpr constants are enabled by default, passing captured non-scalar constant arrays into compiled executables instead of embedding them inline; JAX_USE_SIMPLIFIED_JAXPR_CONSTANTS=False can temporarily restore the old behavior. JAX now always uses the Shardy partitioner after removing legacy GSPMD support. The experimental serialized-executable format advances to version 3, and deserialize_and_load in 0.12.1 cannot load executables saved with JAX 0.12.0 or earlier.

More news