140 wires and counting

$ follow Google

Keep up with Google in about 3 minutes: what actually shipped — the commits, pull requests, releases, and security advisories that matter.

or

fair warning: these emails are deeply technical. diffs, version numbers, CVEs, benchmark deltas. if that's not your idea of a good read, this isn't your newsletter.

Folds into your digest — weekly by default, monthly if you prefer. Unsubscribe in one click.

$ status

wire 2026-06-28
stories 10

© 2026 RepoJournal Home Showcase How it works Privacy

$ the-wire · showcase

JAX TIGHTENS GRADIENT CHECKPOINTING, FIXES SCIPY PARITY BUGS

By RepoJournal · Filed · About Google · Composed from the cited sources · methodology

JAX's custom_vjp3 primitive now supports checkpoint_name in forward passes, unlocking memory-efficient gradient computation for complex AD workflows.

The big move is giving custom_vjp3 a remat rule that applies JAX's remat_transform to primal functions [1]. This means you can now checkpoint intermediate values in custom VJP forward passes without breaking the AD chain, which matters for training large models where memory is the bottleneck. Alongside this, JAX made primal_left_tangent_right a hijax primitive so it lowers away in lojax, letting the compiler dead-code-eliminate unused gradients [1].

The team also shipped fixes to scipy.special parity bugs that have been lingering [2] [3]. lax.special.gammainc and lax.special.gammaincc now return 0.0 and 1.0 respectively when a=infinity instead of NaN, matching scipy's behavior. Same fix went into igamma and igammac for a=inf with finite x. These are edge cases but they matter if you're doing numerical computation or porting scipy code to JAX.

Test coverage for remat3 got a refresh to close gaps [4], so expect fewer surprises in production rematerialization rules. Custom AD primitives and scipy compatibility are both stability plays, not headline features, but both are the kind of fixes that prevent silent correctness bugs.

Quick answers

What shipped in Google on June 28, 2026?
JAX's custom_vjp3 primitive now supports checkpoint_name in forward passes, unlocking memory-efficient gradient computation for complex AD workflows. In total, 6 commits and 4 pull requests landed.
Who contributed to Google on June 28, 2026?
2 developers shipped this update, including mattjj and copybara-service[bot].
What were the notable Google updates?
support checkpoint_name in custom_vjp fwd if custom_vjp3=1, Fix `lax.special.gammainc` and `lax.special.gammaincc` for a = infinity, and Fix `lax.special.igamma` and `lax.special.igammac` for `a = inf` & `x` finite.