$ the-wire · showcase
JAX fixes silent wrong-index results and gcd non-termination
By RepoJournal · Filed · About Google · Composed from the cited sources · methodology
JAX shipped three correctness fixes today that matter more than any feature here: a wrong-element read on long axes indexed with int8, a hung while_loop on minimum signed integers, and inaccurate tail sampling in truncated_normal.
Upcast narrow signed integer indices before negative-index wrapping. google/jax
Indexing an axis longer than 127 elements with a narrow signed index like jnp.int8(-1) used to wrap the axis length into the narrow dtype and overflow, selecting the wrong element. Narrow signed indices are now promoted to a dtype wide enough to hold index + size, so negative indexing on large axes returns the element you asked for instead of a silently wrong one.
Support 128-lane A scale layout in tcgen05 on sm_107f. google/jax
On sm_107f, tcgen05.mma now supports a 128-lane Scale Factor A layout in TMEM via bit 26 of the scaled instruction descriptor; with row r in lane r, each warp writes only its own rows instead of every warp holding and storing all 128. That removes the redundant register and tcgen05.st work the 32-lane layout forced on kernel writers targeting that architecture.
Fix non-termination in `jnp.gcd` and `jnp.lcm` for minimum signed integers. google/jax
by Jake VanderPlas
For a signed dtype, abs(iinfo(dtype).min) overflows back to the negative minimum, so jnp.gcd pushed negative remainders into its Euclidean while_loop and the post-modulo x1 < x2 swap fired forever. jnp.lcm inherits the fix since it delegates to gcd.
Add `mode="high"` to `jax.random.truncated_normal` to fix tail sampling accuracy. google/jax
jax.random.truncated_normal gains mode="high", replacing the erf-based inverse-CDF path that drew uniform samples on a fixed-point grid of resolution 2^(-nmant) and rescaled by floats * (b - a) + a. The accuracy losses were concentrated in the tails, so draws near the truncation bounds were the ones worth not trusting.
[jax-buildcop] Disable `mutable_array_test` on CPU under TSan google/jax
The long tail: the google-auth 2.62.2 release prevents race conditions during mTLS cert rotation on 401 responses and warns on an omitted audience in verify_firebase_token, while spanner-sqlalchemy and spanner-django both compile WHERE 1=1 into unqualified DELETE and UPDATE themselves rather than relying on the DB-API driver to rewrite SQL strings. google-cloud-designcenter 0.1.0 is a new package.
Action items
- → Upgrade google-auth to 2.62.2 to pick up the mTLS cert rotation race fix googleapis/google-cloud-python [plan]
- → Upgrade jax to pick up the narrow signed index fix if you index large axes with int8 or int16 google/jax [plan]