It turns out that the support for C++17 is partial in 10.12, and in particular absl::optional and std::optional are not the same thing under 10.12. Increment to 10.14 which is the lowest version that builds successfully with absl::optional == std::optional.
See: 89cdaed655/absl/base/config.h (L528)
Strictly speaking, we could allow 10.13, but not without updating ABSL in the TF repository to incorporate c86347d4ce which fixes the version detection test to permit 10.13 as well.
Adds a non-standard `method` argument to `jnp.linalg.slogdet` to select between the current LU decomposition based implementation (like NumPy) and the QR decomposition implementation.
QR decomposition is more amenable to a high performance batched implementation particularly on TPU hardware because it does not need row pivoting. The same may be true on other hardware also, and having the option is nice either way!
PiperOrigin-RevId: 449271317
An upcoming change to add a more scalable QDWH-based TPU symmetric eigendecomposition requires that we can obtain the TPU eigenvalues unsorted. The option already exists in XLA, so we simply need to plumb it through to the lax primitive.
PiperOrigin-RevId: 448047584
Previously, `jnp.take` defaulted to clamping out-of-bounds indices into range. Now, `jnp.take` returns invalid values (e.g., NaN) for out-of-bounds indices. This change attempts to prevent latent bugs caused by inadvertent out-of-bounds indices.
The previous behavior can be approximated using the "clip" or "wrap" fill modes.
PiperOrigin-RevId: 445130143
This is a reasonably safe change, because it has no effect on the forward pass of a computation: the default behavior (PROMISE_IN_BOUNDS) also drops out-of-bounds scatters.
This change does however affect the transpose (gradient) of a scatter with out-of-bounds indices: the gradient of a PROMISE_IN_BOUNDS scatter is a PROMISE_IN_BOUNDS gather, and a PROMISE_IN_BOUNDS gather clips out-of-bounds indices into range. This is not mathematically correct: a dropped scatter index does not contribute to the primal output, and so its transpose should yield a zero cotangent.
After this change, the gradient of a default scatter is a gather with a fill value of 0: i.e., the indices that were dropped do not make gradient contributions, which is mathematically correct.
Separately, I am working towards switching out-of-bounds gather() operations to also have FILL_OR_DROP semantics, although that change is more disruptive because a number of users have out-of-bounds indices in their gather()s.
Issues: https://github.com/google/jax/issues/278https://github.com/google/jax/issues/9839
PiperOrigin-RevId: 444935241
This change align the behavior of `ravel_multi_index`, `split` and `indices` to their `numpy` counterparts.
Also ensure size argument of `nonzero` should be integer.
The changes with `*space` are only simplification
Previously, out-of-bounds indices were clipped into range, but that behavior is error prone. We would rather fail in a more visible way when out-of-bounds indices are used. Future changes will migrate other JAX indexing operations to have the same semantics.
PiperOrigin-RevId: 443390170
This allows users of jnp.take_along_axis to override the out-of-bounds indexing behavior.
Default to "clip", which for the forward computation is identical to the current behavior. In a future change, we will change this to "fill".
--
391dea76bc8fe264cf26ec93d42147f87847894d by Peter Hawkins <phawkins@google.com>:
Update version numbers after jax/jaxlib 0.3.7 release.
COPYBARA_INTEGRATE_REVIEW=https://github.com/google/jax/pull/10324 from hawkinsp:jaxlib 391dea76bc8fe264cf26ec93d42147f87847894d
PiperOrigin-RevId: 442311051
