Change log

Best viewed here.

jax 0.2.22 (Unreleased)

jax 0.2.21 (Sept 23, 2021)

  • GitHub commits.

  • Breaking Changes

    • jax.api has been removed. Functions that were available as jax.api.* were aliases for functions in jax.*; please use the functions in jax.* instead.

    • jax.partial, jax.lax.partial, and jax.util.partial were accidental exports that have now been removed. Use functools.partial from the Python standard library instead.

    • Boolean scalar indices now raise a TypeError; previously this silently returned wrong results (#7925).

    • Many more jax.numpy functions now require array-like inputs, and will error if passed a list (#7747 #7802 #7907). See #7737 for a discussion of the rationale behind this change.

    • When inside a transformation such as jax.jit, jax.numpy.array always stages the array it produces into the traced computation. Previously jax.numpy.array would sometimes produce a on-device array, even under a jax.jit decorator. This change may break code that used JAX arrays to perform shape or index computations that must be known statically; the workaround is to perform such computations using classic NumPy arrays instead.

  • New features:

jax 0.2.20 (Sept 2, 2021)

  • GitHub commits.

  • Breaking Changes

    • jnp.poly* functions now require array-like inputs (#7732)

    • jnp.unique and other set-like operations now require array-like inputs (#7662)

jaxlib 0.1.71 (Sep 1, 2021)

  • Breaking changes:

    • Support for CUDA 11.0 and CUDA 10.1 has been dropped. Jaxlib now supports CUDA 10.2 and CUDA 11.1+.

jax 0.2.19 (Aug 12, 2021)

  • GitHub commits.

  • Breaking changes:

    • Support for NumPy 1.17 has been dropped, per the deprecation policy. Please upgrade to a supported NumPy version.

    • The jit decorator has been added around the implementation of a number of operators on JAX arrays. This speeds up dispatch times for common operators such as +.

      This change should largely be transparent to most users. However, there is one known behavioral change, which is that large integer constants may now produce an error when passed directly to a JAX operator (e.g., x + 2**40). The workaround is to cast the constant to an explicit type (e.g., np.float64(2**40)).

  • New features:

    • Improved the support for shape polymorphism in jax2tf for operations that need to use a dimension size in array computation, e.g., jnp.mean. (#7317)

  • Bug fixes:

    • Some leaked trace errors from the previous release (#7613)

jaxlib 0.1.70 (Aug 9, 2021)

  • Breaking changes:

    • Support for Python 3.6 has been dropped, per the deprecation policy. Please upgrade to a supported Python version.

    • Support for NumPy 1.17 has been dropped, per the deprecation policy. Please upgrade to a supported NumPy version.

    • The host_callback mechanism now uses one thread per local device for making the calls to the Python callbacks. Previously there was a single thread for all devices. This means that the callbacks may now be called interleaved. The callbacks corresponding to one device will still be called in sequence.

jax 0.2.18 (July 21 2021)

  • GitHub commits.

  • Breaking changes:

    • Support for Python 3.6 has been dropped, per the deprecation policy. Please upgrade to a supported Python version.

    • The minimum jaxlib version is now 0.1.69.

    • The backend argument to jax.dlpack.from_dlpack() has been removed.

  • New features:

    • Added a polar decomposition (jax.scipy.linalg.polar()).

  • Bug fixes:

    • Tightened the checks for lax.argmin and lax.argmax to ensure they are not used with an invalid axis value, or with an empty reduction dimension. (#7196)

jaxlib 0.1.69 (July 9 2021)

  • Fix bugs in TFRT CPU backend that results in incorrect results.

jax 0.2.17 (July 9 2021)

jax 0.2.16 (June 23 2021)

jax 0.2.15 (June 23 2021)

  • GitHub commits.

  • New features:

    • #7042 Turned on TFRT CPU backend with significant dispatch performance improvements on CPU.

    • The jax2tf.convert() supports inequalities and min/max for booleans (#6956).

    • New SciPy function jax.scipy.special.lpmn_values().

  • Breaking changes:

  • Bug fixes:

    • Fixed bug that prevented round-tripping from JAX to TF and back: jax2tf.call_tf(jax2tf.convert) (#6947).

jaxlib 0.1.68 (June 23 2021)

  • Bug fixes:

    • Fixed bug in TFRT CPU backend that gets nans when transfer TPU buffer to CPU.

jax 0.2.14 (June 10 2021)

  • GitHub commits.

  • New features:

    • The jax2tf.convert() now has support for pjit and sharded_jit.

    • A new configuration option JAX_TRACEBACK_FILTERING controls how JAX filters tracebacks.

    • A new traceback filtering mode using __tracebackhide__ is now enabled by default in sufficiently recent versions of IPython.

    • The jax2tf.convert() supports shape polymorphism even when the unknown dimensions are used in arithmetic operations, e.g., jnp.reshape(-1) (#6827).

    • The jax2tf.convert() generates custom attributes with location information in TF ops. The code that XLA generates after jax2tf has the same location information as JAX/XLA.

    • New SciPy function jax.scipy.special.lpmn().

  • Bug fixes:

    • The jax2tf.convert() now ensures that it uses the same typing rules for Python scalars and for choosing 32-bit vs. 64-bit computations as JAX (#6883).

    • The jax2tf.convert() now scopes the enable_xla conversion parameter properly to apply only during the just-in-time conversion (#6720).

    • The jax2tf.convert() now converts lax.dot_general using the XlaDot TensorFlow op, for better fidelity w.r.t. JAX numerical precision (#6717).

    • The jax2tf.convert() now has support for inequality comparisons and min/max for complex numbers (#6892).

jaxlib 0.1.67 (May 17 2021)

jaxlib 0.1.66 (May 11 2021)

  • New features:

    • CUDA 11.1 wheels are now supported on all CUDA 11 versions 11.1 or higher.

      NVidia now promises compatibility between CUDA minor releases starting with CUDA 11.1. This means that JAX can release a single CUDA 11.1 wheel that is compatible with CUDA 11.2 and 11.3.

      There is no longer a separate jaxlib release for CUDA 11.2 (or higher); use the CUDA 11.1 wheel for those versions (cuda111).

    • Jaxlib now bundles libdevice.10.bc in CUDA wheels. There should be no need to point JAX to a CUDA installation to find this file.

    • Added automatic support for static keyword arguments to the jit() implementation.

    • Added support for pretransformation exception traces.

    • Initial support for pruning unused arguments from jit() -transformed computations. Pruning is still a work in progress.

    • Improved the string representation of PyTreeDef objects.

    • Added support for XLA’s variadic ReduceWindow.

  • Bug fixes:

    • Fixed a bug in the remote cloud TPU support when large numbers of arguments are passed to a computation.

    • Fix a bug that meant that JAX garbage collection was not triggered by jit() transformed functions.

jax 0.2.13 (May 3 2021)

  • GitHub commits.

  • New features:

    • When combined with jaxlib 0.1.66, jax.jit() now supports static keyword arguments. A new static_argnames option has been added to specify keyword arguments as static.

    • jax.nonzero() has a new optional size argument that allows it to be used within jit (#6501)

    • jax.numpy.unique() now supports the axis argument (#6532).

    • jax.experimental.host_callback.call() now supports pjit.pjit (#6569).

    • Added jax.scipy.linalg.eigh_tridiagonal() that computes the eigenvalues of a tridiagonal matrix. Only eigenvalues are supported at present.

    • The order of the filtered and unfiltered stack traces in exceptions has been changed. The traceback attached to an exception thrown from JAX-transformed code is now filtered, with an UnfilteredStackTrace exception containing the original trace as the __cause__ of the filtered exception. Filtered stack traces now also work with Python 3.6.

    • If an exception is thrown by code that has been transformed by reverse-mode automatic differentiation, JAX now attempts to attach as a __cause__ of the exception a JaxStackTraceBeforeTransformation object that contains the stack trace that created the original operation in the forward pass. Requires jaxlib 0.1.66.

  • Breaking changes:

    • The following function names have changed. There are still aliases, so this should not break existing code, but the aliases will eventually be removed so please change your code.

    • Similarly, the argument to local_devices() has been renamed from host_id to process_index.

    • Arguments to jax.jit() other than the function are now marked as keyword-only. This change is to prevent accidental breakage when arguments are added to jit.

  • Bug fixes:

    • The jax2tf.convert() now works in presence of gradients for functions with integer inputs (#6360).

    • Fixed assertion failure in jax2tf.call_tf() when used with captured tf.Variable (#6572).

jaxlib 0.1.65 (April 7 2021)

jax 0.2.12 (April 1 2021)

  • GitHub commits.

  • New features

  • Breaking changes:

    • The minimum jaxlib version is now 0.1.64.

    • Some profiler APIs names have been changed. There are still aliases, so this should not break existing code, but the aliases will eventually be removed so please change your code.

    • Omnistaging can no longer be disabled. See omnistaging for more information.

    • Python integers larger than the maximum int64 value will now lead to an overflow in all cases, rather than being silently converted to uint64 in some cases (#6047).

    • Outside X64 mode, Python integers outside the range representable by int32 will now lead to an OverflowError rather than having their value silently truncated.

  • Bug fixes:

    • host_callback now supports empty arrays in arguments and results (#6262).

    • jax.random.randint() clips rather than wraps of out-of-bounds limits, and can now generate integers in the full range of the specified dtype (#5868)

jax 0.2.11 (March 23 2021)

  • GitHub commits.

  • New features:

    • #6112 added context managers: jax.enable_checks, jax.check_tracer_leaks, jax.debug_nans, jax.debug_infs, jax.log_compiles.

    • #6085 added jnp.delete

  • Bug fixes:

    • #6136 generalized jax.flatten_util.ravel_pytree to handle integer dtypes.

    • #6129 fixed a bug with handling some constants like enum.IntEnums

    • #6145 fixed batching issues with incomplete beta functions

    • #6014 fixed H2D transfers during tracing

    • #6165 avoids OverflowErrors when converting some large Python integers to floats

  • Breaking changes:

    • The minimum jaxlib version is now 0.1.62.

jaxlib 0.1.64 (March 18 2021)

jaxlib 0.1.63 (March 17 2021)

jax 0.2.10 (March 5 2021)

  • GitHub commits.

  • New features:

    • jax.scipy.stats.chi2() is now available as a distribution with logpdf and pdf methods.

    • jax.scipy.stats.betabinom() is now available as a distribution with logpmf and pmf methods.

    • Added jax.experimental.jax2tf.call_tf() to call TensorFlow functions from JAX (#5627) and README).

    • Extended the batching rule for lax.pad to support batching of the padding values.

  • Bug fixes:

  • Breaking changes:

    • JAX’s promotion rules were adjusted to make promotion more consistent and invariant to JIT. In particular, binary operations can now result in weakly-typed values when appropriate. The main user-visible effect of the change is that some operations result in outputs of different precision than before; for example the expression jnp.bfloat16(1) + 0.1 * jnp.arange(10) previously returned a float64 array, and now returns a bfloat16 array. JAX’s type promotion behavior is described at Type promotion semantics.

    • jax.numpy.linspace() now computes the floor of integer values, i.e., rounding towards -inf rather than 0. This change was made to match NumPy 1.20.0.

    • jax.numpy.i0() no longer accepts complex numbers. Previously the function computed the absolute value of complex arguments. This change was made to match the semantics of NumPy 1.20.0.

    • Several jax.numpy functions no longer accept tuples or lists in place of array arguments: jax.numpy.pad(), :funcjax.numpy.ravel, jax.numpy.repeat(), jax.numpy.reshape(). In general, jax.numpy functions should be used with scalars or array arguments.

jaxlib 0.1.62 (March 9 2021)

  • New features:

    • jaxlib wheels are now built to require AVX instructions on x86-64 machines by default. If you want to use JAX on a machine that doesn’t support AVX, you can build a jaxlib from source using the --target_cpu_features flag to build.py. --target_cpu_features also replaces --enable_march_native.

jaxlib 0.1.61 (February 12 2021)

jaxlib 0.1.60 (Febuary 3 2021)

  • Bug fixes:

    • Fixed a memory leak when converting CPU DeviceArrays to NumPy arrays. The memory leak was present in jaxlib releases 0.1.58 and 0.1.59.

    • bool, int8, and uint8 are now considered safe to cast to bfloat16 NumPy extension type.

jax 0.2.9 (January 26 2021)

jaxlib 0.1.59 (January 15 2021)

jax 0.2.8 (January 12 2021)

  • GitHub commits.

  • New features:

  • Bug fixes:

    • jax.numpy.arccosh now returns the same branch as numpy.arccosh for complex inputs (#5156)

    • host_callback.id_tap now works for jax.pmap also. There is an optional parameter for id_tap and id_print to request that the device from which the value is tapped be passed as a keyword argument to the tap function (#5182).

  • Breaking changes:

  • New features:

    • New flag for debugging inf, analagous to that for NaN (#5224).

jax 0.2.7 (Dec 4 2020)

  • GitHub commits.

  • New features:

    • Add jax.device_put_replicated

    • Add multi-host support to jax.experimental.sharded_jit

    • Add support for differentiating eigenvalues computed by jax.numpy.linalg.eig

    • Add support for building on Windows platforms

    • Add support for general in_axes and out_axes in jax.pmap

    • Add complex support for jax.numpy.linalg.slogdet

  • Bug fixes:

    • Fix higher-than-second order derivatives of jax.numpy.sinc at zero

    • Fix some hard-to-hit bugs around symbolic zeros in transpose rules

  • Breaking changes:

    • jax.experimental.optix has been deleted, in favor of the standalone optax Python package.

    • indexing of JAX arrays with non-tuple sequences now raises a TypeError. This type of indexing has been deprecated in Numpy since v1.16, and in JAX since v0.2.4. See #4564.

jax 0.2.6 (Nov 18 2020)

  • GitHub commits.

  • New Features:

    • Add support for shape-polymorphic tracing for the jax.experimental.jax2tf converter. See README.md.

  • Breaking change cleanup

    • Raise an error on non-hashable static arguments for jax.jit and xla_computation. See cb48f42.

    • Improve consistency of type promotion behavior (#4744):

      • Adding a complex Python scalar to a JAX floating point number respects the precision of the JAX float. For example, jnp.float32(1) + 1j now returns complex64, where previously it returned complex128.

      • Results of type promotion with 3 or more terms involving uint64, a signed int, and a third type are now independent of the order of arguments. For example: jnp.result_type(jnp.uint64, jnp.int64, jnp.float16) and jnp.result_type(jnp.float16, jnp.uint64, jnp.int64) both return float16, where previously the first returned float64 and the second returned float16.

    • The contents of the (undocumented) jax.lax_linalg linear algebra module are now exposed publicly as jax.lax.linalg.

    • jax.random.PRNGKey now produces the same results in and out of JIT compilation (#4877). This required changing the result for a given seed in a few particular cases:

      • With jax_enable_x64=False, negative seeds passed as Python integers now return a different result outside JIT mode. For example, jax.random.PRNGKey(-1) previously returned [4294967295, 4294967295], and now returns [0, 4294967295]. This matches the behavior in JIT.

      • Seeds outside the range representable by int64 outside JIT now result in an OverflowError rather than a TypeError. This matches the behavior in JIT.

      To recover the keys returned previously for negative integers with jax_enable_x64=False outside JIT, you can use:

      key = random.PRNGKey(-1).at[0].set(0xFFFFFFFF)

    • DeviceArray now raises RuntimeError instead of ValueError when trying to access its value while it has been deleted.

jaxlib 0.1.58 (January 12ish 2021)

  • Fixed a bug that meant JAX sometimes return platform-specific types (e.g., np.cint) instead of standard types (e.g., np.int32). (#4903)

  • Fixed a crash when constant-folding certain int16 operations. (#4971)

  • Added an is_leaf predicate to pytree.flatten().

jaxlib 0.1.57 (November 12 2020)

  • Fixed manylinux2010 compliance issues in GPU wheels.

  • Switched the CPU FFT implementation from Eigen to PocketFFT.

  • Fixed a bug where the hash of bfloat16 values was not correctly initialized and could change (#4651).

  • Add support for retaining ownership when passing arrays to DLPack (#4636).

  • Fixed a bug for batched triangular solves with sizes greater than 128 but not a multiple of 128.

  • Fixed a bug when performing concurrent FFTs on multiple GPUs (#3518).

  • Fixed a bug in profiler where tools are missing (#4427).

  • Dropped support for CUDA 10.0.

jax 0.2.5 (October 27 2020)

jax 0.2.4 (October 19 2020)

  • GitHub commits.

  • Improvements:

    • Add support for remat to jax.experimental.host_callback. See #4608.

  • Deprecations

    • Indexing with non-tuple sequences is now deprecated, following a similar deprecation in Numpy. In a future release, this will result in a TypeError. See #4564.

jaxlib 0.1.56 (October 14, 2020)

jax 0.2.3 (October 14 2020)

  • GitHub commits.

  • The reason for another release so soon is we need to temporarily roll back a new jit fastpath while we look into a performance degradation

jax 0.2.2 (October 13 2020)

jax 0.2.1 (October 6 2020)

jax (0.2.0) (September 23 2020)

jax (0.1.77) (September 15 2020)

jaxlib 0.1.55 (September 8, 2020)

  • Update XLA:

    • Fix bug in DLPackManagedTensorToBuffer (#4196)

jax 0.1.76 (September 8, 2020)

jax 0.1.75 (July 30, 2020)

  • GitHub commits.

  • Bug Fixes:

    • make jnp.abs() work for unsigned inputs (#3914)

  • Improvements:

    • “Omnistaging” behavior added behind a flag, disabled by default (#3370)

jax 0.1.74 (July 29, 2020)

  • GitHub commits.

  • New Features:

    • BFGS (#3101)

    • TPU support for half-precision arithmetic (#3878)

  • Bug Fixes:

    • Prevent some accidental dtype warnings (#3874)

    • Fix a multi-threading bug in custom derivatives (#3845, #3869)

  • Improvements:

    • Faster searchsorted implementation (#3873)

    • Better test coverage for jax.numpy sorting algorithms (#3836)

jaxlib 0.1.52 (July 22, 2020)

  • Update XLA.

jax 0.1.73 (July 22, 2020)

  • GitHub commits.

  • The minimum jaxlib version is now 0.1.51.

  • New Features:

    • jax.image.resize. (#3703)

    • hfft and ihfft (#3664)

    • jax.numpy.intersect1d (#3726)

    • jax.numpy.lexsort (#3812)

    • lax.scan and the scan primitive support an unroll parameter for loop unrolling when lowering to XLA (#3738).

  • Bug Fixes:

    • Fix reduction repeated axis error (#3618)

    • Fix shape rule for lax.pad for input dimensions of size 0. (#3608)

    • make psum transpose handle zero cotangents (#3653)

    • Fix shape error when taking JVP of reduce-prod over size 0 axis. (#3729)

    • Support differentiation through jax.lax.all_to_all (#3733)

    • address nan issue in jax.scipy.special.zeta (#3777)

  • Improvements:

    • Many improvements to jax2tf

    • Reimplement argmin/argmax using a single pass variadic reduction. (#3611)

    • Enable XLA SPMD partitioning by default. (#3151)

    • Add support for 0d transpose convolution (#3643)

    • Make LU gradient work for low-rank matrices (#3610)

    • support multiple_results and custom JVPs in jet (#3657)

    • Generalize reduce-window padding to support (lo, hi) pairs. (#3728)

    • Implement complex convolutions on CPU and GPU. (#3735)

    • Make jnp.take work for empty slices of empty arrays. (#3751)

    • Relax dimension ordering rules for dot_general. (#3778)

    • Enable buffer donation for GPU. (#3800)

    • Add support for base dilation and window dilation to reduce window op… (#3803)

jaxlib 0.1.51 (July 2, 2020)

  • Update XLA.

  • Add new runtime support for host_callback.

jax 0.1.72 (June 28, 2020)

jax 0.1.71 (June 25, 2020)

  • GitHub commits.

  • The minimum jaxlib version is now 0.1.48.

  • Bug fixes:

    • Allow jax.experimental.ode.odeint dynamics functions to close over values with respect to which we’re differentiating #3562.

jaxlib 0.1.50 (June 25, 2020)

  • Add support for CUDA 11.0.

  • Drop support for CUDA 9.2 (we only maintain support for the last four CUDA versions.)

  • Update XLA.

jaxlib 0.1.49 (June 19, 2020)

jaxlib 0.1.48 (June 12, 2020)

  • New features:

    • Adds support for fast traceback collection.

    • Adds preliminary support for on-device heap profiling.

    • Implements np.nextafter for bfloat16 types.

    • Complex128 support for FFTs on CPU and GPU.

  • Bugfixes:

    • Improved float64 tanh accuracy on GPU.

    • float64 scatters on GPU are much faster.

    • Complex matrix multiplication on CPU should be much faster.

    • Stable sorts on CPU should actually be stable now.

    • Concurrency bug fix in CPU backend.

jax 0.1.70 (June 8, 2020)

  • GitHub commits.

  • New features:

    • lax.switch introduces indexed conditionals with multiple branches, together with a generalization of the cond primitive #3318.

jax 0.1.69 (June 3, 2020)

jax 0.1.68 (May 21, 2020)

jax 0.1.67 (May 12, 2020)

  • GitHub commits.

  • New features:

    • Support for reduction over subsets of a pmapped axis using axis_index_groups #2382.

    • Experimental support for printing and calling host-side Python function from compiled code. See id_print and id_tap (#3006).

  • Notable changes:

    • The visibility of names exported from jax.numpy has been tightened. This may break code that was making use of names that were previously exported accidentally.

jaxlib 0.1.47 (May 8, 2020)

  • Fixes crash for outfeed.

jax 0.1.66 (May 5, 2020)

jaxlib 0.1.46 (May 5, 2020)

  • Fixes crash for linear algebra functions on Mac OS X (#432).

  • Fixes an illegal instruction crash caused by using AVX512 instructions when an operating system or hypervisor disabled them (#2906).

jax 0.1.65 (April 30, 2020)

  • GitHub commits.

  • New features:

    • Differentiation of determinants of singular matrices #2809.

  • Bug fixes:

    • Fix odeint() differentiation with respect to time of ODEs with time-dependent dynamics #2817, also add ODE CI testing.

    • Fix lax_linalg.qr() differentiation #2867.

jaxlib 0.1.45 (April 21, 2020)

  • Fixes segfault: #2755

  • Plumb is_stable option on Sort HLO through to Python.

jax 0.1.64 (April 21, 2020)

jaxlib 0.1.44 (April 16, 2020)

  • Fixes a bug where if multiple GPUs of different models were present, JAX would only compile programs suitable for the first GPU.

  • Bugfix for batch_group_count convolutions.

  • Added precompiled SASS for more GPU versions to avoid startup PTX compilation hang.

jax 0.1.63 (April 12, 2020)

  • GitHub commits.

  • Added jax.custom_jvp and jax.custom_vjp from #2026, see the tutorial notebook. Deprecated jax.custom_transforms and removed it from the docs (though it still works).

  • Add scipy.sparse.linalg.cg #2566.

  • Changed how Tracers are printed to show more useful information for debugging #2591.

  • Made jax.numpy.isclose handle nan and inf correctly #2501.

  • Added several new rules for jax.experimental.jet #2537.

  • Fixed jax.experimental.stax.BatchNorm when scale/center isn’t provided.

  • Fix some missing cases of broadcasting in jax.numpy.einsum #2512.

  • Implement jax.numpy.cumsum and jax.numpy.cumprod in terms of a parallel prefix scan #2596 and make reduce_prod differentiable to arbitray order #2597.

  • Add batch_group_count to conv_general_dilated #2635.

  • Add docstring for test_util.check_grads #2656.

  • Add callback_transform #2665.

  • Implement rollaxis, convolve/correlate 1d & 2d, copysign, trunc, roots, and quantile/percentile interpolation options.

jaxlib 0.1.43 (March 31, 2020)

  • Fixed a performance regression for Resnet-50 on GPU.

jax 0.1.62 (March 21, 2020)

  • GitHub commits.

  • JAX has dropped support for Python 3.5. Please upgrade to Python 3.6 or newer.

  • Removed the internal function lax._safe_mul, which implemented the convention 0. * nan == 0.. This change means some programs when differentiated will produce nans when they previously produced correct values, though it ensures nans rather than silently incorrect results are produced for other programs. See #2447 and #1052 for details.

  • Added an all_gather parallel convenience function.

  • More type annotations in core code.

jaxlib 0.1.42 (March 19, 2020)

  • jaxlib 0.1.41 broke cloud TPU support due to an API incompatibility. This release fixes it again.

  • JAX has dropped support for Python 3.5. Please upgrade to Python 3.6 or newer.

jax 0.1.61 (March 17, 2020)

  • GitHub commits.

  • Fixes Python 3.5 support. This will be the last JAX or jaxlib release that supports Python 3.5.

jax 0.1.60 (March 17, 2020)

  • GitHub commits.

  • New features:

    • jax.pmap() has static_broadcast_argnums argument which allows the user to specify arguments that should be treated as compile-time constants and should be broadcasted to all devices. It works analogously to static_argnums in jax.jit().

    • Improved error messages for when tracers are mistakenly saved in global state.

    • Added jax.nn.one_hot() utility function.

    • Added jax.experimental.jet for exponentially faster higher-order automatic differentiation.

    • Added more correctness checking to arguments of jax.lax.broadcast_in_dim().

  • The minimum jaxlib version is now 0.1.41.

jaxlib 0.1.40 (March 4, 2020)

  • Adds experimental support in Jaxlib for TensorFlow profiler, which allows tracing of CPU and GPU computations from TensorBoard.

  • Includes prototype support for multihost GPU computations that communicate via NCCL.

  • Improves performance of NCCL collectives on GPU.

  • Adds TopK, CustomCallWithoutLayout, CustomCallWithLayout, IGammaGradA and RandomGamma implementations.

  • Supports device assignments known at XLA compilation time.

jax 0.1.59 (February 11, 2020)

  • GitHub commits.

  • Breaking changes

    • The minimum jaxlib version is now 0.1.38.

    • Simplified Jaxpr by removing the Jaxpr.freevars and Jaxpr.bound_subjaxprs. The call primitives (xla_call, xla_pmap, sharded_call, and remat_call) get a new parameter call_jaxpr with a fully-closed (no constvars) jaxpr. Also, added a new field call_primitive to primitives.

  • New features:

    • Reverse-mode automatic differentiation (e.g. grad) of lax.cond, making it now differentiable in both modes (#2091)

    • JAX now supports DLPack, which allows sharing CPU and GPU arrays in a zero-copy way with other libraries, such as PyTorch.

    • JAX GPU DeviceArrays now support __cuda_array_interface__, which is another zero-copy protocol for sharing GPU arrays with other libraries such as CuPy and Numba.

    • JAX CPU device buffers now implement the Python buffer protocol, which allows zero-copy buffer sharing between JAX and NumPy.

    • Added JAX_SKIP_SLOW_TESTS environment variable to skip tests known as slow.

jaxlib 0.1.39 (February 11, 2020)

  • Updates XLA.

jaxlib 0.1.38 (January 29, 2020)

  • CUDA 9.0 is no longer supported.

  • CUDA 10.2 wheels are now built by default.

jax 0.1.58 (January 28, 2020)

Notable bug fixes

  • With the Python 3 upgrade, JAX no longer depends on fastcache, which should help with installation.