author int64 658 755k | date stringdate 2012-06-12 08:34:29 2024-07-22 14:51:21 | timezone int64 -46,800 43.2k | hash stringlengths 40 40 | message stringlengths 5 490 | mods listlengths 1 16 | language stringclasses 20
values | license stringclasses 3
values | repo stringlengths 5 68 | original_message stringlengths 12 491 |
|---|---|---|---|---|---|---|---|---|---|
260,411 | 24.02.2021 09:21:36 | -3,600 | 4800fbb28596f360665aa6d253c01c126f312e15 | [jax2tf] Updated documentation | [
{
"change_type": "MODIFY",
"old_path": "jax/experimental/jax2tf/README.md",
"new_path": "jax/experimental/jax2tf/README.md",
"diff": "-# JAX to TensorFlow converter (jax2tf)\n+# JAX and TensorFlow interoperation (jax2tf/call_tf)\n-This package provides an experimental JAX converter that can take a f... | Python | Apache License 2.0 | google/jax | [jax2tf] Updated documentation |
260,299 | 24.02.2021 09:56:01 | 0 | 661914ecfd597afeb9de324a673d1c8edc847a91 | Add some more detail to the sharded_jit docstring | [
{
"change_type": "MODIFY",
"old_path": "jax/interpreters/sharded_jit.py",
"new_path": "jax/interpreters/sharded_jit.py",
"diff": "@@ -290,19 +290,24 @@ def sharded_jit(fun: Callable, in_parts, out_parts, num_partitions: int = None,\nArgs:\nfun: Function to be jitted.\n- in_parts: The input partition... | Python | Apache License 2.0 | google/jax | Add some more detail to the sharded_jit docstring |
260,411 | 24.02.2021 18:26:38 | -3,600 | 192647898140bbda474958ba3b5600a371a65182 | [jax2tf] Fine-tuning the documentation | [
{
"change_type": "MODIFY",
"old_path": "jax/experimental/jax2tf/README.md",
"new_path": "jax/experimental/jax2tf/README.md",
"diff": "# JAX and TensorFlow interoperation (jax2tf/call_tf)\nThis package provides experimental support for interoperation between JAX and TensorFlow.\n+There are two intero... | Python | Apache License 2.0 | google/jax | [jax2tf] Fine-tuning the documentation |
260,719 | 06.02.2021 23:56:07 | -19,080 | c223ef9d07d602152d0d69601edd32869f632f73 | add support for setxor1d | [
{
"change_type": "MODIFY",
"old_path": "docs/jax.numpy.rst",
"new_path": "docs/jax.numpy.rst",
"diff": "@@ -317,6 +317,7 @@ Not every function in NumPy is implemented; contributions are welcome!\nselect\nset_printoptions\nsetdiff1d\n+ setxor1d\nshape\nsign\nsignbit\n"
},
{
"change_type": "MO... | Python | Apache License 2.0 | google/jax | add support for setxor1d |
260,335 | 24.02.2021 20:25:24 | 28,800 | 9789677e85f43f8fa0f1c96e0b91276017af3dcf | reviewer comments, delint | [
{
"change_type": "MODIFY",
"old_path": "docs/autodidax.ipynb",
"new_path": "docs/autodidax.ipynb",
"diff": "\"where we apply primitive functions to numerical inputs to produce numerical\\n\",\n\"outputs, we want to override primitive application and let different values\\n\",\n\"flow through our pro... | Python | Apache License 2.0 | google/jax | reviewer comments, delint |
260,411 | 01.03.2021 21:56:13 | -3,600 | b43cd47b33315a72cfc5e7797d4f505ca2749b5d | Update jax/experimental/jax2tf/README.md | [
{
"change_type": "MODIFY",
"old_path": "jax/experimental/jax2tf/README.md",
"new_path": "jax/experimental/jax2tf/README.md",
"diff": "This package provides experimental support for interoperation between JAX and TensorFlow.\nThere are two interoperation directions:\n-(a) `jax2tf.convert`: using JAX ... | Python | Apache License 2.0 | google/jax | Update jax/experimental/jax2tf/README.md
Co-authored-by: 8bitmp3 <19637339+8bitmp3@users.noreply.github.com> |
260,411 | 01.03.2021 23:27:11 | -3,600 | 410f24af6bf625c2194e1fc42f72a12d7682bf3e | [jax2tf] Fix conversion for integer division for TF 1
It seems that TF 1 has a different broadcasting behavior than TF2. Change
the conversion of integer division to use `tf.not_equal` instead of `!=`. | [
{
"change_type": "MODIFY",
"old_path": "jax/experimental/jax2tf/jax2tf.py",
"new_path": "jax/experimental/jax2tf/jax2tf.py",
"diff": "@@ -956,8 +956,9 @@ tf_impl[lax.iota_p] = _iota\ndef _div(lhs, rhs):\nif lhs.dtype.is_integer:\nquotient = tf.math.floordiv(lhs, rhs)\n- select = tf.math.logical_and(... | Python | Apache License 2.0 | google/jax | [jax2tf] Fix conversion for integer division for TF 1
Bug: #5831
It seems that TF 1 has a different broadcasting behavior than TF2. Change
the conversion of integer division to use `tf.not_equal` instead of `!=`. |
260,335 | 02.03.2021 22:54:49 | 28,800 | 4a0f6e35e6d609de23a9ebfdfb76b4ce654fb313 | relax tolerance in jax2tf qr to avoid flakiness | [
{
"change_type": "MODIFY",
"old_path": "jax/experimental/jax2tf/tests/jax2tf_limitations.py",
"new_path": "jax/experimental/jax2tf/tests/jax2tf_limitations.py",
"diff": "@@ -949,7 +949,7 @@ class Jax2TfLimitation(primitive_harness.Limitation):\n# compiling with TF is expected to have worse performan... | Python | Apache License 2.0 | google/jax | relax tolerance in jax2tf qr to avoid flakiness
PiperOrigin-RevId: 360596515 |
260,411 | 03.03.2021 12:37:01 | -3,600 | 469e1f05e14964e5f0f0095e808cf2cd5c974fd7 | [jax2tf] Updated the limitations
Update the documentation to reflect the improvements in TF support for dtypes | [
{
"change_type": "MODIFY",
"old_path": "jax/experimental/jax2tf/g3doc/jax_primitives_coverage.md",
"new_path": "jax/experimental/jax2tf/g3doc/jax_primitives_coverage.md",
"diff": "# Primitives with limited JAX support\n-*Last generated on: 2021-02-02* (YYYY-MM-DD)\n+*Last generated on: 2021-03-03* (... | Python | Apache License 2.0 | google/jax | [jax2tf] Updated the limitations
Update the documentation to reflect the improvements in TF support for dtypes |
260,496 | 04.03.2021 16:59:40 | -3,600 | 70444df6ded1d5f5f8131dfdf9b6a102cf5042b1 | Add python to codeblocks | [
{
"change_type": "MODIFY",
"old_path": "jax/experimental/jax2tf/README.md",
"new_path": "jax/experimental/jax2tf/README.md",
"diff": "@@ -74,7 +74,7 @@ Just like for regular TensorFlow functions, it is possible to include in the\nSavedModel multiple versions of a function for different input shapes,... | Python | Apache License 2.0 | google/jax | Add python to codeblocks |
260,287 | 04.03.2021 13:11:19 | 0 | 84c00653491577c9c085df71141ef2fe807904c3 | Only vectorize xmap axes that have only one element per resource
To make the jaxpr much less noisy. size-1 vmaps are quite pointless. | [
{
"change_type": "MODIFY",
"old_path": "jax/experimental/maps.py",
"new_path": "jax/experimental/maps.py",
"diff": "@@ -562,10 +562,9 @@ def make_xmap_callable(fun: lu.WrappedFun,\nclass EvaluationPlan(NamedTuple):\n\"\"\"Encapsulates preprocessing common to top-level xmap invocations and its transl... | Python | Apache License 2.0 | google/jax | Only vectorize xmap axes that have only one element per resource
To make the jaxpr much less noisy. size-1 vmaps are quite pointless. |
260,287 | 04.03.2021 18:08:45 | 0 | 6884f21b60bfb4ea181ab943cadfb7b38dfcaedb | Fix batching formula of xmap
Turns out that once you insert multiple dimensions things become much
more tricky than in the case of batching a one-dimensional map. Also
strenghten our tests to make sure we don't depend too much on the
semantics of the einsum batching rule. | [
{
"change_type": "MODIFY",
"old_path": "jax/experimental/maps.py",
"new_path": "jax/experimental/maps.py",
"diff": "@@ -686,17 +686,24 @@ def _batch_trace_process_xmap(self, primitive, f: lu.WrappedFun, tracers, params\nfor d, in_axis in zip(dims, params['in_axes']))\nf, mapped_dims_out = batching.b... | Python | Apache License 2.0 | google/jax | Fix batching formula of xmap
Turns out that once you insert multiple dimensions things become much
more tricky than in the case of batching a one-dimensional map. Also
strenghten our tests to make sure we don't depend too much on the
semantics of the einsum batching rule. |
260,335 | 05.03.2021 10:12:58 | 28,800 | 206acc1ed26950ec9c5a0834d6a06763ffceb383 | update jax for pypi | [
{
"change_type": "MODIFY",
"old_path": "docs/CHANGELOG.md",
"new_path": "docs/CHANGELOG.md",
"diff": "# Change Log\n-<!---\n-This is a comment.\n+<!--\nRemember to align the itemized text with the first line of an item within a list.\nPLEASE REMEMBER TO CHANGE THE '..master' WITH AN ACTUAL TAG in GI... | Python | Apache License 2.0 | google/jax | update jax for pypi |
260,335 | 05.03.2021 12:47:20 | 28,800 | b3fed79ad607ec71d563533f418b2fca203c2441 | add test for issue once jaxlib is updated | [
{
"change_type": "MODIFY",
"old_path": "tests/lax_control_flow_test.py",
"new_path": "tests/lax_control_flow_test.py",
"diff": "@@ -2566,6 +2566,36 @@ class LaxControlFlowTest(jtu.JaxTestCase):\nscan_v(jnp.ones([1]), jnp.arange(5).reshape((1, 5))),\n(jnp.array([1.]), jnp.array([[0., 1., 2., 3., 4.]]... | Python | Apache License 2.0 | google/jax | add test for issue #5900, once jaxlib is updated |
260,287 | 05.03.2021 17:59:16 | 0 | 2c7c86a4ba00d286361561b3d3a39d2bed287557 | Reenable multi-axis all_to_all | [
{
"change_type": "MODIFY",
"old_path": "jax/_src/lax/parallel.py",
"new_path": "jax/_src/lax/parallel.py",
"diff": "@@ -911,12 +911,17 @@ def _all_to_all_batcher(vals_in, dims_in, *, axis_name, split_axis, concat_axis,\ndef _all_to_all_batched_collective(frame, vals_in, dims_in,\naxis_name, split_ax... | Python | Apache License 2.0 | google/jax | Reenable multi-axis all_to_all |
260,287 | 08.03.2021 12:27:01 | 0 | ec29275d7e40d8908fb85a9b8cf0a6719e1f9fda | Substitute axis names in nested jaxprs
Previously any collectives buried inside control flow would fail to
compile with xmap, because it would not traverse those with its name
substitution. This adds a "catch-all" default substitution rule which
recursively applies to all jaxpr found in the params (at the top level). | [
{
"change_type": "MODIFY",
"old_path": "jax/core.py",
"new_path": "jax/core.py",
"diff": "@@ -1529,7 +1529,31 @@ def used_axis_names(primitive: Primitive, params: ParamDict) -> Set[AxisName]:\ndef subst_axis_names(primitive: Primitive, params: ParamDict, subst: AxisSubst) -> ParamDict:\nif primitive... | Python | Apache License 2.0 | google/jax | Substitute axis names in nested jaxprs
Previously any collectives buried inside control flow would fail to
compile with xmap, because it would not traverse those with its name
substitution. This adds a "catch-all" default substitution rule which
recursively applies to all jaxpr found in the params (at the top level). |
260,389 | 24.02.2021 11:03:11 | 28,800 | ba0f785a1fbfd7edb602345c52545075ba981dab | allow named axes on while_loop condition aval | [
{
"change_type": "MODIFY",
"old_path": "jax/_src/lax/control_flow.py",
"new_path": "jax/_src/lax/control_flow.py",
"diff": "@@ -271,7 +271,9 @@ def while_loop(cond_fun: Callable[[T], bool],\nif not treedef_is_leaf(cond_tree) or len(cond_jaxpr.out_avals) != 1:\nmsg = \"cond_fun must return a boolean ... | Python | Apache License 2.0 | google/jax | allow named axes on while_loop condition aval |
260,389 | 04.03.2021 04:13:04 | 28,800 | 47f17f04804ea7318cf0e7d8bd8a8e046ab10ffd | add (broken) test | [
{
"change_type": "MODIFY",
"old_path": "tests/lax_control_flow_test.py",
"new_path": "tests/lax_control_flow_test.py",
"diff": "@@ -335,6 +335,20 @@ class LaxControlFlowTest(jtu.JaxTestCase):\nexpected = np.array([4, 3, 4, 3])\nself.assertAllClose(ans, expected, check_dtypes=False)\n+ def testWhileL... | Python | Apache License 2.0 | google/jax | add (broken) test |
260,335 | 08.03.2021 19:49:10 | 28,800 | 2b9ffb1fb3a72fea413c0f034ce4a31baf657631 | make axis_index bind respect dynamic traces | [
{
"change_type": "MODIFY",
"old_path": "jax/_src/lax/parallel.py",
"new_path": "jax/_src/lax/parallel.py",
"diff": "@@ -1097,13 +1097,9 @@ def _all_gather_abstract_eval(x, *, all_gather_dimension, axis_name, axis_index_\ndef _all_gather_transpose_rule(cts, x, *, all_gather_dimension, axis_name, axis... | Python | Apache License 2.0 | google/jax | make axis_index bind respect dynamic traces |
260,510 | 09.03.2021 15:21:07 | 28,800 | ddaef193fe5188ba8cc83062baafd4976d40c974 | Add scan and while rule for jax.experimental.callback transformation | [
{
"change_type": "MODIFY",
"old_path": "jax/experimental/callback.py",
"new_path": "jax/experimental/callback.py",
"diff": "# See the License for the specific language governing permissions and\n# limitations under the License.\n+import itertools as it\n+\nfrom typing import Any, Callable, Dict, Seq... | Python | Apache License 2.0 | google/jax | Add scan and while rule for jax.experimental.callback transformation |
260,335 | 10.03.2021 12:20:36 | 28,800 | d11bba9cafcaea1c469f5a49e7abf1cfefbd5967 | tweak broken tests | [
{
"change_type": "MODIFY",
"old_path": "tests/api_test.py",
"new_path": "tests/api_test.py",
"diff": "@@ -2134,9 +2134,9 @@ class APITest(jtu.JaxTestCase):\nwith self.assertRaisesRegex(core.ConcretizationTypeError, msg):\nf()\n- # TODO(jakevdp): re-enable this if possible.\n- @unittest.skipIf(True, ... | Python | Apache License 2.0 | google/jax | tweak broken tests |
260,293 | 11.03.2021 11:32:43 | -3,600 | 6470094f04ba7c02c8574051ec3128eff3f11def | Fix import of datasets in differentially private SGD example | [
{
"change_type": "MODIFY",
"old_path": "examples/differentially_private_sgd.py",
"new_path": "examples/differentially_private_sgd.py",
"diff": "@@ -79,7 +79,7 @@ from jax.experimental import optimizers\nfrom jax.experimental import stax\nfrom jax.tree_util import tree_flatten, tree_unflatten\nimport... | Python | Apache License 2.0 | google/jax | Fix import of datasets in differentially private SGD example |
260,482 | 06.03.2021 11:53:01 | -3,600 | d743aa5803be9264cbcfcd98f679954309d6517e | Added 'where' keyword to 'jnp.{mean, var, std}' | [
{
"change_type": "MODIFY",
"old_path": "jax/_src/numpy/lax_numpy.py",
"new_path": "jax/_src/numpy/lax_numpy.py",
"diff": "@@ -2013,15 +2013,15 @@ def min(a, axis: Optional[Union[int, Tuple[int, ...]]] = None, out=None,\n@_wraps(np.all)\ndef all(a, axis: Optional[Union[int, Tuple[int, ...]]] = None, ... | Python | Apache License 2.0 | google/jax | Added 'where' keyword to 'jnp.{mean, var, std}' |
260,490 | 15.03.2021 23:26:31 | 14,400 | 9d28b67022030d324f501339e1032ce72de2e8ab | Fixed two small typos in jax.lax. | [
{
"change_type": "MODIFY",
"old_path": "jax/_src/lax/lax.py",
"new_path": "jax/_src/lax/lax.py",
"diff": "@@ -3848,7 +3848,7 @@ def _slice_shape_rule(operand, *, start_indices, limit_indices, strides):\nraise TypeError(msg.format(start_indices, operand.shape))\nif len(start_indices) != len(limit_ind... | Python | Apache License 2.0 | google/jax | Fixed two small typos in jax.lax. |
260,411 | 16.03.2021 11:38:57 | -3,600 | 840d5162030576d45edb7830408fdf867b8ff6b8 | [jax2tf] Removed more traces of support for batch polymorphism
See issue
* Also cleanup the examples | [
{
"change_type": "MODIFY",
"old_path": "jax/experimental/jax2tf/README.md",
"new_path": "jax/experimental/jax2tf/README.md",
"diff": "@@ -152,7 +152,7 @@ for CUDA. One must be mindful to install a version of CUDA that is compatible\nwith both [jaxlib](https://github.com/google/jax/blob/master/README... | Python | Apache License 2.0 | google/jax | [jax2tf] Removed more traces of support for batch polymorphism
See issue #6080
* Also cleanup the examples |
260,287 | 15.03.2021 13:23:21 | 0 | 92bca61737a434deb7500c86d8fe944a42a7e3fb | Fix incorrect handling of constants in xmap's partial_eval rule | [
{
"change_type": "MODIFY",
"old_path": "jax/experimental/maps.py",
"new_path": "jax/experimental/maps.py",
"diff": "@@ -659,7 +659,7 @@ core.axis_substitution_rules[xmap_p] = _xmap_axis_subst\ndef _dynamic_jaxpr_process_xmap(self, primitive, f, tracers, params):\nfrom jax.interpreters.partial_eval i... | Python | Apache License 2.0 | google/jax | Fix incorrect handling of constants in xmap's partial_eval rule |
260,335 | 17.03.2021 12:40:23 | 25,200 | 0181d03902dbc8e251e1b6b190d801f4f29c751e | add a memory leak test for jit jaxpr construction
Tweak implementation for `_inline_literals` not to include a class
defined in a function, since that seemed to cause leaking! | [
{
"change_type": "MODIFY",
"old_path": "jax/interpreters/partial_eval.py",
"new_path": "jax/interpreters/partial_eval.py",
"diff": "@@ -986,11 +986,8 @@ class JaxprStackFrame:\ndef _inline_literals(jaxpr, constvals):\nconsts = dict(zip(jaxpr.constvars, constvals))\nnewvar = core.gensym()\n- class va... | Python | Apache License 2.0 | google/jax | add a memory leak test for jit jaxpr construction
Tweak implementation for `_inline_literals` not to include a class
defined in a function, since that seemed to cause leaking! |
260,411 | 17.03.2021 08:40:48 | -3,600 | dace15572b2249684da8aa1b813f695ad6f21f28 | [jax2tf] Added instructions for using OSS TensorFlow model server. | [
{
"change_type": "MODIFY",
"old_path": "jax/experimental/jax2tf/examples/mnist_lib.py",
"new_path": "jax/experimental/jax2tf/examples/mnist_lib.py",
"diff": "@@ -178,8 +178,8 @@ class PureJaxMNIST:\ntest_acc = PureJaxMNIST.accuracy(PureJaxMNIST.predict, params, test_ds)\nlogging.info(\nf\"{PureJaxMN... | Python | Apache License 2.0 | google/jax | [jax2tf] Added instructions for using OSS TensorFlow model server. |
260,411 | 18.03.2021 08:53:00 | -3,600 | 917db30d9bc1c1eadb4c85f930caa6a8c6ac1714 | Fix links and reviewer-suggested edits | [
{
"change_type": "MODIFY",
"old_path": "jax/experimental/jax2tf/examples/serving/README.md",
"new_path": "jax/experimental/jax2tf/examples/serving/README.md",
"diff": "@@ -2,29 +2,29 @@ Using jax2tf with TensorFlow serving\n====================================\nThis is a supplement to the\n-[example... | Python | Apache License 2.0 | google/jax | Fix links and reviewer-suggested edits |
260,411 | 18.03.2021 10:23:28 | -3,600 | 8f846a3aa96074bf5c9acdb7ccc88f9223cd429e | [host_callback] Enable configuring the outfeed receiver buffer size
The C++ runtime is already configurable. We are adding here the ability
to pass the configuration parameter by environment variable and commmand-line
flag.
Fixes: | [
{
"change_type": "MODIFY",
"old_path": "jax/experimental/host_callback.py",
"new_path": "jax/experimental/host_callback.py",
"diff": "@@ -229,7 +229,8 @@ the send operations were performed on the device.\nThe host callback functions for multiple devices may be interleaved.\nThe data from the devices... | Python | Apache License 2.0 | google/jax | [host_callback] Enable configuring the outfeed receiver buffer size
The C++ runtime is already configurable. We are adding here the ability
to pass the configuration parameter by environment variable and commmand-line
flag.
Fixes: #5782 |
260,411 | 18.03.2021 11:09:33 | -3,600 | 52408b35a8943ee69ba709f6ba7f9792924d27be | Move the flag definitions to config
Without this when running tests the flag parsing may happen before the
host_callback module is loaded and then the host_callback flags may
be left undefined. | [
{
"change_type": "MODIFY",
"old_path": "jax/config.py",
"new_path": "jax/config.py",
"diff": "@@ -191,3 +191,18 @@ flags.DEFINE_bool(\n'Enabling leak checking may have performance impacts: some caching '\n'is disabled, and other overheads may be added.'),\n)\n+\n+flags.DEFINE_bool(\n+ 'jax_host_call... | Python | Apache License 2.0 | google/jax | Move the flag definitions to config
Without this when running tests the flag parsing may happen before the
host_callback module is loaded and then the host_callback flags may
be left undefined. |
260,424 | 18.03.2021 17:32:33 | 0 | d86dd24bf8c9341fbe2ebb54900fc71606f5d09b | Make sublevel weak-referable, and enable the leak checker on sublevels.
Reimplement Sublevel to not inherit from `int`.
See docs on weakref: "CPython implementation detail: Other built-in types
such as tuple and int do not support weak references even when subclassed." | [
{
"change_type": "MODIFY",
"old_path": "jax/core.py",
"new_path": "jax/core.py",
"diff": "@@ -688,7 +688,23 @@ class TraceStack:\nnew.dynamic = self.dynamic\nreturn new\n-class Sublevel(int): pass\n+\n+@total_ordering\n+class Sublevel:\n+\n+ def __init__(self, level: int):\n+ self.level = level\n+\n... | Python | Apache License 2.0 | google/jax | Make sublevel weak-referable, and enable the leak checker on sublevels.
Reimplement Sublevel to not inherit from `int`.
See docs on weakref: "CPython implementation detail: Other built-in types
such as tuple and int do not support weak references even when subclassed." |
260,287 | 18.03.2021 11:53:10 | 25,200 | 79040f109ea8ef7d7dc299314f2af742dd174ec3 | Reenable multi-host xmap
Apparently the implementation has regressed quite a bit since I've last tried it.
Also added some tests to the internal CI (not visible on GitHub) to make sure that
we don't regress again. | [
{
"change_type": "MODIFY",
"old_path": "jax/experimental/maps.py",
"new_path": "jax/experimental/maps.py",
"diff": "@@ -16,7 +16,7 @@ import threading\nimport contextlib\nimport numpy as np\nimport itertools as it\n-from collections import OrderedDict\n+from collections import OrderedDict, namedtupl... | Python | Apache License 2.0 | google/jax | Reenable multi-host xmap
Apparently the implementation has regressed quite a bit since I've last tried it.
Also added some tests to the internal CI (not visible on GitHub) to make sure that
we don't regress again.
PiperOrigin-RevId: 363709635 |
260,424 | 18.03.2021 20:08:33 | 0 | 276645dd593466be695492073af6aa4d111ab1ab | Fix error when running a jvp of a jit of a custom_vjp.
Error before:
NotImplementedError: XLA translation rule for primitive 'custom_lin' not found
Error after:
TypeError: can't apply forward-mode autodiff (jvp) to a custom_vjp function. | [
{
"change_type": "MODIFY",
"old_path": "jax/custom_derivatives.py",
"new_path": "jax/custom_derivatives.py",
"diff": "@@ -671,6 +671,7 @@ xla.initial_style_translations[custom_vjp_call_jaxpr_p] = \\\nxla.lower_fun_initial_style(_custom_vjp_call_jaxpr_impl)\nbatching.primitive_batchers[ad.custom_lin_... | Python | Apache License 2.0 | google/jax | Fix error when running a jvp of a jit of a custom_vjp.
Error before:
NotImplementedError: XLA translation rule for primitive 'custom_lin' not found
Error after:
TypeError: can't apply forward-mode autodiff (jvp) to a custom_vjp function. |
260,335 | 18.03.2021 18:05:22 | 25,200 | ee4ec867ec94cc490f1213dddb6a65b99a624554 | make constant handlers follow type mro
fixes | [
{
"change_type": "MODIFY",
"old_path": "jax/lib/xla_bridge.py",
"new_path": "jax/lib/xla_bridge.py",
"diff": "@@ -29,11 +29,13 @@ from absl import logging\nlogging._warn_preinit_stderr = 0\nfrom ..config import flags\n-from jax._src import util\n+from jax._src import util, traceback_util\nfrom .. im... | Python | Apache License 2.0 | google/jax | make constant handlers follow type mro
fixes #6129 |
260,335 | 18.03.2021 18:56:47 | 25,200 | 90f8d1c8686be7e54132873d708c21a3ca2f0bcc | xla.flatten_shape don't use recursive generators
This is an attempt to fix a strange memory leak failure which I can't
reproduce directly but seems to fail consistently in Google internal CI. | [
{
"change_type": "MODIFY",
"old_path": "jax/interpreters/xla.py",
"new_path": "jax/interpreters/xla.py",
"diff": "@@ -634,18 +634,19 @@ def flatten_shape(s: XlaShape) -> Sequence[Tuple[Sequence[int], XlaShape]]:\nAn iterable of pairs of indices and shapes for each array within the shape\ntree.\n\"\"... | Python | Apache License 2.0 | google/jax | xla.flatten_shape don't use recursive generators
This is an attempt to fix a strange memory leak failure which I can't
reproduce directly but seems to fail consistently in Google internal CI. |
260,335 | 18.03.2021 21:46:46 | 25,200 | 9802f3378edb66edaca52bfdcfc0cf773bd697a1 | add simple single-primitive eager benchmarks | [
{
"change_type": "MODIFY",
"old_path": "benchmarks/api_benchmark.py",
"new_path": "benchmarks/api_benchmark.py",
"diff": "@@ -19,6 +19,7 @@ import google_benchmark\nimport jax\nimport jax.numpy as jnp\nimport numpy as np\n+from jax import lax\npartial = functools.partial\n@@ -36,6 +37,40 @@ def requ... | Python | Apache License 2.0 | google/jax | add simple single-primitive eager benchmarks |
260,335 | 19.03.2021 10:09:14 | 25,200 | 5c6ff67e4e042b64518ac03f7599039c99c35daa | generalize ravel_pytree to handle int types, add tests | [
{
"change_type": "MODIFY",
"old_path": "jax/flatten_util.py",
"new_path": "jax/flatten_util.py",
"diff": "# See the License for the specific language governing permissions and\n# limitations under the License.\n+import warnings\n+\n+import numpy as np\nfrom .tree_util import tree_flatten, tree_unfla... | Python | Apache License 2.0 | google/jax | generalize ravel_pytree to handle int types, add tests |
260,335 | 09.03.2021 15:58:39 | 28,800 | bf15ba5310d5f9009571928f70548bcbc7e856c3 | don't device transfer in convert_element_type | [
{
"change_type": "MODIFY",
"old_path": "jax/_src/lax/lax.py",
"new_path": "jax/_src/lax/lax.py",
"diff": "@@ -437,10 +437,8 @@ def convert_element_type(operand: Array, new_dtype: DType = None,\nmsg = \"Casting complex values to real discards the imaginary part\"\nwarnings.warn(msg, np.ComplexWarning... | Python | Apache License 2.0 | google/jax | don't device transfer in convert_element_type
Co-authored-by: Qiao Zhang <zhangqiaorjc@google.com> |
260,631 | 19.03.2021 16:35:12 | 25,200 | 4f8814a760122450287eefe30a8e9fd16a83a412 | Copybara import of the project:
by Matthew Johnson
don't device transfer in convert_element_type | [
{
"change_type": "MODIFY",
"old_path": "jax/_src/lax/lax.py",
"new_path": "jax/_src/lax/lax.py",
"diff": "@@ -437,8 +437,10 @@ def convert_element_type(operand: Array, new_dtype: DType = None,\nmsg = \"Casting complex values to real discards the imaginary part\"\nwarnings.warn(msg, np.ComplexWarning... | Python | Apache License 2.0 | google/jax | Copybara import of the project:
--
bf15ba5310d5f9009571928f70548bcbc7e856c3 by Matthew Johnson <mattjj@google.com>:
don't device transfer in convert_element_type
Co-authored-by: Qiao Zhang <zhangqiaorjc@google.com>
PiperOrigin-RevId: 363995032 |
260,335 | 19.03.2021 22:35:31 | 25,200 | 57d5c6af5f6fb43579747d5102c2d1985f9f5279 | add clz primitive | [
{
"change_type": "MODIFY",
"old_path": "jax/_src/lax/lax.py",
"new_path": "jax/_src/lax/lax.py",
"diff": "@@ -335,6 +335,10 @@ def population_count(x: Array) -> Array:\nr\"\"\"Elementwise popcount, count the number of set bits in each element.\"\"\"\nreturn population_count_p.bind(x)\n+def clz(x: Ar... | Python | Apache License 2.0 | google/jax | add clz primitive |
260,335 | 19.03.2021 23:13:20 | 25,200 | 97aca25ec48be0696e44cbf2f9547e44416e3ee3 | tell jax2tf to ignore clz | [
{
"change_type": "MODIFY",
"old_path": "jax/experimental/jax2tf/jax2tf.py",
"new_path": "jax/experimental/jax2tf/jax2tf.py",
"diff": "@@ -797,7 +797,7 @@ for unexpected in xla.call_translations: # Call primitives are inlined\n# Primitives that are not yet implemented must be explicitly declared here... | Python | Apache License 2.0 | google/jax | tell jax2tf to ignore clz |
260,335 | 21.03.2021 13:39:57 | 25,200 | af59542d00fee4394ff24b70d41500e155a05a2a | Re-applying the changes in after they had to be rolled-back. | [
{
"change_type": "MODIFY",
"old_path": "jax/_src/lax/lax.py",
"new_path": "jax/_src/lax/lax.py",
"diff": "@@ -441,10 +441,8 @@ def convert_element_type(operand: Array, new_dtype: DType = None,\nmsg = \"Casting complex values to real discards the imaginary part\"\nwarnings.warn(msg, np.ComplexWarning... | Python | Apache License 2.0 | google/jax | Re-applying the changes in #6014, after they had to be rolled-back.
PiperOrigin-RevId: 364200195 |
260,335 | 21.03.2021 15:53:24 | 25,200 | 8c3125c172b5868c29fe58863457362cbd144f3d | fix convert_element_type on large Py int inputs | [
{
"change_type": "MODIFY",
"old_path": "jax/_src/lax/lax.py",
"new_path": "jax/_src/lax/lax.py",
"diff": "@@ -2658,6 +2658,12 @@ lt_p = naryop(_fixed_dtype(np.bool_), [_any, _any], 'lt')\nad.defjvp_zero(lt_p)\n+def _convert_element_type_impl(operand, *, new_dtype, weak_type):\n+ if dtypes.is_python_... | Python | Apache License 2.0 | google/jax | fix convert_element_type on large Py int inputs |
260,335 | 21.03.2021 19:38:12 | 25,200 | fe4d12c10fb0cb50b811b10ecd0b417db1cb886f | move logic to traceable | [
{
"change_type": "MODIFY",
"old_path": "jax/_src/lax/lax.py",
"new_path": "jax/_src/lax/lax.py",
"diff": "@@ -428,8 +428,8 @@ def convert_element_type(operand: Array, new_dtype: DType = None,\nif hasattr(operand, '__jax_array__'):\noperand = operand.__jax_array__()\n- # Note: don't canonicalize old_... | Python | Apache License 2.0 | google/jax | move logic to traceable |
260,335 | 21.03.2021 19:41:04 | 25,200 | 214d273d8cbf05b7c8ccf3b96f778fb581425e52 | undo changes to host_callback (not needed anymore) | [
{
"change_type": "MODIFY",
"old_path": "jax/_src/lax/lax.py",
"new_path": "jax/_src/lax/lax.py",
"diff": "@@ -443,7 +443,9 @@ def convert_element_type(operand: Array, new_dtype: DType = None,\n# Python has big integers, but convert_element_type(2 ** 100, np.float32) need\n# not be an error since the... | Python | Apache License 2.0 | google/jax | undo changes to host_callback (not needed anymore) |
260,510 | 09.03.2021 18:40:25 | 28,800 | 252bd6c0c8cb9f337eb29415fdd56033b53a3385 | Add support for custom derivatives in jax.experimental.callback | [
{
"change_type": "MODIFY",
"old_path": "jax/experimental/callback.py",
"new_path": "jax/experimental/callback.py",
"diff": "@@ -21,13 +21,15 @@ import jax.numpy as jnp\nfrom jax import core\nfrom jax.core import Trace, Tracer, jaxpr_as_fun\nfrom jax import lax\n+from jax import custom_derivatives as... | Python | Apache License 2.0 | google/jax | Add support for custom derivatives in jax.experimental.callback |
260,677 | 22.03.2021 21:29:39 | 0 | 809d6895824fd82cecc60a67d70bdfe913e7efb1 | Increase threshold for switching to unbatched triangular solve on GPU | [
{
"change_type": "MODIFY",
"old_path": "jax/_src/lax/linalg.py",
"new_path": "jax/_src/lax/linalg.py",
"diff": "@@ -707,7 +707,7 @@ def _triangular_solve_gpu_translation_rule(trsm_impl,\nif conjugate_a and not transpose_a:\na = xops.Conj(a)\nconjugate_a = False\n- if batch > 1 and m <= 32 and n <= 3... | Python | Apache License 2.0 | google/jax | Increase threshold for switching to unbatched triangular solve on GPU |
260,335 | 23.03.2021 19:13:15 | 25,200 | 70cd62d5454c07d6a7b1141c597afca594eaaebc | disable disable_omnistaging | [
{
"change_type": "MODIFY",
"old_path": "jax/config.py",
"new_path": "jax/config.py",
"diff": "@@ -154,19 +154,14 @@ class Config:\nraise Exception(\"can't re-enable omnistaging after it's been disabled\")\ndef disable_omnistaging(self):\n+ return\n+\n+ def temporary_hack_do_not_call_me(self):\nif se... | Python | Apache License 2.0 | google/jax | disable disable_omnistaging |
260,335 | 23.03.2021 20:58:52 | 25,200 | 89768a3d2898ee7e08425f90ccf1435449c07ee6 | add jax_default_matmul_precision flag & context mngr | [
{
"change_type": "MODIFY",
"old_path": "jax/__init__.py",
"new_path": "jax/__init__.py",
"diff": "@@ -31,7 +31,8 @@ del _cloud_tpu_init\n# flake8: noqa: F401\nfrom .config import (config, enable_checks, check_tracer_leaks, checking_leaks,\n- debug_nans, debug_infs, log_compiles)\n+ debug_nans, debug... | Python | Apache License 2.0 | google/jax | add jax_default_matmul_precision flag & context mngr |
260,335 | 24.03.2021 16:25:50 | 25,200 | a261c768f2c216a54673c4c7a89336f399a50292 | simplify cotangent dropping in ad.backward_pass
Since variables in jaxprs are only assigned-to once, when we transpose
them we end up reading a variable's cotangent value only once. That
means we can pop the cotangent environment's reference to a cotangent
value in read_cotangent. | [
{
"change_type": "MODIFY",
"old_path": "jax/interpreters/ad.py",
"new_path": "jax/interpreters/ad.py",
"diff": "import functools\nimport itertools as it\n-from typing import Any, Callable, Dict, Set, List\n+from typing import Any, Callable, Dict\nfrom . import partial_eval as pe\nfrom ..config impor... | Python | Apache License 2.0 | google/jax | simplify cotangent dropping in ad.backward_pass
Since variables in jaxprs are only assigned-to once, when we transpose
them we end up reading a variable's cotangent value only once. That
means we can pop the cotangent environment's reference to a cotangent
value in read_cotangent. |
260,335 | 24.03.2021 23:00:51 | 25,200 | 278b3ae16e8d883e6a2763519fe13350e2e78c10 | add test for backward pass ref dropping | [
{
"change_type": "MODIFY",
"old_path": "tests/api_test.py",
"new_path": "tests/api_test.py",
"diff": "@@ -2452,6 +2452,23 @@ class APITest(jtu.JaxTestCase):\nself.assertIn('precision=HIGH', str(jaxpr))\nself.assertEqual(prev_val, config._read(\"jax_default_matmul_precision\"))\n+ def test_backward_p... | Python | Apache License 2.0 | google/jax | add test for backward pass ref dropping |
260,424 | 25.03.2021 19:15:45 | 0 | 25a2e1a66ab1abd0361664cb1c77d0ea6a272c78 | Fix the changelog link in the README. | [
{
"change_type": "MODIFY",
"old_path": "README.md",
"new_path": "README.md",
"diff": "| [**Transformations**](#transformations)\n| [**Install guide**](#installation)\n| [**Neural net libraries**](#neural-network-libraries)\n-| [**Change logs**](https://jax.readthedocs.io/en/latest/CHANGELOG.html)\n+... | Python | Apache License 2.0 | google/jax | Fix the changelog link in the README. |
260,335 | 25.03.2021 22:15:18 | 25,200 | 848fed8b873e41c7d4b746aa5bcfffe5700c8005 | Work around CPython bug exposed by changes to C++ JIT dispatch path. | [
{
"change_type": "MODIFY",
"old_path": "jax/api.py",
"new_path": "jax/api.py",
"diff": "@@ -133,9 +133,17 @@ float0 = dtypes.float0\ndef _check_callable(fun):\nif not callable(fun):\nraise TypeError(f\"Expected a callable value, got {fun}\")\n- if inspect.isgeneratorfunction(fun):\n+ if _isgenerator... | Python | Apache License 2.0 | google/jax | Work around CPython bug https://bugs.python.org/issue33261 exposed by changes to C++ JIT dispatch path.
PiperOrigin-RevId: 365189779 |
260,299 | 26.03.2021 10:50:24 | 0 | 0a3ba6f2cea609dfbdddea3ef82cbb10fc02e5a7 | Instantiate zero outputs of linear_transpose | [
{
"change_type": "MODIFY",
"old_path": "jax/api.py",
"new_path": "jax/api.py",
"diff": "@@ -1963,7 +1963,9 @@ def linear_transpose(fun: Callable, *primals) -> Callable:\nraise TypeError(\"cotangent type does not match function output, \"\nf\"expected {out_avals} but got {out_cotangents}\")\ndummies ... | Python | Apache License 2.0 | google/jax | Instantiate zero outputs of linear_transpose |
260,335 | 27.03.2021 19:34:42 | 25,200 | 5a055b1dad0adfe147d67c2b59d35af6e7719322 | add WIP disclaimer to autodidax, fix some typos | [
{
"change_type": "MODIFY",
"old_path": "docs/autodidax.ipynb",
"new_path": "docs/autodidax.ipynb",
"diff": "\"\\n\",\n\"Ever want to learn how JAX works, but the implementation seemed impenetrable?\\n\",\n\"Well, you're in luck! By reading this tutorial, you'll learn every big idea in\\n\",\n- \"JAX... | Python | Apache License 2.0 | google/jax | add WIP disclaimer to autodidax, fix some typos |
260,470 | 27.03.2021 21:47:07 | 25,200 | 0d19b7c082c555ceb6fa963791a23cf27eb1c33a | Mirror the minor spelling fixes over the {.ipynb, .md, .py}. | [
{
"change_type": "MODIFY",
"old_path": "docs/autodidax.ipynb",
"new_path": "docs/autodidax.ipynb",
"diff": "\"interpreter will build a jaxpr on the fly while tracking data dependencies. To\\n\",\n\"do so, it builds a bipartite directed acyclic graph (DAG) between\\n\",\n\"`PartialEvalTracer` nodes, ... | Python | Apache License 2.0 | google/jax | Mirror the minor spelling fixes over the {.ipynb, .md, .py}. |
260,335 | 28.03.2021 10:32:02 | 25,200 | 8547c71bfd26f96d00ddb8bbe825d90014e2adf5 | simplify public lax.convert_element_type api
Specifically:
1. don't expose weak_type in the public api, as it's jax-internal
2. don't make new_dtype optional, which could make bugs easier
This change keeps the public API simpler, and also makes
convert_element_type match the ConvertElementType HLO. As an internal
API we can call lax._convert_element_type just like before. | [
{
"change_type": "MODIFY",
"old_path": "jax/_src/lax/lax.py",
"new_path": "jax/_src/lax/lax.py",
"diff": "@@ -409,9 +409,9 @@ def lt(x: Array, y: Array) -> Array:\nr\"\"\"Elementwise less-than: :math:`x < y`.\"\"\"\nreturn lt_p.bind(x, y)\n-def convert_element_type(operand: Array, new_dtype: DType =... | Python | Apache License 2.0 | google/jax | simplify public lax.convert_element_type api
Specifically:
1. don't expose weak_type in the public api, as it's jax-internal
2. don't make new_dtype optional, which could make bugs easier
This change keeps the public API simpler, and also makes
convert_element_type match the ConvertElementType HLO. As an internal
API we can call lax._convert_element_type just like before. |
260,335 | 28.03.2021 11:17:41 | 25,200 | 4253c929ec06060a5150362d885979cdc1f97323 | use correct batching util in custom_vjp_call_jaxpr
fixes | [
{
"change_type": "MODIFY",
"old_path": "jax/custom_derivatives.py",
"new_path": "jax/custom_derivatives.py",
"diff": "@@ -654,7 +654,7 @@ def _custom_vjp_call_jaxpr_vmap(\nfwd_args_batched = [0 if b else not_mapped for b in args_batched]\nfwd_out_dims = lambda: out_dims2[0]\n- batched_bwd = batching... | Python | Apache License 2.0 | google/jax | use correct batching util in custom_vjp_call_jaxpr
fixes #5832 |
260,411 | 29.03.2021 14:15:27 | -10,800 | 5fce79ed441f40b73d29fd95ec312bc9ed3fa912 | Update fft_test.py
Trivial change to re-trigger the presubmit checks | [
{
"change_type": "MODIFY",
"old_path": "tests/fft_test.py",
"new_path": "tests/fft_test.py",
"diff": "@@ -119,7 +119,6 @@ class FftTest(jtu.JaxTestCase):\ndef testIrfftTranspose(self):\n# regression test for https://github.com/google/jax/issues/6223\n-\ndef build_matrix(linear_func, size):\nreturn j... | Python | Apache License 2.0 | google/jax | Update fft_test.py
Trivial change to re-trigger the presubmit checks |
260,411 | 29.03.2021 14:20:15 | -10,800 | 514709963d93cc4a0424d6b354b12d49c367ff63 | [jax2tf] Fixed typos, and updated the jax2tf documentation | [
{
"change_type": "MODIFY",
"old_path": "docs/developer.md",
"new_path": "docs/developer.md",
"diff": "@@ -239,7 +239,7 @@ $ jupytext --sync docs/notebooks/*\n```\nAlternatively, you can run this command via the [pre-commit](https://pre-commit.com/)\n-framework by executing the folloing in the main J... | Python | Apache License 2.0 | google/jax | [jax2tf] Fixed typos, and updated the jax2tf documentation |
260,411 | 29.03.2021 15:03:38 | -10,800 | e0a87ef062378a5526aafd86c66eb78f5daacf88 | [jax2tf] Update limitations due to updates in TF | [
{
"change_type": "MODIFY",
"old_path": "jax/experimental/jax2tf/g3doc/jax_primitives_coverage.md",
"new_path": "jax/experimental/jax2tf/g3doc/jax_primitives_coverage.md",
"diff": "# Primitives with limited JAX support\n-*Last generated on: 2021-03-23* (YYYY-MM-DD)\n+*Last generated on: 2021-03-29* (... | Python | Apache License 2.0 | google/jax | [jax2tf] Update limitations due to updates in TF |
260,335 | 29.03.2021 16:18:29 | 25,200 | 399f330f5a5d4bf0d30eb7dce7ef53c8a812d824 | disable_omnistaging error, enable_omnistaging warn | [
{
"change_type": "MODIFY",
"old_path": "jax/config.py",
"new_path": "jax/config.py",
"diff": "@@ -61,8 +61,7 @@ class Config:\nself._contextmanager_flags = set()\nself._update_hooks = {}\n- # TODO(mattjj): delete these when only omnistaging is available\n- self.omnistaging_enabled = bool_env('JAX_OM... | Python | Apache License 2.0 | google/jax | disable_omnistaging error, enable_omnistaging warn |
260,411 | 29.03.2021 17:21:56 | -10,800 | d323ad0f2b44a3178e02867992ba7243d7e0f70c | [host_callback] Add support for tapping empty arrays
We make sure that both the inputs and the outputs of
callbacks can contain empty arrays.
Most platforms do not support empty infeed, so we ensure
we do not send those. | [
{
"change_type": "MODIFY",
"old_path": "CHANGELOG.md",
"new_path": "CHANGELOG.md",
"diff": "@@ -25,7 +25,8 @@ PLEASE REMEMBER TO CHANGE THE '..master' WITH AN ACTUAL TAG in GITHUB LINK.\nfor more information.\n* Python integers larger than the maximum `int64` value will now lead to an overflow\nin a... | Python | Apache License 2.0 | google/jax | [host_callback] Add support for tapping empty arrays
We make sure that both the inputs and the outputs of
callbacks can contain empty arrays.
Most platforms do not support empty infeed, so we ensure
we do not send those. |
260,424 | 23.03.2021 19:47:58 | 0 | be70820ca147ca3e9a2df7d53a1c24a8e9899908 | Thread through transform name. | [
{
"change_type": "MODIFY",
"old_path": "jax/_src/lax/control_flow.py",
"new_path": "jax/_src/lax/control_flow.py",
"diff": "@@ -63,14 +63,18 @@ T = TypeVar('T')\nArray = Any\n@cache()\n-def _initial_style_open_jaxpr(fun: Callable, in_tree, in_avals):\n+def _initial_style_open_jaxpr(fun: Callable, in... | Python | Apache License 2.0 | google/jax | Thread through transform name. |
260,335 | 28.03.2021 19:58:01 | 25,200 | aa2472db0c6f14b737710d6f5ee6dfc42be808eb | add scatter_add jet rule, fixes
could use a better test though... | [
{
"change_type": "MODIFY",
"old_path": "jax/_src/lax/lax.py",
"new_path": "jax/_src/lax/lax.py",
"diff": "@@ -4555,6 +4555,7 @@ def _scatter_add_jvp(primals, tangents, *, update_jaxpr, update_consts,\ndimension_numbers, indices_are_sorted, unique_indices):\noperand, scatter_indices, updates = primal... | Python | Apache License 2.0 | google/jax | add scatter_add jet rule, fixes #5365
could use a better test though... |
260,335 | 31.03.2021 13:50:34 | 25,200 | 632876d773b5451ac83bb3a05bdffb164a7036cc | Copybara import of the project:
by Matthew Johnson
remove deprecated custom_transforms code | [
{
"change_type": "MODIFY",
"old_path": "jax/__init__.py",
"new_path": "jax/__init__.py",
"diff": "@@ -43,12 +43,7 @@ from .api import (\ncustom_gradient,\ncustom_jvp,\ncustom_vjp,\n- custom_transforms,\ndefault_backend,\n- defjvp,\n- defjvp_all,\n- defvjp,\n- defvjp_all,\ndevice_count,\ndevice_get,\... | Python | Apache License 2.0 | google/jax | Copybara import of the project:
--
35fcf2e2fd5b4c56cbb591f4c8bf01222a23dfe5 by Matthew Johnson <mattjj@google.com>:
remove deprecated custom_transforms code
PiperOrigin-RevId: 366108489 |
260,335 | 02.04.2021 12:52:37 | 25,200 | d4af8cd08afbbd2929d4834cfe6100d825e47b6e | improve djax dim indexing type checking | [
{
"change_type": "MODIFY",
"old_path": "jax/experimental/djax.py",
"new_path": "jax/experimental/djax.py",
"diff": "@@ -261,10 +261,16 @@ def typecheck_type(env, aval):\nif not (isinstance(d.name.aval, AbsArray) and\nisinstance(d.name.aval._eltTy, BoundedIntTy)):\nraise TypeError(f'dim var of unexpe... | Python | Apache License 2.0 | google/jax | improve djax dim indexing type checking
Co-authored-by: Edward Loper <edloper@google.com> |
260,335 | 02.04.2021 13:25:53 | 25,200 | 9205b6f1258f419a1b8256c737f53a25582cba19 | remove dead lax._dot_using_sum_of_products | [
{
"change_type": "MODIFY",
"old_path": "jax/_src/lax/lax.py",
"new_path": "jax/_src/lax/lax.py",
"diff": "@@ -693,8 +693,8 @@ def dot_general(lhs: Array, rhs: Array, dimension_numbers: DotDimensionNumbers,\nAn array containing the result.\n\"\"\"\ncontract_dims_seq, batch_dims_seq = dimension_number... | Python | Apache License 2.0 | google/jax | remove dead lax._dot_using_sum_of_products |
260,709 | 31.03.2021 18:35:15 | 25,200 | a19098d4627e352ceee0a0c534e847709a334b72 | Reimplement as JAX Primitive | [
{
"change_type": "MODIFY",
"old_path": "build/BUILD.bazel",
"new_path": "build/BUILD.bazel",
"diff": "@@ -39,6 +39,7 @@ py_binary(\n]) + if_cuda([\n\"//jaxlib:cublas_kernels\",\n\"//jaxlib:cusolver_kernels\",\n+ \"//jaxlib:cuda_lu_pivot_kernels\",\n\"//jaxlib:cuda_prng_kernels\",\n]) + if_rocm([\n\"... | Python | Apache License 2.0 | google/jax | Reimplement as JAX Primitive |
260,411 | 05.04.2021 15:56:38 | -10,800 | 6176ac1cb606e0fcbbb711f3999930f0d7270142 | [jax2tf] Fix bug in dot_general.
The case when there were batch dimensions but RHS has only
one inner dimmension was handled incorrectly. Add test also. | [
{
"change_type": "MODIFY",
"old_path": "jax/experimental/jax2tf/jax2tf.py",
"new_path": "jax/experimental/jax2tf/jax2tf.py",
"diff": "@@ -1262,7 +1262,7 @@ def _dot_general(lhs, rhs, dimension_numbers, precision, preferred_element_type)\ndel precision\ndel preferred_element_type\n(lhs_contracting, r... | Python | Apache License 2.0 | google/jax | [jax2tf] Fix bug in dot_general.
The case when there were batch dimensions but RHS has only
one inner dimmension was handled incorrectly. Add test also. |
260,411 | 05.04.2021 19:29:41 | -10,800 | 4e2e30da0cdae166fb981d0752bd7b2e5c47b5b3 | Increase tolerange for dot_general | [
{
"change_type": "MODIFY",
"old_path": "jax/experimental/jax2tf/tests/jax2tf_limitations.py",
"new_path": "jax/experimental/jax2tf/tests/jax2tf_limitations.py",
"diff": "@@ -464,6 +464,9 @@ class Jax2TfLimitation(primitive_harness.Limitation):\ncustom_numeric(\ndtypes=[np.complex64, np.float32], dev... | Python | Apache License 2.0 | google/jax | Increase tolerange for dot_general |
260,416 | 06.04.2021 17:08:33 | -7,200 | 73bc03794c9b9531ffccaf0efcb294f055c8ec6b | Fix jnp.flip for axis tuples | [
{
"change_type": "MODIFY",
"old_path": "jax/_src/numpy/lax_numpy.py",
"new_path": "jax/_src/numpy/lax_numpy.py",
"diff": "@@ -1111,7 +1111,8 @@ def flip(m, axis: Optional[Union[int, Tuple[int, ...]]] = None):\n_check_arraylike(\"flip\", m)\nif axis is None:\nreturn lax.rev(m, list(range(len(shape(m)... | Python | Apache License 2.0 | google/jax | Fix jnp.flip for axis tuples |
260,335 | 02.04.2021 13:23:43 | 25,200 | abf2d69262c581813bd8da4728b669a5ecd7cd10 | djax: add analogue of lower_fun, dot | [
{
"change_type": "MODIFY",
"old_path": "jax/_src/lax/lax.py",
"new_path": "jax/_src/lax/lax.py",
"diff": "@@ -3174,11 +3174,15 @@ def _dot_general_shape_rule(lhs, rhs, *, dimension_numbers, precision,\n\"shape, got {} and {}.\")\nraise TypeError(msg.format(lhs_contracting_shape, rhs_contracting_shap... | Python | Apache License 2.0 | google/jax | djax: add analogue of lower_fun, dot |
260,709 | 06.04.2021 13:38:36 | 25,200 | e186b9e591d4b6d3bda2bda841ba6d4335aee888 | Mark lu_pivots_to_permutation as "not yet implemented" in JAX2TF. | [
{
"change_type": "MODIFY",
"old_path": "jax/experimental/jax2tf/jax2tf.py",
"new_path": "jax/experimental/jax2tf/jax2tf.py",
"diff": "@@ -800,6 +800,7 @@ tf_not_yet_impl = [\n\"infeed\", \"outfeed\", \"pmax_p\",\n\"pmin\", \"ppermute\", \"psum\", \"pmax\", \"pgather\",\n\"axis_index\", \"pdot\", \"a... | Python | Apache License 2.0 | google/jax | Mark lu_pivots_to_permutation as "not yet implemented" in JAX2TF. |
260,411 | 07.04.2021 11:24:31 | -10,800 | dce31e9631f5cba7d4cc292072697dfb173c3fb6 | [jax2tf] Fix handling of float0 | [
{
"change_type": "MODIFY",
"old_path": "CHANGELOG.md",
"new_path": "CHANGELOG.md",
"diff": "@@ -9,9 +9,13 @@ PLEASE REMEMBER TO CHANGE THE '..master' WITH AN ACTUAL TAG in GITHUB LINK.\n-->\n## jax 0.2.13 (unreleased)\n+* [GitHub commits](https://github.com/google/jax/compare/jax-v0.2.12...master).\... | Python | Apache License 2.0 | google/jax | [jax2tf] Fix handling of float0 |
260,287 | 25.03.2021 11:31:55 | 0 | ba8430605dbd583210854c0bc0c8bbddf6034048 | Fix lax.all_gather inside xmap
The batching rule didn't properly handle tupled axis names. | [
{
"change_type": "MODIFY",
"old_path": "jax/_src/lax/parallel.py",
"new_path": "jax/_src/lax/parallel.py",
"diff": "@@ -1119,7 +1119,11 @@ def _all_gather_batcher(vals_in, dims_in, *, all_gather_dimension, axis_name, ax\ndef _all_gather_batched_collective(frame, vals_in, dims_in, all_gather_dimensio... | Python | Apache License 2.0 | google/jax | Fix lax.all_gather inside xmap
The batching rule didn't properly handle tupled axis names. |
260,411 | 04.04.2021 16:23:24 | -10,800 | 56e41b7cd76aa85f1e1b7c2a8d39efd818a05987 | Add support for cummax | [
{
"change_type": "MODIFY",
"old_path": "jax/_src/lax/lax.py",
"new_path": "jax/_src/lax/lax.py",
"diff": "@@ -5432,10 +5432,9 @@ def reduce_window_shape_tuple(operand_shape, window_dimensions, window_strides,\noperand_shape = _dilate_shape(operand_shape, base_dilation)\nif window_dilation is not Non... | Python | Apache License 2.0 | google/jax | Add support for cummax |
260,411 | 04.04.2021 17:05:18 | -10,800 | 4f9ac031d74c871be2f95c38cc3d605d2282a161 | Add some support for convolutions | [
{
"change_type": "MODIFY",
"old_path": "jax/_src/lax/lax.py",
"new_path": "jax/_src/lax/lax.py",
"diff": "@@ -2786,7 +2786,7 @@ def _conv_general_dilated_shape_rule(\n\"must be a positive integer, got {}.\")\nraise ValueError(msg.format(batch_group_count))\nlhs_batch_count = lhs.shape[dimension_numb... | Python | Apache License 2.0 | google/jax | Add some support for convolutions |
260,411 | 05.04.2021 10:13:02 | -10,800 | 3ae73a91a4c7e4b219a878fcb2a34cb4d7a14858 | Improve harness selection for shape_poly_test | [
{
"change_type": "MODIFY",
"old_path": "jax/experimental/jax2tf/jax2tf.py",
"new_path": "jax/experimental/jax2tf/jax2tf.py",
"diff": "@@ -1943,7 +1943,6 @@ def _sort(*operands: TfVal, dimension: int, is_stable: bool,\nif not _enable_xla:\nraise _xla_path_disabled_error(\"sort\")\nassert 1 <= num_key... | Python | Apache License 2.0 | google/jax | Improve harness selection for shape_poly_test |
260,411 | 05.04.2021 11:08:46 | -10,800 | cbe5f54ccaf8a8242dba62446b298613979c9e60 | Added support for lax.pad, and more error checking | [
{
"change_type": "MODIFY",
"old_path": "jax/_src/lax/lax.py",
"new_path": "jax/_src/lax/lax.py",
"diff": "@@ -3487,15 +3487,22 @@ def _pad_dtype_rule(operand, padding_value, *, padding_config):\ndef _pad_shape_rule(operand, padding_value, *, padding_config):\ndel padding_value\n+ op_shape = np.shape... | Python | Apache License 2.0 | google/jax | Added support for lax.pad, and more error checking |
260,411 | 05.04.2021 15:00:15 | -10,800 | 99d5f09b29dfbe7da8ea87b90274746a7ec22677 | Fix select and eigh | [
{
"change_type": "MODIFY",
"old_path": "jax/_src/lax/lax.py",
"new_path": "jax/_src/lax/lax.py",
"diff": "@@ -3812,7 +3812,7 @@ def _select_batch_rule(batched_args, batch_dims, **unused_kwargs):\nreturn select(pred, on_true, on_false), ot_bdim\npred = batching.bdim_at_front(pred, pred_bdim, size) if... | Python | Apache License 2.0 | google/jax | Fix select and eigh |
260,411 | 05.04.2021 15:56:30 | -10,800 | 0386ee38ff83d70354b7aefc6548ef5eadc21ede | Add more tests for dot_general | [
{
"change_type": "MODIFY",
"old_path": "jax/experimental/jax2tf/tests/shape_poly_test.py",
"new_path": "jax/experimental/jax2tf/tests/shape_poly_test.py",
"diff": "@@ -505,6 +505,9 @@ _VMAP_NOT_POLY_YET = {\n# In the random._gamma_impl we do reshape(-1, 2) for the keys\n\"random_gamma\",\n+ # In lin... | Python | Apache License 2.0 | google/jax | Add more tests for dot_general |
260,411 | 05.04.2021 19:23:30 | -10,800 | 14737e365eb07b82fc9738ae5fcf49e316f1c995 | Rewrite for python 3.8 | [
{
"change_type": "MODIFY",
"old_path": "jax/_src/numpy/lax_numpy.py",
"new_path": "jax/_src/numpy/lax_numpy.py",
"diff": "@@ -1314,7 +1314,8 @@ def _compute_newshape(a, newshape):\nexcept: iterable = False\nelse: iterable = True\nnewshape = core.canonicalize_shape(newshape if iterable else [newshape... | Python | Apache License 2.0 | google/jax | Rewrite for python 3.8 |
260,411 | 06.04.2021 11:43:06 | -10,800 | d9468c7513b60d35003f4032b2ab7449ad2242ed | Cleanup the API, and more documentation | [
{
"change_type": "MODIFY",
"old_path": "docs/developer.md",
"new_path": "docs/developer.md",
"diff": "@@ -129,7 +129,7 @@ sets up symbolic links from site-packages into the repository.\nTo run all the JAX tests, we recommend using `pytest-xdist`, which can run tests in\nparallel. First, install `pyt... | Python | Apache License 2.0 | google/jax | Cleanup the API, and more documentation |
260,631 | 08.04.2021 10:42:25 | 25,200 | 438b56c4830d09e1bdf95f4ab7dc12195560f417 | Fix typo in rng_bit_generator comment. | [
{
"change_type": "MODIFY",
"old_path": "jax/_src/lax/lax.py",
"new_path": "jax/_src/lax/lax.py",
"diff": "@@ -6183,7 +6183,7 @@ def rng_bit_generator(key,\n\"\"\"Stateless PRNG bit generator. Experimental and its use is discouraged.\nReturns uniformly distributed random bits with the specified shape... | Python | Apache License 2.0 | google/jax | Fix typo in rng_bit_generator comment.
PiperOrigin-RevId: 367460802 |
260,303 | 08.04.2021 18:24:49 | 25,200 | e3cb192e1d63c91b5a4e4fbd248accb37773e467 | Make mesh thread-safe | [
{
"change_type": "MODIFY",
"old_path": "jax/experimental/maps.py",
"new_path": "jax/experimental/maps.py",
"diff": "@@ -149,7 +149,7 @@ def mesh(devices: np.ndarray, axis_names: Sequence[ResourceAxisName]):\nout_axes=['left', 'right', ...],\naxis_resources={'left': 'x', 'right': 'y'})(x, x.T)\n\"\"\... | Python | Apache License 2.0 | google/jax | Make mesh thread-safe |
260,411 | 09.04.2021 11:10:32 | -10,800 | 7667fc3be71e36ba779177fc0f889e014ebcfc87 | [jax2tf] Added support for shape-polymorphic reductionsxs | [
{
"change_type": "MODIFY",
"old_path": "jax/_src/numpy/lax_numpy.py",
"new_path": "jax/_src/numpy/lax_numpy.py",
"diff": "@@ -1975,7 +1975,7 @@ def _reduction(a, name, np_fun, op, init_val, has_identity=True,\naxis = core.concrete_or_error(None, axis, f\"axis argument to jnp.{name}().\")\nif initial... | Python | Apache License 2.0 | google/jax | [jax2tf] Added support for shape-polymorphic reductionsxs |
260,411 | 09.04.2021 13:46:28 | -10,800 | 8815425e369412dc71d6b888556d045caddd0f55 | Cleanup of the dispatch to shape polymorphic dimension handlers | [
{
"change_type": "MODIFY",
"old_path": "jax/core.py",
"new_path": "jax/core.py",
"diff": "@@ -1226,7 +1226,11 @@ class DimensionHandler:\nDimension sizes are normally integer constants, but can also be symbolic,\ne.g., masking.Poly or jax2tf.shape_poly.DimVar.\n- The base class works for integers on... | Python | Apache License 2.0 | google/jax | Cleanup of the dispatch to shape polymorphic dimension handlers |
260,335 | 09.04.2021 14:43:13 | 25,200 | 60828e9b19eacdc20b6f62975c58ea143bc10b6e | raise error if vmap/pmap in_axes are booleans
fixes | [
{
"change_type": "MODIFY",
"old_path": "jax/api.py",
"new_path": "jax/api.py",
"diff": "@@ -1259,14 +1259,12 @@ def vmap(fun: F, in_axes=0, out_axes=0, axis_name=None) -> F:\n# rather than raising an error. https://github.com/google/jax/issues/2367\nin_axes = tuple(in_axes)\n- in_axes_, out_axes_ = ... | Python | Apache License 2.0 | google/jax | raise error if vmap/pmap in_axes are booleans
fixes #6372 |
260,335 | 09.04.2021 18:10:20 | 25,200 | 5f6bce4bfe5bdf35019845e2ad9a3cef4b82ca64 | add 'open in colab' button, add numpy<1.18 compat | [
{
"change_type": "MODIFY",
"old_path": "docs/autodidax.ipynb",
"new_path": "docs/autodidax.ipynb",
"diff": "]\n},\n{\n- \"cell_type\": \"code\",\n- \"execution_count\": null,\n+ \"cell_type\": \"markdown\",\n\"metadata\": {\n- \"lines_to_next_cell\": 0\n+ \"lines_to_next_cell\": 2\n},\n- \"outputs\"... | Python | Apache License 2.0 | google/jax | add 'open in colab' button, add numpy<1.18 compat
Co-authored-by: Edward Loper <edloper@google.com> |
260,496 | 09.04.2021 16:31:46 | -7,200 | 9e3001b9b342247ce8758ba3b81da6907dea33b4 | Some updates to jax2tf | [
{
"change_type": "MODIFY",
"old_path": "jax/experimental/jax2tf/examples/tf_js/quickdraw/README.md",
"new_path": "jax/experimental/jax2tf/examples/tf_js/quickdraw/README.md",
"diff": "@@ -12,24 +12,28 @@ and is a part of the code featured in the aforementioned blog post.\n## Training the model\nYou ... | Python | Apache License 2.0 | google/jax | Some updates to jax2tf |
260,424 | 12.04.2021 12:59:12 | -3,600 | b4f66d267694d055f2fbfdbbb06da87ccfe28428 | Fix handling of ad.Zero in _select_and_scatter_add_transpose.
Fixes | [
{
"change_type": "MODIFY",
"old_path": "jax/_src/lax/lax.py",
"new_path": "jax/_src/lax/lax.py",
"diff": "@@ -5573,6 +5573,8 @@ def _select_and_scatter_add_transpose(\nt, source, operand, *, select_prim, window_dimensions, window_strides,\npadding):\nassert ad.is_undefined_primal(source) and not ad.... | Python | Apache License 2.0 | google/jax | Fix handling of ad.Zero in _select_and_scatter_add_transpose.
Fixes #6403. |
260,510 | 12.04.2021 14:10:11 | 25,200 | 577c6011916f0e9f2ff51788bc765e07cb323d1b | Update references in custom interpreter notebook | [
{
"change_type": "MODIFY",
"old_path": "docs/notebooks/Writing_custom_interpreters_in_Jax.ipynb",
"new_path": "docs/notebooks/Writing_custom_interpreters_in_Jax.ipynb",
"diff": "},\n\"outputs\": [],\n\"source\": [\n- \"def examine_jaxpr(typed_jaxpr):\\n\",\n- \" jaxpr = typed_jaxpr.jaxpr\\n\",\n+ \"... | Python | Apache License 2.0 | google/jax | Update references in custom interpreter notebook |
260,287 | 14.04.2021 14:01:28 | 0 | c13efc12234df22ef19e8b6e5f41cecfbc1c37eb | Add JVP of xmap | [
{
"change_type": "MODIFY",
"old_path": "jax/experimental/maps.py",
"new_path": "jax/experimental/maps.py",
"diff": "@@ -35,6 +35,7 @@ from ..interpreters import partial_eval as pe\nfrom ..interpreters import pxla\nfrom ..interpreters import xla\nfrom ..interpreters import batching\n+from ..interpret... | Python | Apache License 2.0 | google/jax | Add JVP of xmap |
260,411 | 15.04.2021 09:50:00 | -10,800 | 3c6f3e50373a17be481886db68a4a875eafc0f3c | [jax2tf] Added documentation for shape polymorphism | [
{
"change_type": "MODIFY",
"old_path": "jax/experimental/jax2tf/README.md",
"new_path": "jax/experimental/jax2tf/README.md",
"diff": "@@ -40,6 +40,8 @@ More involved examples, including using jax2tf with\nFlax models and their use with TensorFlow Hub and Keras, are described in the\n[examples direct... | Python | Apache License 2.0 | google/jax | [jax2tf] Added documentation for shape polymorphism |
260,411 | 15.04.2021 09:57:46 | -10,800 | 29929a8b39916d2ddb82aac1b007e44e9eba5d2f | Update quickdraw.py
Minor cosmetic changes. The main reason I did this was to trigger another CI run. | [
{
"change_type": "MODIFY",
"old_path": "jax/experimental/jax2tf/examples/tf_js/quickdraw/quickdraw.py",
"new_path": "jax/experimental/jax2tf/examples/tf_js/quickdraw/quickdraw.py",
"diff": "@@ -44,9 +44,9 @@ flags.DEFINE_integer(\"num_epochs\", 5,\n(\"Number of epochs to train for.\"))\nflags.DEFINE... | Python | Apache License 2.0 | google/jax | Update quickdraw.py
Minor cosmetic changes. The main reason I did this was to trigger another CI run. |
260,287 | 12.04.2021 12:49:35 | 0 | 2d95d5ad2b5aba1fcefe077eb808f03d4a160311 | Small updates to abstract eval rules (AWN related)
I've been reading the AWN-related PRs and have found a few places that
could be improved a little. | [
{
"change_type": "MODIFY",
"old_path": "jax/_src/lax/parallel.py",
"new_path": "jax/_src/lax/parallel.py",
"diff": "@@ -629,7 +629,9 @@ def _allreduce_abstract_eval(*args, axes, axis_index_groups):\nnamed_shapes = [{name: size for name, size in arg.named_shape.items()\nif name not in named_axes} for... | Python | Apache License 2.0 | google/jax | Small updates to abstract eval rules (AWN related)
I've been reading the AWN-related PRs and have found a few places that
could be improved a little. |
260,411 | 15.04.2021 15:48:00 | -10,800 | eff49fe106fa399eac0dd848a84c1927e8a76877 | [jax2tf] Upgraded examples and tests to allow shape polymorphism | [
{
"change_type": "MODIFY",
"old_path": "jax/experimental/jax2tf/examples/README.md",
"new_path": "jax/experimental/jax2tf/examples/README.md",
"diff": "@@ -68,7 +68,7 @@ one using Flax (`FlaxMNIST`). Other Flax models can be arranged similarly,\nand the same strategy should work for other neural-net... | Python | Apache License 2.0 | google/jax | [jax2tf] Upgraded examples and tests to allow shape polymorphism |
260,287 | 15.04.2021 09:01:49 | 25,200 | 3c003a68fcc2ecea575a58bae033b951fc0e08a5 | Split up the HLO compilation from HLO generation in mesh_callable. | [
{
"change_type": "MODIFY",
"old_path": "jax/interpreters/pxla.py",
"new_path": "jax/interpreters/pxla.py",
"diff": "@@ -1430,7 +1430,20 @@ def mesh_callable(fun: lu.WrappedFun,\n# TODO: Warn about unused donations?\nbuilt = c.Build(out_tuple)\n- # 4. Compile the HLO\n+ return compile_and_wrap_mesh_h... | Python | Apache License 2.0 | google/jax | Split up the HLO compilation from HLO generation in mesh_callable.
PiperOrigin-RevId: 368649553 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.