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
260,335
30.06.2020 22:19:16
25,200
107689e91f0940327a807e2a9890da3071e6feac
improve vmap axis spec structure mismatch errors * improve vmap axis spec structure mismatch errors fixes * deflake
[ { "change_type": "MODIFY", "old_path": "jax/api.py", "new_path": "jax/api.py", "diff": "@@ -867,10 +867,11 @@ def vmap(fun: Callable, in_axes=0, out_axes=0) -> Callable:\nargs_flat, in_tree = tree_flatten(args)\nf = lu.wrap_init(fun)\nflat_fun, out_tree = flatten_fun_nokwargs(f, in_tree)\n- in_axes_...
Python
Apache License 2.0
google/jax
improve vmap axis spec structure mismatch errors (#3619) * improve vmap axis spec structure mismatch errors fixes #3613 * deflake
260,335
01.07.2020 11:26:44
25,200
ba1b5ce8de01985133ce0cf960c112ffb81dd7fb
skip some ode tests on gpu for speed
[ { "change_type": "MODIFY", "old_path": "jax/test_util.py", "new_path": "jax/test_util.py", "diff": "@@ -55,8 +55,7 @@ flags.DEFINE_integer(\nflags.DEFINE_bool(\n'jax_skip_slow_tests',\nbool_env('JAX_SKIP_SLOW_TESTS', False),\n- help=\n- 'Skip tests marked as slow (> 5 sec).'\n+ help='Skip tests mark...
Python
Apache License 2.0
google/jax
skip some ode tests on gpu for speed (#3629)
260,335
01.07.2020 12:22:39
25,200
1bd04e2d2f3ed0cbbe4dae8bfec09651930db66c
skip gamma tests on tpu for compilation speed
[ { "change_type": "MODIFY", "old_path": "tests/random_test.py", "new_path": "tests/random_test.py", "diff": "@@ -402,6 +402,7 @@ class LaxRandomTest(jtu.JaxTestCase):\nnp.array([0.2, 1., 5.]),\n]\nfor dtype in [np.float32, np.float64]))\n+ @jtu.skip_on_devices(\"tpu\") # TODO(mattjj): slow compilatio...
Python
Apache License 2.0
google/jax
skip gamma tests on tpu for compilation speed (#3631)
260,335
01.07.2020 14:15:48
25,200
65c4d755def093ab9c4388d04b10a6ff0f3178f6
fix bug in categorical test, disable on tpu * fix bug in categorical test, disable on tpu Disabling on TPU pending a TPU compilation bug. * unskip a test
[ { "change_type": "MODIFY", "old_path": "jax/lax/lax.py", "new_path": "jax/lax/lax.py", "diff": "@@ -4363,6 +4363,10 @@ ad.defjvp_zero(argmin_p)\nxla.backend_specific_translations['gpu'][argmin_p] = xla.lower_fun(\npartial(_argminmax_gpu_translation_rule, _reduce_min),\nmultiple_results=False)\n+# TO...
Python
Apache License 2.0
google/jax
fix bug in categorical test, disable #3611 on tpu (#3633) * fix bug in categorical test, disable #3611 on tpu Disabling #3611 on TPU pending a TPU compilation bug. * unskip a test
260,411
02.07.2020 10:21:23
-10,800
448c635a9ea00f32b2c20c277a49f62ea9029f3b
[jax2tf] Update the tf-nightly version to 2020-07-01
[ { "change_type": "MODIFY", "old_path": ".github/workflows/ci-build.yaml", "new_path": ".github/workflows/ci-build.yaml", "diff": "@@ -52,6 +52,7 @@ jobs:\n- python-version: 3.6\nos: ubuntu-latest\nenable-x64: enable-x64\n+ # Test with numpy version that matches Google-internal version\npackage-overr...
Python
Apache License 2.0
google/jax
[jax2tf] Update the tf-nightly version to 2020-07-01 (#3635)
260,411
02.07.2020 10:33:00
-10,800
a5ed161550825f3b3c52b764b39db8e15b4c9f1e
Release jaxlib 0.1.51
[ { "change_type": "MODIFY", "old_path": "WORKSPACE", "new_path": "WORKSPACE", "diff": "@@ -28,10 +28,10 @@ http_archive(\n# and update the sha256 with the result.\nhttp_archive(\nname = \"org_tensorflow\",\n- sha256 = \"3e2387ae5d069de6fe1ccd4faa8d0dc1d63799d59df10f5bb93d03152ba9b158\",\n- strip_pref...
Python
Apache License 2.0
google/jax
Release jaxlib 0.1.51 (#3636)
260,411
02.07.2020 12:47:22
-10,800
bfe8f8732b9b90a3a0e7d79c40964b759833bdea
Change the version of tensorflow, the one in jaxlib 0.1.51 does not build
[ { "change_type": "MODIFY", "old_path": "WORKSPACE", "new_path": "WORKSPACE", "diff": "@@ -28,10 +28,10 @@ http_archive(\n# and update the sha256 with the result.\nhttp_archive(\nname = \"org_tensorflow\",\n- sha256 = \"ab1b8edb072269a4a513385fbf03a6dc2a22cda5f7cb833aa4962e692d3caf61\",\n- strip_pref...
Python
Apache License 2.0
google/jax
Change the version of tensorflow, the one in jaxlib 0.1.51 does not build (#3637)
260,411
02.07.2020 18:53:58
-10,800
166e795d638349530accacb1a48261203b29ae67
Updated minimum jaxlib to 0.1.51
[ { "change_type": "MODIFY", "old_path": "README.md", "new_path": "README.md", "diff": "@@ -410,7 +410,7 @@ PYTHON_VERSION=cp37 # alternatives: cp36, cp37, cp38\nCUDA_VERSION=cuda100 # alternatives: cuda100, cuda101, cuda102, cuda110\nPLATFORM=manylinux2010_x86_64 # alternatives: manylinux2010_x86_64\...
Python
Apache License 2.0
google/jax
Updated minimum jaxlib to 0.1.51 (#3642)
260,335
02.07.2020 14:29:17
25,200
d10cf0e38f9281e16e7e84fa4fe289f6c5ae2654
fix prng key reuse in differential privacy example fix prng key reuse in differential privacy example
[ { "change_type": "MODIFY", "old_path": "examples/differentially_private_sgd.py", "new_path": "examples/differentially_private_sgd.py", "diff": "@@ -64,7 +64,6 @@ Example invocations:\n--learning_rate=.25 \\\n\"\"\"\n-from functools import partial\nimport itertools\nimport time\nimport warnings\n@@ -...
Python
Apache License 2.0
google/jax
fix prng key reuse in differential privacy example (#3646) fix prng key reuse in differential privacy example
260,335
03.07.2020 10:00:25
25,200
796df9c550618c15dfec157af306b08fe58eac13
make psum transpose handle zero cotangents make psum transpose handle zero cotangents fixes
[ { "change_type": "MODIFY", "old_path": "jax/lax/lax_parallel.py", "new_path": "jax/lax/lax_parallel.py", "diff": "@@ -364,6 +364,12 @@ def _notuple_psum_translation_rule(c, *args, replica_groups):\nreturn psum(val)\nreturn xops.Tuple(c, list(map(_translate, args)))\n+def _psum_transpose_rule(cts, ax...
Python
Apache License 2.0
google/jax
make psum transpose handle zero cotangents (#3653) make psum transpose handle zero cotangents fixes #3651
260,335
03.07.2020 20:54:25
25,200
49cfe2687c9a23d355c799a01a931280e44e6018
improve concreteness error message for nn.one_hot * improve nn.one_hot and jax.numpy.arange errors fixes * deflake * debug
[ { "change_type": "MODIFY", "old_path": "jax/core.py", "new_path": "jax/core.py", "diff": "@@ -831,17 +831,15 @@ def concretization_function_error(fun, context=\"\"):\nreturn error\n-def concrete_or_error(typ: Type, val: Any, context=\"\"):\n- \"\"\"Like typ(val), but gives the context in the error m...
Python
Apache License 2.0
google/jax
improve concreteness error message for nn.one_hot (#3656) * improve nn.one_hot and jax.numpy.arange errors fixes #3654 * deflake * debug
260,411
06.07.2020 09:04:02
-10,800
8f93607330b36f1bfdb8e030657b6d73452d0cb3
Fix broken links to deleted notebook Fixes
[ { "change_type": "MODIFY", "old_path": "jax/tree_util.py", "new_path": "jax/tree_util.py", "diff": "@@ -31,7 +31,7 @@ The primary purpose of this module is to enable the interoperability between\nuser defined data structures and JAX transformations (e.g. `jit`). This is not\nmeant to be a general pu...
Python
Apache License 2.0
google/jax
Fix broken links to deleted notebook (#3663) Fixes #3662.
260,335
06.07.2020 15:11:01
25,200
e5ba5f1e6707fde12b3d95de0bd71da8a2dd453b
support multiple_results and custom JVPs in jet
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jet.py", "new_path": "jax/experimental/jet.py", "diff": "@@ -26,6 +26,7 @@ from jax.tree_util import (register_pytree_node, tree_structure,\ntreedef_is_leaf, tree_flatten, tree_unflatten)\nimport jax.linear_util as lu\nfrom jax.interpreters im...
Python
Apache License 2.0
google/jax
support multiple_results and custom JVPs in jet (#3657)
260,335
07.07.2020 00:30:08
25,200
d2ebb6eb19b0a824303e22a492d4b72f15ba1d88
fix ppermute test bugs found by
[ { "change_type": "MODIFY", "old_path": "jax/lax/lax_parallel.py", "new_path": "jax/lax/lax_parallel.py", "diff": "@@ -196,29 +196,26 @@ def ppermute(x, axis_name, perm):\npartial(ppermute_p.bind, axis_name=axis_name, perm=tuple(perm)), x)\ndef pshuffle(x, axis_name, perm):\n- \"\"\"Perform a collect...
Python
Apache License 2.0
google/jax
fix ppermute test bugs found by @jekbradbury (#3675)
260,335
07.07.2020 14:48:54
25,200
1034f29de778b288c32271cd283fb44e9b5adaf4
fix bad pmap tests from
[ { "change_type": "MODIFY", "old_path": "tests/pmap_test.py", "new_path": "tests/pmap_test.py", "diff": "@@ -564,15 +564,10 @@ class PmapTest(jtu.JaxTestCase):\nx = np.arange(device_count, dtype=np.float32)\n- ans = g(x)\n- expected = np.roll(np.pi + np.arange(device_count), 1)\n- self.assertAllClose...
Python
Apache License 2.0
google/jax
fix bad pmap tests from #3675 (#3685)
260,411
08.07.2020 16:08:54
-10,800
fdd7f0c8574b1e0e4366ee428153d73bc1788510
Added concurrent id_tap tests, disabled for GPU
[ { "change_type": "MODIFY", "old_path": "tests/host_callback_test.py", "new_path": "tests/host_callback_test.py", "diff": "@@ -336,6 +336,77 @@ where: 3\nhcb.barrier_wait() # Wait for receivers to be done\nself.assertEqual(3, tap_count)\n+ @parameterized.named_parameters(\n+ jtu.cases_from_list(\n+ d...
Python
Apache License 2.0
google/jax
Added concurrent id_tap tests, disabled for GPU (#3690)
260,651
10.07.2020 20:34:59
-10,800
412b9d5209619c6cc6f5af7c6926cab24b919c3e
hfft and ihfft implementation
[ { "change_type": "MODIFY", "old_path": "docs/jax.numpy.rst", "new_path": "docs/jax.numpy.rst", "diff": "@@ -312,6 +312,8 @@ jax.numpy.fft\nirfft2\nrfftn\nirfftn\n+ hfft\n+ ihfft\nfftfreq\nrfftfreq\nfftshift\n" }, { "change_type": "MODIFY", "old_path": "jax/numpy/fft.py", "new_path": ...
Python
Apache License 2.0
google/jax
hfft and ihfft implementation (#3664)
260,335
11.07.2020 20:47:22
25,200
51ca57d5fc02abe486cecc1f1a5478bd2bf6a100
check matmul inputs aren't scalar also dot_general shape rule should check dimension numbers are in range fixes
[ { "change_type": "MODIFY", "old_path": "jax/lax/lax.py", "new_path": "jax/lax/lax.py", "diff": "@@ -2657,6 +2657,20 @@ def _precision_config(precision):\ndef _dot_general_shape_rule(lhs, rhs, *, dimension_numbers, precision):\n(lhs_contracting, rhs_contracting), (lhs_batch, rhs_batch) = dimension_nu...
Python
Apache License 2.0
google/jax
check matmul inputs aren't scalar (#3725) also dot_general shape rule should check dimension numbers are in range fixes #3718
260,651
14.07.2020 01:31:47
-10,800
a6ab742f3635ed7c6d8c499c4be8aeb2a4181922
Improve np.intersect1d
[ { "change_type": "MODIFY", "old_path": "jax/numpy/lax_numpy.py", "new_path": "jax/numpy/lax_numpy.py", "diff": "@@ -1249,10 +1249,9 @@ def _intersect1d_sorted_mask(ar1, ar2, return_indices=False):\nHelper function for intersect1d which is jit-able\n\"\"\"\nar = concatenate((ar1, ar2))\n-\nif return_...
Python
Apache License 2.0
google/jax
Improve np.intersect1d (#3739)
260,500
14.07.2020 18:37:09
-3,600
f6f97554f9c9689caccae8ff383a51620f0151f8
A jit-able version of np.repeat. A new keyword argument has been added to np.repeat, total_repeat_length, that can optionally be supplied to make np.repeat jit-able.
[ { "change_type": "MODIFY", "old_path": "jax/numpy/lax_numpy.py", "new_path": "jax/numpy/lax_numpy.py", "diff": "@@ -2505,72 +2505,61 @@ def indices(dimensions, dtype=int32, sparse=False):\nreturn stack(output, 0) if output else array([], dtype=dtype)\n-def _repeat_scalar(a, repeats, axis=None):\n- i...
Python
Apache License 2.0
google/jax
A jit-able version of np.repeat. (#3670) A new keyword argument has been added to np.repeat, total_repeat_length, that can optionally be supplied to make np.repeat jit-able.
260,308
15.07.2020 00:45:49
-7,200
f02d5b4694ae62a5a05a193d89e825bda2b79737
Support differentiation through jax.lax.all_to_all * Support differentiation through jax.lax.all_to_all Credit to for the solution. * Test gradient of all_to_all We are testing all_to_all through pswapaxes, since general all_to_all is problematic according to * Removed trailing spaces
[ { "change_type": "MODIFY", "old_path": "jax/lax/lax_parallel.py", "new_path": "jax/lax/lax_parallel.py", "diff": "@@ -441,6 +441,9 @@ def _all_to_all_split_axis_rule(vals, which_mapped, split_axis, concat_axis,\nout = _moveaxis(1, concat_axis + 1, out)\nreturn out, True\n+def _all_to_all_transpose_r...
Python
Apache License 2.0
google/jax
Support differentiation through jax.lax.all_to_all (#3733) * Support differentiation through jax.lax.all_to_all Credit to @levskaya for the solution. * Test gradient of all_to_all We are testing all_to_all through pswapaxes, since general all_to_all is problematic according to https://github.com/google/jax/issues/1332. * Removed trailing spaces
260,298
15.07.2020 06:56:19
-7,200
68c8dc781e9623afa01f5d0ec9cdde93c7e2bc6d
Added installation note for jax2tf with GPU support.
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/README.md", "new_path": "jax/experimental/jax2tf/README.md", "diff": "@@ -77,3 +77,10 @@ tf.saved_model.save(my_model, '/some/directory')\n# Restoring (note: the restored model does *not* require JAX to run, just XLA).\nrestored_model =...
Python
Apache License 2.0
google/jax
Added installation note for jax2tf with GPU support. (#3750)
260,298
15.07.2020 14:12:42
-7,200
d55c47d2cfc031b1340cca776ece33faf54c8060
[jax2tf] Fix error in sorting with TF graphs.
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/jax2tf.py", "new_path": "jax/experimental/jax2tf/jax2tf.py", "diff": "@@ -1104,9 +1104,9 @@ def _sort(*operand: TfVal, dimension: int, is_stable: bool, num_keys: int) -> Tu\nraise NotImplementedError(\"TODO: implement stable version of ...
Python
Apache License 2.0
google/jax
[jax2tf] Fix error in sorting with TF graphs. (#3764)
260,411
15.07.2020 18:19:22
-10,800
23c279f033a4fc6b836f4c9f12d82a8bcc1dda17
[jax2tf] Relax tolerance for JaxPrimitiveTest.test_unary_elementwise_lgamma_float32
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/tests/primitives_test.py", "new_path": "jax/experimental/jax2tf/tests/primitives_test.py", "diff": "@@ -171,7 +171,12 @@ class JaxPrimitiveTest(tf_test_util.JaxToTfTestCase):\n# non-special cases are equal\nself.assertAllClose(result_ja...
Python
Apache License 2.0
google/jax
[jax2tf] Relax tolerance for JaxPrimitiveTest.test_unary_elementwise_lgamma_float32 (#3767)
260,266
15.07.2020 19:19:40
-3,600
150d028d9d381b47ff58b9b71dc7167ef3f6dc7f
Update scipy.ndimage.map_coordinates docstring
[ { "change_type": "MODIFY", "old_path": "jax/scipy/ndimage.py", "new_path": "jax/scipy/ndimage.py", "diff": "@@ -111,10 +111,10 @@ def _map_coordinates(input, coordinates, order, mode, cval):\n@_wraps(scipy.ndimage.map_coordinates, lax_description=textwrap.dedent(\"\"\"\\\n- Only linear interpolation...
Python
Apache License 2.0
google/jax
Update scipy.ndimage.map_coordinates docstring (#3762)
260,298
16.07.2020 15:44:20
-7,200
7b18a4e857f0a11f6203fe8fff7a4d87f22ce107
[jax2tf] Fix interface of ConvertAndCompare function.
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/tests/tf_test_util.py", "new_path": "jax/experimental/jax2tf/tests/tf_test_util.py", "diff": "@@ -62,7 +62,6 @@ class JaxToTfTestCase(jtu.JaxTestCase):\ndef ConvertAndCompare(self, func_jax: Callable, *args,\ncustom_assert: Optional[Cal...
Python
Apache License 2.0
google/jax
[jax2tf] Fix interface of ConvertAndCompare function. (#3776)
260,355
18.07.2020 21:36:23
-3,600
d43393405abe30b2da2561bd29b4d8b08a6063bf
Add a TypeVar to while_loop definition.
[ { "change_type": "MODIFY", "old_path": "jax/lax/lax_control_flow.py", "new_path": "jax/lax/lax_control_flow.py", "diff": "@@ -22,7 +22,7 @@ import functools\nimport inspect\nimport itertools\nimport operator\n-from typing import Callable, Sequence\n+from typing import Callable, Sequence, TypeVar\nim...
Python
Apache License 2.0
google/jax
Add a TypeVar to while_loop definition. (#3792)
260,424
20.07.2020 14:59:13
-7,200
dd3cb8213507f5f57f23b554862f7e703abb7c6e
Enable buffer donation for GPU.
[ { "change_type": "MODIFY", "old_path": "jax/interpreters/pxla.py", "new_path": "jax/interpreters/pxla.py", "diff": "@@ -795,7 +795,7 @@ def parallel_callable(fun, backend, axis_name, axis_size, global_axis_size,\nelse:\nout_tuple = build_out_tuple()\nbackend = xb.get_backend(backend)\n- if backend.p...
Python
Apache License 2.0
google/jax
Enable buffer donation for GPU. (#3800)
260,335
21.07.2020 06:48:55
25,200
3a3c8ea8f21332bce8bd44bdd62deec059e0fe62
tweak jnp.repeat not to use jnp on shapes tweak jnp.repeat not to use jnp on shapes
[ { "change_type": "MODIFY", "old_path": "jax/numpy/lax_numpy.py", "new_path": "jax/numpy/lax_numpy.py", "diff": "@@ -2543,16 +2543,18 @@ def repeat(a, repeats, axis=None, *, total_repeat_length=None):\na = ravel(a)\naxis = 0\n- repeats = array(repeats)\n+ # If total_repeat_length is not given, can't ...
Python
Apache License 2.0
google/jax
tweak jnp.repeat not to use jnp on shapes (#3810) tweak jnp.repeat not to use jnp on shapes
260,476
21.07.2020 15:41:08
14,400
74d363e552ca916d48eb116366eca3bb7eccf23f
fix extremely minor typo "ijnputs" -> "inputs"
[ { "change_type": "MODIFY", "old_path": "jax/random.py", "new_path": "jax/random.py", "diff": "@@ -280,7 +280,7 @@ def fold_in(key, data):\ndata: a 32bit integer representing data to be folded in to the key.\nReturns:\n- A new PRNGKey that is a deterministic function of the ijnputs and is\n+ A new PR...
Python
Apache License 2.0
google/jax
fix extremely minor typo (#3815) "ijnputs" -> "inputs"
260,335
23.07.2020 19:38:56
25,200
67ad5eb5a1bf4db3c61a28d7f7233f971e4c1263
add result_shape option to xla_computation add result_shape option to xla_computation
[ { "change_type": "MODIFY", "old_path": "jax/api.py", "new_path": "jax/api.py", "diff": "@@ -228,7 +228,8 @@ def xla_computation(fun: Callable,\naxis_env: Optional[Sequence[Tuple[AxisName, int]]] = None,\nbackend: Optional[str] = None,\ntuple_args: bool = False,\n- instantiate_const_outputs: bool = T...
Python
Apache License 2.0
google/jax
add result_shape option to xla_computation (#3844) add result_shape option to xla_computation
260,335
23.07.2020 19:49:04
25,200
cc9528d97dba5c88c15c089a4bac069d7407a61b
fix thread locality bug in custom_derivatives * fix thread locality bug in custom_derivatives fixes
[ { "change_type": "MODIFY", "old_path": "jax/core.py", "new_path": "jax/core.py", "diff": "@@ -635,6 +635,11 @@ class TraceState(threading.local):\nself.substack = [Sublevel(0)]\nself.initial_style = False\n+ def set_state(self, other: 'TraceState') -> None:\n+ self.trace_stack = other.trace_stack\n+...
Python
Apache License 2.0
google/jax
fix thread locality bug in custom_derivatives (#3845) * fix thread locality bug in custom_derivatives fixes #3843
260,335
23.07.2020 20:59:12
25,200
e2424e3b242a521dc02eaa11228861ddc354b41b
attempt to fix CI failure (from test?)
[ { "change_type": "MODIFY", "old_path": "tests/api_test.py", "new_path": "tests/api_test.py", "diff": "@@ -2534,13 +2534,17 @@ class CustomJVPTest(jtu.JaxTestCase):\ndef run():\nreturn jax.grad(unroll)(jnp.array(1.0), jnp.array([1.0]))\n+ expected = run()\n+\n# we just don't want this to crash\nn_wor...
Python
Apache License 2.0
google/jax
attempt to fix CI failure (from #3845 test?) (#3846)
260,411
24.07.2020 10:14:08
-10,800
86732620c9c996c60ffd2666e18e56b3142c437d
[jax2tf] Expand the support for jax.remat. In the simplest forms, remat is already handled by the `process_call`. But when the `remat` has outputs computed from values captured from an outer environment, we need to also implement `post_process_call`.
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/jax2tf.py", "new_path": "jax/experimental/jax2tf/jax2tf.py", "diff": "@@ -305,8 +305,17 @@ class TensorFlowTrace(core.Trace):\nvals_out: Sequence[TfValOrUnit] = f.call_wrapped(*vals)\nreturn [TensorFlowTracer(self, v) for v in vals_out]...
Python
Apache License 2.0
google/jax
[jax2tf] Expand the support for jax.remat. (#3828) In the simplest forms, remat is already handled by the `process_call`. But when the `remat` has outputs computed from values captured from an outer environment, we need to also implement `post_process_call`.
260,411
24.07.2020 10:14:34
-10,800
55cd8279c857456a0797639899259906dce570cf
[jax2tf] Selectively disable some primitive tests on TPU. Currently, all primitive tests are disabled on TPU, but only some cases fail. This PR simply disables the selected failing tests on TPU, so that we can enable the rest by default. Future PRs will address each failure individually.
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/tests/primitives_test.py", "new_path": "jax/experimental/jax2tf/tests/primitives_test.py", "diff": "@@ -119,6 +119,9 @@ class JaxPrimitiveTest(tf_test_util.JaxToTfTestCase):\nelif harness.params[\"dtype\"] is dtypes.bfloat16:\nraise uni...
Python
Apache License 2.0
google/jax
[jax2tf] Selectively disable some primitive tests on TPU. (#3835) Currently, all primitive tests are disabled on TPU, but only some cases fail. This PR simply disables the selected failing tests on TPU, so that we can enable the rest by default. Future PRs will address each failure individually.
260,335
24.07.2020 12:52:52
25,200
57c2fcb0cdb58768790fed13e6dd171bba235d90
tweak parallel collective translation rule api also remove standard_parallel_primitive helper function, which wasn't very helpful
[ { "change_type": "MODIFY", "old_path": "jax/interpreters/pxla.py", "new_path": "jax/interpreters/pxla.py", "diff": "@@ -944,7 +944,7 @@ def get_num_partitions(*partitions):\nreturn num_partitions_set.pop()\n-class ResultToPopulate(object): pass\n+class ResultToPopulate: pass\nresult_to_populate = Re...
Python
Apache License 2.0
google/jax
tweak parallel collective translation rule api (#3853) also remove standard_parallel_primitive helper function, which wasn't very helpful
260,335
24.07.2020 15:54:21
25,200
5e91965723ac16dda361435d1bfcd4dc99201537
delete standard_parallel_primitive helper delete standard_parallel_primitive helper meant to include this in it's even in the PR message!
[ { "change_type": "MODIFY", "old_path": "jax/lax/__init__.py", "new_path": "jax/lax/__init__.py", "diff": "@@ -334,5 +334,4 @@ from .lax_parallel import (\npsum,\npsum_p,\npswapaxes,\n- standard_pmap_primitive,\n)\n" }, { "change_type": "MODIFY", "old_path": "jax/lax/lax_parallel.py", ...
Python
Apache License 2.0
google/jax
delete standard_parallel_primitive helper (#3858) delete standard_parallel_primitive helper meant to include this in #3853, it's even in the PR message!
260,335
26.07.2020 22:38:14
25,200
c9d8acd2e921efdbc01b2208fa9f790a9c312933
put core trace state in a threading.local class this is a refinement of the fix in so that we no longer need TraceState.set_state (and so that is easier to adapt)
[ { "change_type": "MODIFY", "old_path": "jax/core.py", "new_path": "jax/core.py", "diff": "@@ -39,7 +39,7 @@ from .pprint_util import pp, vcat, PrettyPrint\n# TODO(dougalm): compilation cache breaks the leak detector. Consisder solving.\ncheck_leaks = False\n-\"\"\"Disables internal invariant checks....
Python
Apache License 2.0
google/jax
put core trace state in a threading.local class (#3869) this is a refinement of the fix in #3845, so that we no longer need TraceState.set_state (and so that #3370 is easier to adapt)
260,326
28.07.2020 02:45:36
-19,080
3aa37d3af4324f53c72270296ca905189ab6fd99
Replicating sort_complex functionality from np.sort_complex to jax.numpy
[ { "change_type": "MODIFY", "old_path": "docs/jax.numpy.rst", "new_path": "docs/jax.numpy.rst", "diff": "@@ -255,6 +255,7 @@ Not every function in NumPy is implemented; contributions are welcome!\nsinh\nsometrue\nsort\n+ sort_complex\nsplit\nsqrt\nsquare\n" }, { "change_type": "MODIFY", "...
Python
Apache License 2.0
google/jax
Replicating sort_complex functionality from np.sort_complex to jax.numpy (#3870)
260,335
27.07.2020 21:51:12
25,200
616d63b19ae08c6d58f1de8d7c41fe78ebded157
fix vmap error, fixes
[ { "change_type": "MODIFY", "old_path": "jax/api.py", "new_path": "jax/api.py", "diff": "@@ -907,7 +907,7 @@ def _mapped_axis_size(tree, vals, dims, name):\n# TODO(mattjj,phawkins): add a way to inspect pytree kind more directly\nif tree == tree_flatten((core.unit,) * tree.num_leaves)[1]:\nlines1 = [...
Python
Apache License 2.0
google/jax
fix vmap error, fixes #3877 (#3879)
260,299
28.07.2020 16:27:07
-7,200
7506a3e5f0137f48415def878bea71c5977dfbd0
Fix flaky generated_fun_test.py test
[ { "change_type": "MODIFY", "old_path": "tests/generated_fun_test.py", "new_path": "tests/generated_fun_test.py", "diff": "@@ -186,7 +186,7 @@ def inner_prod(xs, ys):\nreturn sum(jnp.sum(x * y) for x, y in xys)\ndef jvp_fd(fun, args, tangents):\n- EPS = 1e-4\n+ EPS = 1e-3\ndef eval_eps(eps):\nreturn ...
Python
Apache License 2.0
google/jax
Fix flaky generated_fun_test.py test (#3885)
260,335
28.07.2020 19:46:00
25,200
30980742c57f3b507d74b5ae79e9a75b5f57bc08
refine population_count type check * refine population_count type check fixes * allow signed/unsigned ints for population_count https://cs.opensource.google/tensorflow/tensorflow/+/master:tensorflow/compiler/xla/service/shape_inference.cc;l=314?q=xla%20f:shape_inference.cc * make lax_reference.population_count handle signed
[ { "change_type": "MODIFY", "old_path": "jax/lax/lax.py", "new_path": "jax/lax/lax.py", "diff": "@@ -2242,7 +2242,7 @@ ad.defjvp_zero(or_p)\nxor_p = standard_naryop([_bool_or_int, _bool_or_int], 'xor')\nad.defjvp_zero(xor_p)\n-population_count_p = standard_unop(_bool_or_int, 'population_count')\n+pop...
Python
Apache License 2.0
google/jax
refine population_count type check (#3887) * refine population_count type check fixes #3886 * allow signed/unsigned ints for population_count https://cs.opensource.google/tensorflow/tensorflow/+/master:tensorflow/compiler/xla/service/shape_inference.cc;l=314?q=xla%20f:shape_inference.cc * make lax_reference.population_count handle signed
260,335
30.07.2020 00:42:32
25,200
b3806ce87438ecbeb52ffa5ee7755adaf21eef2b
add placeholder enable_omnistaging method
[ { "change_type": "MODIFY", "old_path": "jax/config.py", "new_path": "jax/config.py", "diff": "@@ -113,6 +113,8 @@ class Config(object):\nself.complete_absl_config(absl.flags)\nalready_configured_with_absl = True\n+ def enable_omnistaging(self):\n+ pass # placeholder\nclass NameSpace(object):\ndef __...
Python
Apache License 2.0
google/jax
add placeholder enable_omnistaging method (#3907)
260,335
30.07.2020 22:27:01
25,200
146cf49fa04ce8e476da6981f52688b63bef7f4e
delay backend check for xla_computation
[ { "change_type": "MODIFY", "old_path": "jax/interpreters/xla.py", "new_path": "jax/interpreters/xla.py", "diff": "@@ -369,7 +369,10 @@ def jaxpr_literals(jaxpr):\ndef jaxpr_subcomp(c, jaxpr, backend, axis_env, consts, name_stack, *args):\n- platform = xb.get_backend(backend).platform\n+ if backend n...
Python
Apache License 2.0
google/jax
delay backend check for xla_computation (#3920)
260,335
31.07.2020 15:11:01
25,200
843d710116138fba67728fafb53a4eb667cbf55c
allow mask to return output logical shape When the `mask` argument `out_shape` is not provided, or when it has value `None`, return the output logical shape to the user.
[ { "change_type": "MODIFY", "old_path": "jax/api.py", "new_path": "jax/api.py", "diff": "@@ -1241,7 +1241,7 @@ def _papply(fun):\nreturn papply_fun, axis_name\n-def mask(fun: Callable, in_shapes, out_shape) -> Callable:\n+def mask(fun: Callable, in_shapes, out_shape=None) -> Callable:\n_check_callabl...
Python
Apache License 2.0
google/jax
allow mask to return output logical shape (#3929) When the `mask` argument `out_shape` is not provided, or when it has value `None`, return the output logical shape to the user.
260,335
31.07.2020 22:20:58
25,200
ff96de935b186609da473cf6f769901a17f77aa9
add dummy eval context
[ { "change_type": "MODIFY", "old_path": "jax/core.py", "new_path": "jax/core.py", "diff": "@@ -733,6 +733,10 @@ def initial_style_staging():\nfinally:\ntrace_state.initial_style = prev\n+@contextmanager\n+def eval_context():\n+ yield # dummy implementation for forward compatibility\n+\n# ------------...
Python
Apache License 2.0
google/jax
add dummy eval context (#3932)
260,355
01.08.2020 21:33:11
-3,600
8a8bb702d29c8036807d53b6110fc2fe2051559a
Catch invalid (negative) in/out axes in vmap. Catch invalid (negative) in/out axes in vmap.
[ { "change_type": "MODIFY", "old_path": "jax/api.py", "new_path": "jax/api.py", "diff": "@@ -869,16 +869,22 @@ def vmap(fun: Callable, in_axes=0, out_axes=0) -> Callable:\n# rather than raising an error. https://github.com/google/jax/issues/2367\nin_axes = tuple(in_axes)\n- if not all(isinstance(l, (...
Python
Apache License 2.0
google/jax
Catch invalid (negative) in/out axes in vmap. (#3926) Catch invalid (negative) in/out axes in vmap.
260,285
03.08.2020 17:17:48
-7,200
7de784afbf719590583e362102a2bcc0fa5037a0
Fix _CheckAgainstNumpy arg order
[ { "change_type": "MODIFY", "old_path": "tests/lax_numpy_test.py", "new_path": "tests/lax_numpy_test.py", "diff": "@@ -2582,7 +2582,7 @@ class LaxBackedNumpyTests(jtu.JaxTestCase):\nargs_maker = lambda: [rng.randn(3, 4).astype(\"float32\")]\nnp_op = lambda x: np.asarray(x).astype(jnp.int32)\njnp_op =...
Python
Apache License 2.0
google/jax
Fix _CheckAgainstNumpy arg order (#3935)
260,298
04.08.2020 09:24:53
-7,200
ea6d1dafc2f698a831cfdbb7faf1f4dab47ed102
[jax2tf] First draft of conversion of population_count. tf.raw_ops.PopulationCount returns a Tensor with elements of type uint8. In JAX, the type of the result is the same as the type of the input, hence the need to use tf.cast in the conversion function.
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/jax2tf.py", "new_path": "jax/experimental/jax2tf/jax2tf.py", "diff": "@@ -364,7 +364,7 @@ for unexpected in [\n# Primitives that are not yet implemented must be explicitly declared here.\ntf_not_yet_impl = [\n- lax.population_count_p, l...
Python
Apache License 2.0
google/jax
[jax2tf] First draft of conversion of population_count. (#3890) tf.raw_ops.PopulationCount returns a Tensor with elements of type uint8. In JAX, the type of the result is the same as the type of the input, hence the need to use tf.cast in the conversion function.
260,287
04.08.2020 12:41:30
0
60db0a073747bd11d654e31a380f23ee9de95ab3
Improve a comment and a docstring in partial_eval
[ { "change_type": "MODIFY", "old_path": "jax/interpreters/partial_eval.py", "new_path": "jax/interpreters/partial_eval.py", "diff": "@@ -605,6 +605,12 @@ def partial_eval_jaxpr(jaxpr: TypedJaxpr, unknowns: Sequence[bool],\n`jaxpr_known` = lambda ki, ui: let ka = ki + 2\nin (ki + 3, *, ka)\n'jaxpr_unk...
Python
Apache License 2.0
google/jax
Improve a comment and a docstring in partial_eval
260,285
06.08.2020 03:36:46
-7,200
2e873b7d050a09dd99881d9339353a18042cf7b3
Fix jnp.right_shift incorrect on unsigned ints
[ { "change_type": "MODIFY", "old_path": "jax/core.py", "new_path": "jax/core.py", "diff": "@@ -506,7 +506,9 @@ class Tracer:\ndef __rxor__(self, other): return self.aval._rxor(self, other)\ndef __invert__(self): return self.aval._invert(self)\ndef __lshift__(self, other): return self.aval._lshift(sel...
Python
Apache License 2.0
google/jax
Fix jnp.right_shift incorrect on unsigned ints (#3958)
260,299
06.08.2020 18:37:32
-7,200
09f4339c165f3885583144644b8fa9d33788e3b3
Rm two unused lines from lax_parallel.psum_bind
[ { "change_type": "MODIFY", "old_path": "jax/lax/lax_parallel.py", "new_path": "jax/lax/lax_parallel.py", "diff": "@@ -955,8 +955,6 @@ def omnistaging_enabler() -> None:\n# tracing time.\n@psum_p.def_custom_bind\ndef psum_bind(*args, axis_name, **params):\n- if len(args) == 1 and not isinstance(args[...
Python
Apache License 2.0
google/jax
Rm two unused lines from lax_parallel.psum_bind (#3978)
260,411
07.08.2020 13:30:26
-10,800
e39420b55d715a6e73aae0691936945975775451
[jax2tf] Another attempt to work around the nested XLA compilation bug * [jax2tf] Another attempt to work around the nested compilation bug for XLA This time we abandon trying to force compilation, and instead we use XlaDynamicSlice.
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/README.md", "new_path": "jax/experimental/jax2tf/README.md", "diff": "@@ -92,36 +92,6 @@ with both [jaxlib](../../../../../#pip-installation) and\n## Known limitations\n-#### Errors due to nested invocations of XLA\n-\n-In some rare cas...
Python
Apache License 2.0
google/jax
[jax2tf] Another attempt to work around the nested XLA compilation bug (#3965) * [jax2tf] Another attempt to work around the nested compilation bug for XLA This time we abandon trying to force compilation, and instead we use XlaDynamicSlice.
260,298
07.08.2020 17:22:17
-7,200
b756f29457a717ab47b791538dc029681e8f5c80
[jax2tf] implementation of select_and_gather_add conversion. * [jax2tf] First draft of select_and_gather_add conversion. The case where we need to reduce precision for packing is left unimplemented, as packing is anyway a temporary hack until XLA implements ReduceWindow on tuples (see b/73062247).
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/jax2tf.py", "new_path": "jax/experimental/jax2tf/jax2tf.py", "diff": "@@ -23,6 +23,7 @@ from jax import ad_util\nfrom jax import api\nfrom jax import core\nfrom jax import custom_derivatives\n+from jax import dtypes\nfrom jax import lax...
Python
Apache License 2.0
google/jax
[jax2tf] implementation of select_and_gather_add conversion. (#3987) * [jax2tf] First draft of select_and_gather_add conversion. The case where we need to reduce precision for packing is left unimplemented, as packing is anyway a temporary hack until XLA implements ReduceWindow on tuples (see b/73062247).
260,355
08.08.2020 17:22:54
-3,600
038c85dad0ca6d7b39ccee3f8838d29630f97e8f
Improve type annotations for `jit` and `vmap`.
[ { "change_type": "MODIFY", "old_path": "jax/api.py", "new_path": "jax/api.py", "diff": "@@ -29,7 +29,7 @@ import functools\nimport inspect\nimport itertools as it\nimport threading\n-from typing import Any, Callable, Iterable, Optional, Sequence, Tuple, Union\n+from typing import Any, Callable, Iter...
Python
Apache License 2.0
google/jax
Improve type annotations for `jit` and `vmap`. (#3938)
260,335
10.08.2020 07:20:09
25,200
d46ea969533dbfe4451517ac9a5e51dbdeee6d5d
add helper for flax to be omnistaging-compatible
[ { "change_type": "MODIFY", "old_path": "jax/interpreters/partial_eval.py", "new_path": "jax/interpreters/partial_eval.py", "diff": "@@ -1075,6 +1075,15 @@ def trace_to_jaxpr_final(fun: lu.WrappedFun, in_avals: Sequence[AbstractValue]):\ndel master\nreturn jaxpr, out_avals, consts\n+def partial_eval_...
Python
Apache License 2.0
google/jax
add helper for flax to be omnistaging-compatible (#4004)
260,299
10.08.2020 19:09:34
-7,200
89e8a0886b17385556074f1c31d5048981794bb8
Fix warnings in pmap_test.py Also add note to developer documentation re: testing pmap.
[ { "change_type": "MODIFY", "old_path": "docs/developer.rst", "new_path": "docs/developer.rst", "diff": "@@ -116,6 +116,14 @@ the tests of ``jax.numpy.pad`` using::\nThe Colab notebooks are tested for errors as part of the documentation build.\n+Note that to run the full pmap tests on a (multi-core) ...
Python
Apache License 2.0
google/jax
Fix warnings in pmap_test.py (#3977) Also add note to developer documentation re: testing pmap.
260,335
10.08.2020 18:11:57
25,200
6a3b920507dcae7a4e4dfa513155222fa0c6feb1
make make_jaxpr work on tracer example args (don't use xla.abstractify)
[ { "change_type": "MODIFY", "old_path": "jax/api.py", "new_path": "jax/api.py", "diff": "@@ -1625,7 +1625,7 @@ def make_jaxpr(fun: Callable,\nwrapped, _ = argnums_partial(wrapped, dyn_argnums, args)\njax_args, in_tree = tree_flatten((args, kwargs))\njaxtree_fun, out_tree = flatten_fun(wrapped, in_tre...
Python
Apache License 2.0
google/jax
make make_jaxpr work on tracer example args (#4014) (don't use xla.abstractify)
260,335
10.08.2020 20:48:03
25,200
09d8ac14de8c9edc2c9a14becb4963bfdebda605
use fewer internal APIs in custom interp 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+ \"nbformat\": 4,\n+ \"nbformat_minor\": 0,\n+ \"metadata\": {\n+ \"colab\": {\n+ \"name\": \"Writing custom interpret...
Python
Apache License 2.0
google/jax
use fewer internal APIs in custom interp notebook (#4016)
260,411
11.08.2020 12:39:54
-10,800
0b99ca896d8cfecd737af71f1a140b9956143d4c
[jax2tf] Disable the CI tests for jax2tf. We do this to see if this reduces the incidence of errors fetching the tf-nightly package. These tests are being run when we import the code in Google.
[ { "change_type": "MODIFY", "old_path": ".github/workflows/ci-build.yaml", "new_path": ".github/workflows/ci-build.yaml", "diff": "@@ -110,9 +110,6 @@ jobs:\necho \"JAX_NUM_GENERATED_CASES=$JAX_NUM_GENERATED_CASES\"\necho \"JAX_ENABLE_X64=$JAX_ENABLE_X64\"\necho \"JAX_OMNISTAGING=$JAX_OMNISTAGING\"\n...
Python
Apache License 2.0
google/jax
[jax2tf] Disable the CI tests for jax2tf. (#4019) We do this to see if this reduces the incidence of errors fetching the tf-nightly package. These tests are being run when we import the code in Google.
260,287
11.08.2020 11:45:58
-7,200
cd54dd977897f868c3f55723604054d43db801e8
Implement the invertible decorator in terms of custom_vjp This simplifies the implementation significantly, as we can piggyback off of all the logic for custom derivatives. For example, the previous implementation didn't support differentiating with respect to a subset of function parameters, but the new one does.
[ { "change_type": "MODIFY", "old_path": "jax/api.py", "new_path": "jax/api.py", "diff": "@@ -2044,7 +2044,7 @@ def custom_gradient(fun):\ndef _ensure_tuple(x: Union[int, Iterable[int]]) -> Tuple[int, ...]:\nreturn (x,) if isinstance(x, int) else tuple(x)\n-def invertible(fun: Callable, concrete: bool...
Python
Apache License 2.0
google/jax
Implement the invertible decorator in terms of custom_vjp (#3957) This simplifies the implementation significantly, as we can piggyback off of all the logic for custom derivatives. For example, the previous implementation didn't support differentiating with respect to a subset of function parameters, but the new one does.
260,299
11.08.2020 16:09:36
-7,200
1e07712955939d6f8f461fc259b12a20808782b3
Fix typos in api.py docstrings
[ { "change_type": "MODIFY", "old_path": "jax/api.py", "new_path": "jax/api.py", "diff": "@@ -118,7 +118,7 @@ def jit(fun: Callable[..., T],\nstatic_argnums: An int or collection of ints specifying which positional\narguments to treat as static (compile-time constant). Operations that only\ndepend on ...
Python
Apache License 2.0
google/jax
Fix typos in api.py docstrings (#4021)
260,335
11.08.2020 20:36:51
25,200
c564aca77710df0599715d4231b7d5b7dd46984a
skip more ode tests on gpu, b/c slow to compile
[ { "change_type": "MODIFY", "old_path": "tests/ode_test.py", "new_path": "tests/ode_test.py", "diff": "@@ -59,7 +59,7 @@ class ODETest(jtu.JaxTestCase):\njtu.check_grads(integrate, (y0, ts, *args), modes=[\"rev\"], order=2,\natol=tol, rtol=tol)\n- @jtu.skip_on_devices(\"tpu\")\n+ @jtu.skip_on_devices...
Python
Apache License 2.0
google/jax
skip more ode tests on gpu, b/c slow to compile (#4028)
260,411
12.08.2020 09:20:26
-10,800
ece05d2813942594cc904c306ac485b4846882bd
[host_callback] Add support for custom gradients inside control-flow * [host_callback] Add support for custom gradients inside control-flow Specifically, add support for rewriting JAXPR that contains custom_vjp_call_jaxpr_p and custom_jvp_call_jaxpr_p. Fixes
[ { "change_type": "MODIFY", "old_path": "jax/experimental/host_callback.py", "new_path": "jax/experimental/host_callback.py", "diff": "@@ -136,6 +136,7 @@ import itertools\nfrom jax import api\nfrom jax import core\n+from jax import custom_derivatives\nfrom jax import lax\nfrom jax.lib import pytree\...
Python
Apache License 2.0
google/jax
[host_callback] Add support for custom gradients inside control-flow (#4029) * [host_callback] Add support for custom gradients inside control-flow Specifically, add support for rewriting JAXPR that contains custom_vjp_call_jaxpr_p and custom_jvp_call_jaxpr_p. Fixes #4015.
260,701
12.08.2020 19:52:42
-3,600
6e4ec7cb81be0dc06dd0bf0ffe789abb8a2ae87a
Fix broadcasting in random.uniform and randint.
[ { "change_type": "MODIFY", "old_path": "jax/random.py", "new_path": "jax/random.py", "diff": "@@ -353,8 +353,8 @@ def uniform(key: jnp.ndarray,\nshape. Default ().\ndtype: optional, a float dtype for the returned values (default float64 if\njax_enable_x64 is true, otherwise float32).\n- minval: opti...
Python
Apache License 2.0
google/jax
Fix broadcasting in random.uniform and randint. (#4035)
260,411
13.08.2020 13:02:22
-10,800
206d2f9890dac4da6855933ae486f8a270d477dd
[host_callback] Fix handling of params[donated_invars] There was some confusion when rewriting xla_call primitives and the donated_invars was not set to be equal to the length of the invars. Fixes: Fixes:
[ { "change_type": "MODIFY", "old_path": "jax/experimental/host_callback.py", "new_path": "jax/experimental/host_callback.py", "diff": "@@ -654,7 +654,10 @@ def _rewrite_eqn(eqn: core.JaxprEqn, eqns: List[core.JaxprEqn],\ndict(\neqn.params,\ncall_jaxpr=_rewrite_jaxpr(call_jaxpr, True,\n- True)), eqn.s...
Python
Apache License 2.0
google/jax
[host_callback] Fix handling of params[donated_invars] (#4040) There was some confusion when rewriting xla_call primitives and the donated_invars was not set to be equal to the length of the invars. Fixes: #4015 Fixes: #3863
260,298
13.08.2020 12:27:16
-7,200
3458f1fc1643eb5b1f71c598a16baf7e6d739307
[jax2tf] Change np.array dtype to jnp.float32 when target dtype is bfloat16. * Change np.array dtype to jnp.float32 when target dtype is bfloat16. This ensures that the conversion for tf.convert_to_tensor is well defined without losing precision. Fixes
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/jax2tf.py", "new_path": "jax/experimental/jax2tf/jax2tf.py", "diff": "@@ -57,11 +57,39 @@ def _is_tfval(v: TfVal) -> bool:\nif isinstance(v, (tf.Tensor, tf.Variable)):\nreturn True\ntry:\n- tf.convert_to_tensor(v)\n+ # Note: this conver...
Python
Apache License 2.0
google/jax
[jax2tf] Change np.array dtype to jnp.float32 when target dtype is bfloat16. (#3959) * Change np.array dtype to jnp.float32 when target dtype is bfloat16. This ensures that the conversion for tf.convert_to_tensor is well defined without losing precision. Fixes #3942.
260,411
13.08.2020 15:20:07
-10,800
7cf236f83f0494f2ef0777f0ba450df20973dde2
[jax2tf] Clean up test_qr and disable (temporarily) some tests for shirt_right on TPU This is a follow-up from that enabled some more integer types for TPU for JAX. There are still some silent errors to investigate for shift_right on TPU.
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/tests/primitive_harness.py", "new_path": "jax/experimental/jax2tf/tests/primitive_harness.py", "diff": "@@ -413,7 +413,7 @@ lax_linalg_qr = tuple(\nshape=shape,\ndtype=dtype,\nfull_matrices=full_matrices)\n- for dtype in jtu.dtypes.all\...
Python
Apache License 2.0
google/jax
[jax2tf] Clean up test_qr and disable (temporarily) some tests for shirt_right (#4044) on TPU This is a follow-up from #4032 that enabled some more integer types for TPU for JAX. There are still some silent errors to investigate for shift_right on TPU.
260,298
14.08.2020 09:58:41
-7,200
41f1eb412c2522be34795cd47148f3e39405088c
[jax2tf] Added note about nextafter support in TensorFlow. The TF implementation of nextafter only supports float32/64 params.
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/tests/primitives_test.py", "new_path": "jax/experimental/jax2tf/tests/primitives_test.py", "diff": "@@ -361,9 +361,11 @@ class JaxPrimitiveTest(tf_test_util.JaxToTfTestCase):\nif dtype is np.float32 and jtu.device_under_test() == \"tpu\...
Python
Apache License 2.0
google/jax
[jax2tf] Added note about nextafter support in TensorFlow. (#4056) The TF implementation of nextafter only supports float32/64 params.
260,298
14.08.2020 12:24:57
-7,200
bd14f2343aece428df85808d23cace1ede0b2821
[jax2tf] Fix bfloat16 failures on CPU/GPU with latest tf-nightly.
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/tests/primitives_test.py", "new_path": "jax/experimental/jax2tf/tests/primitives_test.py", "diff": "@@ -282,8 +282,13 @@ class JaxPrimitiveTest(tf_test_util.JaxToTfTestCase):\ndef test_unary_elementwise(self, harness: primitive_harness....
Python
Apache License 2.0
google/jax
[jax2tf] Fix bfloat16 failures on CPU/GPU with latest tf-nightly. (#4060)
260,272
14.08.2020 21:05:46
-7,200
b44a1bacaf14833f44907dbd13399bf04d162fee
Note flush-denormal-to-zero behavior in nextafter. Since imports of functions from lax.py in lax_numpy.py previously didn't copy docstrings, add option to extract docstring from the lax function. To avoid duplication of the short description, discard first line of the lax function's docstring, and pass the rest as lax_description to _wraps.
[ { "change_type": "MODIFY", "old_path": "jax/lax/lax.py", "new_path": "jax/lax/lax.py", "diff": "@@ -113,7 +113,18 @@ def sign(x: Array) -> Array:\nreturn sign_p.bind(x)\ndef nextafter(x1: Array, x2: Array) -> Array:\n- r\"\"\"Returns the next representable value after `x1` in the direction of `x2`.\...
Python
Apache License 2.0
google/jax
Note flush-denormal-to-zero behavior in nextafter. (#4049) Since imports of functions from lax.py in lax_numpy.py previously didn't copy docstrings, add option to extract docstring from the lax function. To avoid duplication of the short description, discard first line of the lax function's docstring, and pass the rest as lax_description to _wraps.
260,588
15.08.2020 17:47:28
-7,200
1316562b0b96eb4ce10b0661bb113e1be3b9406c
Canonicalize result dtype to fix double precision problem in ldexp
[ { "change_type": "MODIFY", "old_path": "jax/numpy/lax_numpy.py", "new_path": "jax/numpy/lax_numpy.py", "diff": "@@ -757,7 +757,7 @@ _INT_DTYPES = {\n@_wraps(np.ldexp)\n@jit\ndef ldexp(x1, x2):\n- dtype = _result_dtype(np.ldexp, x1, x2)\n+ dtype = dtypes.canonicalize_dtype(_result_dtype(np.ldexp, x1,...
Python
Apache License 2.0
google/jax
Canonicalize result dtype to fix double precision problem in ldexp (#4069)
260,335
15.08.2020 22:55:18
25,200
8232f2deee5d1e0c78496699f35225b2f5d6226b
adapt _TempAxisName for unhashable objs adapt _TempAxisName for unhashable objs
[ { "change_type": "MODIFY", "old_path": "jax/api.py", "new_path": "jax/api.py", "diff": "@@ -364,7 +364,7 @@ def xla_computation(fun: Callable,\nreturn xla.AxisEnv(nreps, (), (), None)\nelse:\nnreps = nreps * prod(size for name, size in axis_env)\n- names, sizes = zip(*axis_env)\n+ names, sizes = unz...
Python
Apache License 2.0
google/jax
adapt _TempAxisName for unhashable objs (#4077) adapt _TempAxisName for unhashable objs
260,411
17.08.2020 09:50:47
-10,800
22b92c5122ab5af6f5e4560f9be08f5649ae7653
Increase tolerance for CPU test LaxBackedNumpyTests::testCorrCoef
[ { "change_type": "MODIFY", "old_path": "tests/lax_numpy_test.py", "new_path": "tests/lax_numpy_test.py", "diff": "@@ -3623,7 +3623,10 @@ class LaxBackedNumpyTests(jtu.JaxTestCase):\nself._CheckAgainstNumpy(\nnp_fun, jnp_fun, args_maker, check_dtypes=False,\ntol=1e-2 if jtu.device_under_test() == \"t...
Python
Apache License 2.0
google/jax
Increase tolerance for CPU test LaxBackedNumpyTests::testCorrCoef (#4080)
260,298
17.08.2020 12:57:41
-7,200
ec90c3587adb569491bf54c7b447844491925337
[jax2tf] Fix bfloat16 bug in select_and_gather_add conversion. * [jax2tf] Fix bfloat16 bug in select_and_gather_add conversion. This fix makes it possible to run bfloat16 tests for the jax2tf conversion of select_and_gather_add.
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/jax2tf.py", "new_path": "jax/experimental/jax2tf/jax2tf.py", "diff": "@@ -799,8 +799,8 @@ def _select_and_gather_add(tangents: TfVal,\npadding: Sequence[Tuple[int, int]]):\n# Note: this function follows the pattern in\n# jax.lax._select...
Python
Apache License 2.0
google/jax
[jax2tf] Fix bfloat16 bug in select_and_gather_add conversion. (#4058) * [jax2tf] Fix bfloat16 bug in select_and_gather_add conversion. This fix makes it possible to run bfloat16 tests for the jax2tf conversion of select_and_gather_add.
260,298
17.08.2020 16:32:34
-7,200
4c22e012d27ca865d429492f0e5956791907a981
[jax2tf] Explictly raise an error when attempting to convert _select_and_scatter_add_p.
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/jax2tf.py", "new_path": "jax/experimental/jax2tf/jax2tf.py", "diff": "@@ -395,7 +395,6 @@ for unexpected in [\n# Primitives that are not yet implemented must be explicitly declared here.\ntf_not_yet_impl = [\nlax.reduce_p, lax.reduce_wi...
Python
Apache License 2.0
google/jax
[jax2tf] Explictly raise an error when attempting to convert _select_and_scatter_add_p. (#4084)
260,287
17.08.2020 20:11:43
-7,200
1ba4e06c9101f5fcc3fa587e76b1247b72a5fa1b
Initial version of gmap Co-autored-by: Matthew Johnson
[ { "change_type": "ADD", "old_path": null, "new_path": "jax/experimental/general_map.py", "diff": "+# Copyright 2020 Google LLC\n+#\n+# Licensed under the Apache License, Version 2.0 (the \"License\");\n+# you may not use this file except in compliance with the License.\n+# You may obtain a copy of t...
Python
Apache License 2.0
google/jax
Initial version of gmap (#4006) Co-autored-by: Matthew Johnson <mattjj@google.com>
260,287
18.08.2020 12:02:28
-7,200
36f3a369c73f9d293b600dc97cbd3be4e31bf708
Separate axis splitting from collective handling This makes the vmap collective handling a bit more flexible and allowed me to add ppermute support.
[ { "change_type": "MODIFY", "old_path": "jax/interpreters/batching.py", "new_path": "jax/interpreters/batching.py", "diff": "@@ -133,23 +133,20 @@ class BatchTrace(Trace):\nif all(bdim is not_mapped for bdim in dims_in):\nreturn primitive.bind(*vals_in, **params)\nelif config.omnistaging_enabled and ...
Python
Apache License 2.0
google/jax
Separate axis splitting from collective handling (#4082) This makes the vmap collective handling a bit more flexible and allowed me to add ppermute support.
260,485
18.08.2020 16:40:45
14,400
29f7fa753a24225800a7d7ba64db6441ca6a4dac
Add implementation of jax.numpy.trim_zeros
[ { "change_type": "MODIFY", "old_path": "docs/jax.numpy.rst", "new_path": "docs/jax.numpy.rst", "diff": "@@ -342,6 +342,7 @@ Not every function in NumPy is implemented; contributions are welcome!\ntril\ntril_indices\ntril_indices_from\n+ trim_zeros\ntriu\ntriu_indices\ntriu_indices_from\n" }, { ...
Python
Apache License 2.0
google/jax
Add implementation of jax.numpy.trim_zeros (#4027)
260,268
18.08.2020 18:49:29
14,400
afeefa6f1fb8c5fdd296e5bedf4d4e1c1df70b45
Add typing and namedtuple to `optimizers.py`, improve documentation.
[ { "change_type": "MODIFY", "old_path": "jax/experimental/optimizers.py", "new_path": "jax/experimental/optimizers.py", "diff": "@@ -64,8 +64,21 @@ Notice that an optimizer implementation has a lot of flexibility in the form of\nopt_state: it just has to be a pytree of JaxTypes (so that it can be pas...
Python
Apache License 2.0
google/jax
Add typing and namedtuple to `optimizers.py`, improve documentation. (#3570)
260,321
19.08.2020 19:36:28
-3,600
b4efb31f1a9c38b6e635b1b096adc496866fb235
Docs: Fix broken link in quickstart
[ { "change_type": "MODIFY", "old_path": "docs/notebooks/quickstart.ipynb", "new_path": "docs/notebooks/quickstart.ipynb", "diff": "\"id\": \"Xpy1dSgNqCP4\"\n},\n\"source\": [\n- \"We'll be generating random data in the following examples. One big difference between NumPy and JAX is how you generate r...
Python
Apache License 2.0
google/jax
Docs: Fix broken link in quickstart (#4102)
260,335
19.08.2020 15:51:40
25,200
9ba282f321c7b68d0a3f5b70c3c352bcd43a4b91
add axis_index to supported multi-host collectives also make the error message less confusing
[ { "change_type": "MODIFY", "old_path": "jax/interpreters/pxla.py", "new_path": "jax/interpreters/pxla.py", "diff": "@@ -726,8 +726,9 @@ def parallel_callable(fun, backend, axis_name, axis_size, global_axis_size,\nif is_multi_host_pmap:\nused_collectives = set(xla.jaxpr_collectives(jaxpr))\nif not us...
Python
Apache License 2.0
google/jax
add axis_index to supported multi-host collectives (#4107) also make the error message less confusing
260,298
20.08.2020 09:05:30
-7,200
22b593b6b9888dafb5308bc8e689d2d272ab623e
[jax2tf] General conversion of reduce_window. * [jax2tf] General conversion of reduce_window. Much like scatter_p, the conversion works as well as the underlying reduction function (e.g. reduce_window_max is not properly converted for the int8 dtype, because tf.math.maximum is not defined for int8 tensors).
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/jax2tf.py", "new_path": "jax/experimental/jax2tf/jax2tf.py", "diff": "@@ -781,9 +781,6 @@ tf_impl[lax.argmax_p] = functools.partial(_argminmax, tf.math.argmax)\n_add_fn = tf.function(tf.math.add)\n_ge_fn = tf.function(tf.math.greater_eq...
Python
Apache License 2.0
google/jax
[jax2tf] General conversion of reduce_window. (#4093) * [jax2tf] General conversion of reduce_window. Much like scatter_p, the conversion works as well as the underlying reduction function (e.g. reduce_window_max is not properly converted for the int8 dtype, because tf.math.maximum is not defined for int8 tensors).
260,441
20.08.2020 16:46:55
-7,200
1e8ac248635a14249dd17269f8e165698603d6ea
Add rademacher, maxwell, double_sided_maxwell and weibull_min to jax.random.
[ { "change_type": "MODIFY", "old_path": "jax/random.py", "new_path": "jax/random.py", "diff": "@@ -1349,3 +1349,132 @@ def _t(key, df, shape, dtype):\nhalf_df = lax.div(df, two)\ng = gamma(key_n, half_df, shape, dtype)\nreturn n * jnp.sqrt(half_df / g)\n+\n+\n+def rademacher(key, shape, dtype=dtypes....
Python
Apache License 2.0
google/jax
Add rademacher, maxwell, double_sided_maxwell and weibull_min to jax.random. (#4104)
260,335
20.08.2020 14:44:26
25,200
306314590e0c4565178320b13acac40adaee1f8d
use xla.backend_compile function in pxla.py * use xla.backend_compile function in pxla.py Not only is this useful for profiling, it also helps us do google-internal logging for the XLA team.
[ { "change_type": "MODIFY", "old_path": "jax/api.py", "new_path": "jax/api.py", "diff": "@@ -302,12 +302,9 @@ def _cpp_jit(\noptions.parameter_is_tupled_arguments = True\nreturn (\n- # We call xla._backend_compile to ensure the XLA computation appears\n+ # We call xla.backend_compile to ensure the XL...
Python
Apache License 2.0
google/jax
use xla.backend_compile function in pxla.py (#4113) * use xla.backend_compile function in pxla.py Not only is this useful for profiling, it also helps us do google-internal logging for the XLA team.
260,335
21.08.2020 11:10:53
25,200
b2a239ca9e7ffa9008e5b1db6adcac66f2bf02f2
don't force backend creation in xla_computation
[ { "change_type": "MODIFY", "old_path": "jax/api.py", "new_path": "jax/api.py", "diff": "@@ -611,17 +611,13 @@ def _xla_computation(\nelse:\nout_tuple = build_out_tuple()\n- backend_ = xb.get_backend(backend)\n- if backend_.platform in (\"gpu\", \"tpu\"):\n- donated_invars = xla.set_up_aliases(c, xla...
Python
Apache License 2.0
google/jax
don't force backend creation in xla_computation
260,644
21.08.2020 20:19:51
-7,200
9d733dd7d7edb3660493654c7aa807e996f99a37
Doc: change suggested way of starting the profiler
[ { "change_type": "MODIFY", "old_path": "docs/profiling.md", "new_path": "docs/profiling.md", "diff": "@@ -37,11 +37,13 @@ from a running program.\n```python\nimport jax.profiler\n- jax.profiler.start_server(9999)\n+ server = jax.profiler.start_server(9999)\n```\nThis starts the profiler server that ...
Python
Apache License 2.0
google/jax
Doc: change suggested way of starting the profiler (#4120)
260,335
21.08.2020 17:43:15
25,200
66a02b69713a773a837b7f9160dacb6c9a6d397f
only construct one axis_index_p primitive Before this change, there were two versions, one used with omnistaging and one without. But that made bookkeeping hard and buggy. This change defines the axis_index_p primitive in core.py. Some of its rules are still changed when omnistaging is enabled.
[ { "change_type": "MODIFY", "old_path": "jax/core.py", "new_path": "jax/core.py", "diff": "@@ -1436,7 +1436,7 @@ axis_frame = None\ndef omnistaging_enabler() -> None:\nglobal thread_local_state, call_bind, find_top_trace, initial_style_staging, \\\nnew_master, reset_trace_state, extend_axis_env, axis...
Python
Apache License 2.0
google/jax
only construct one axis_index_p primitive Before this change, there were two versions, one used with omnistaging and one without. But that made bookkeeping hard and buggy. This change defines the axis_index_p primitive in core.py. Some of its rules are still changed when omnistaging is enabled.
260,335
21.08.2020 19:58:06
25,200
56b3688db90cc6c943a15eda4ca7e7c673c17e14
make random.choice error when shape isn't sequence fixes
[ { "change_type": "MODIFY", "old_path": "jax/random.py", "new_path": "jax/random.py", "diff": "@@ -556,6 +556,9 @@ def choice(key, a, shape=(), replace=True, p=None):\nReturns:\nAn array of shape `shape` containing samples from `a`.\n\"\"\"\n+ if not isinstance(shape, Sequence):\n+ raise TypeError(\"...
Python
Apache License 2.0
google/jax
make random.choice error when shape isn't sequence fixes #4124
260,298
24.08.2020 21:29:22
-7,200
0cc3802b0453efa85b45d98099590302608936a9
Fix documentation of scatter_* operations. * Fix documentation of scatter_* operations. This commit changes the documentation of the `unique_indices` parameter to scatter to better capture its intended meaning in XLA.
[ { "change_type": "MODIFY", "old_path": "jax/lax/lax.py", "new_path": "jax/lax/lax.py", "diff": "@@ -868,8 +868,9 @@ def scatter_add(operand: Array, scatter_indices: Array, updates: Array,\nrelate.\nindices_are_sorted: whether `scatter_indices` is known to be sorted. If\ntrue, may improve performance...
Python
Apache License 2.0
google/jax
Fix documentation of scatter_* operations. (#4138) * Fix documentation of scatter_* operations. This commit changes the documentation of the `unique_indices` parameter to scatter to better capture its intended meaning in XLA.
260,510
24.08.2020 21:08:23
25,200
774b5f688e69a33396309a1d0dadb58783fbba86
Remove frame check assertion in `extend_axis_env`.
[ { "change_type": "MODIFY", "old_path": "jax/core.py", "new_path": "jax/core.py", "diff": "@@ -1579,8 +1579,7 @@ def omnistaging_enabler() -> None:\ntry:\nyield\nfinally:\n- frame_ = thread_local_state.trace_state.axis_env.pop()\n- assert frame is frame_\n+ thread_local_state.trace_state.axis_env.pop...
Python
Apache License 2.0
google/jax
Remove frame check assertion in `extend_axis_env`.
260,335
25.08.2020 05:38:41
25,200
f4b05bc9ea18bda3dfab1cab77b2e0a36379a1c7
make pe.abstract_eval_fun use omnistaging
[ { "change_type": "MODIFY", "old_path": "jax/interpreters/partial_eval.py", "new_path": "jax/interpreters/partial_eval.py", "diff": "@@ -315,11 +315,10 @@ call_param_updaters: Dict[core.Primitive, Callable] = {}\ndef abstract_eval_fun(fun, *avals, **params):\n- pvals_in = [PartialVal.unknown(a) for a...
Python
Apache License 2.0
google/jax
make pe.abstract_eval_fun use omnistaging (#4139)
260,335
26.08.2020 10:21:56
25,200
1d93991003cba20235c21a51e313efa5e06a218e
allow random.choice to accept ndarray input * allow random.choice to accept ndarray `a` follow-up to to allow ndarray inputs to be passed * add jax.random.choice tests to cover ndarray input * don't use callables in test params it can mess with pytest-xdist because of hashing by id
[ { "change_type": "MODIFY", "old_path": "jax/random.py", "new_path": "jax/random.py", "diff": "@@ -98,6 +98,14 @@ def _is_prng_key(key: jnp.ndarray) -> bool:\n### utilities\n+# TODO(mattjj,jakevdp): add more info to error message, use this utility more\n+def _asarray(x):\n+ \"\"\"A more restrictive j...
Python
Apache License 2.0
google/jax
allow random.choice to accept ndarray input (#4145) * allow random.choice to accept ndarray `a` follow-up to #4137 to allow ndarray inputs to be passed * add jax.random.choice tests to cover ndarray input * don't use callables in test params it can mess with pytest-xdist because of hashing by id
260,298
27.08.2020 09:46:32
-7,200
1dc71b2f41bc17ef6d0854661fb64e9a668babea
[jax2tf] Add testing for add/mul/min/max conversion. * [jax2tf] Add testing for add/mul/min/max conversion. Only certain types are supported for each of the operations above. This commit adds previously missing tests to make this explicit.
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/tests/primitive_harness.py", "new_path": "jax/experimental/jax2tf/tests/primitive_harness.py", "diff": "@@ -22,6 +22,7 @@ from typing import Any, Callable, Dict, Iterable, Optional, NamedTuple, Sequence\nfrom absl import testing\nfrom j...
Python
Apache License 2.0
google/jax
[jax2tf] Add testing for add/mul/min/max conversion. (#4142) * [jax2tf] Add testing for add/mul/min/max conversion. Only certain types are supported for each of the operations above. This commit adds previously missing tests to make this explicit.
260,298
27.08.2020 09:47:19
-7,200
80114e51d643e945f14b66bd576208cd30d8e691
Add a boolean to _check_shapelike to accept or reject shapes * Add a boolean to _check_shapelike to accept or reject shapes corresponding to arrays of 0 elements. (Fixes google/jax#3972). * Added test for failures referenced in issue 3972.
[ { "change_type": "MODIFY", "old_path": "jax/lax/lax.py", "new_path": "jax/lax/lax.py", "diff": "@@ -4748,8 +4748,10 @@ def _reduce_window_chooser_jvp_rule(prim, g, operand, *, window_dimensions,\ndef _common_reduce_window_shape_rule(operand, window_dimensions,\nwindow_strides, padding, base_dilation...
Python
Apache License 2.0
google/jax
Add a boolean to _check_shapelike to accept or reject shapes (#4108) * Add a boolean to _check_shapelike to accept or reject shapes corresponding to arrays of 0 elements. (Fixes google/jax#3972). * Added test for failures referenced in issue 3972.
260,298
27.08.2020 11:04:32
-7,200
4d7396aa02f1d8f739b60e18ed75de4efcb4a427
Implement a proper shape checking rule for scatter. The implementation is based on the corresponding shape inference code in `tensorflow/compiler/xla/service/shape_inference.cc`. The tests added in `tests/lax_test.py` are similarly mirroring the corresponding tests in tensorflow, with slight adaptations for the particular setting of JAX.
[ { "change_type": "MODIFY", "old_path": "jax/lax/lax.py", "new_path": "jax/lax/lax.py", "diff": "@@ -3952,7 +3952,131 @@ def _scatter_dtype_rule(operand, scatter_indices, updates, **kwargs):\n_check_same_dtypes(\"scatter\", False, operand.dtype, updates.dtype)\nreturn dtypes.canonicalize_dtype(operan...
Python
Apache License 2.0
google/jax
Implement a proper shape checking rule for scatter. (#4144) The implementation is based on the corresponding shape inference code in `tensorflow/compiler/xla/service/shape_inference.cc`. The tests added in `tests/lax_test.py` are similarly mirroring the corresponding tests in tensorflow, with slight adaptations for the particular setting of JAX.
260,287
18.08.2020 09:14:38
0
7210d6f5d01ba7c64c7552628f37d66523feac9b
Add support for binding axis_name in gmap This allows executing collectives over the gmapped axes. This requires some extra manipulation of the gmapped jaxpr, since gmap exposes a single logical axis name, but evaluates the program using multiple "physical" axes. This also fixes some bugs around handling `multiple_returns` in vmap collective implementation.
[ { "change_type": "MODIFY", "old_path": "jax/core.py", "new_path": "jax/core.py", "diff": "@@ -1579,7 +1579,8 @@ def omnistaging_enabler() -> None:\ntry:\nyield\nfinally:\n- thread_local_state.trace_state.axis_env.pop()\n+ frame_ = thread_local_state.trace_state.axis_env.pop()\n+ assert frame is fram...
Python
Apache License 2.0
google/jax
Add support for binding axis_name in gmap This allows executing collectives over the gmapped axes. This requires some extra manipulation of the gmapped jaxpr, since gmap exposes a single logical axis name, but evaluates the program using multiple "physical" axes. This also fixes some bugs around handling `multiple_returns` in vmap collective implementation.
260,298
28.08.2020 16:27:32
-7,200
04f9ff7ff4ddb08649e94542900cd2c14755ca06
Addition of one more conclusive polynomial comparison case. * Addition of one more conclusive polynomial comparison case. In the case when the difference between two polynomials is a constant, it is possible to conclusively compare them. This commit adds such a case to masking.Poly.__ge__. * Added a few relevant tests in tests.masking_test.test_Poly_compare.
[ { "change_type": "MODIFY", "old_path": "jax/interpreters/masking.py", "new_path": "jax/interpreters/masking.py", "diff": "@@ -200,6 +200,10 @@ class Poly(dict):\nreturn False # See above.\nelif self == other:\nreturn True\n+ else:\n+ diff = self - other\n+ if diff.is_constant:\n+ return int(diff) >=...
Python
Apache License 2.0
google/jax
Addition of one more conclusive polynomial comparison case. (#4167) * Addition of one more conclusive polynomial comparison case. In the case when the difference between two polynomials is a constant, it is possible to conclusively compare them. This commit adds such a case to masking.Poly.__ge__. * Added a few relevant tests in tests.masking_test.test_Poly_compare.
260,287
28.08.2020 20:03:39
-7,200
a33f4dd8c8853a988e97b57542627869533a2ee7
Add support for axis_index inside vmap Also, reorganize the code to put all `axis_index` related functions in `lax_parallel.py`, next to all other parallel collectives.
[ { "change_type": "MODIFY", "old_path": "jax/core.py", "new_path": "jax/core.py", "diff": "@@ -645,7 +645,7 @@ class TraceStack:\nreturn new\nclass Sublevel(int): pass\n-AxisEnvFrame = namedtuple('AxisEnvFrame', ['name', 'size', 'tag'])\n+AxisEnvFrame = namedtuple('AxisEnvFrame', ['name', 'size', 'ma...
Python
Apache License 2.0
google/jax
Add support for axis_index inside vmap (#4168) Also, reorganize the code to put all `axis_index` related functions in `lax_parallel.py`, next to all other parallel collectives.
260,411
30.08.2020 12:38:14
-10,800
634c6259df26484a7e716db47105c589057d3580
More renaming of master to main in JAX internals
[ { "change_type": "MODIFY", "old_path": "jax/core.py", "new_path": "jax/core.py", "diff": "@@ -645,7 +645,7 @@ class TraceStack:\nreturn new\nclass Sublevel(int): pass\n-AxisEnvFrame = namedtuple('AxisEnvFrame', ['name', 'size', 'master_trace'])\n+AxisEnvFrame = namedtuple('AxisEnvFrame', ['name', 's...
Python
Apache License 2.0
google/jax
More renaming of master to main in JAX internals (#4179)