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
260,299
26.03.2021 11:14:43
0
90f7e06bcc3c06ff038b07637b5babaab23bd778
Allow integer inputs/outputs of linear_transpose Refine check to only allow float/complex -> float/complex or int -> int functions (i.e. no mixing float/int inputs/outputs).
[ { "change_type": "MODIFY", "old_path": "jax/_src/api.py", "new_path": "jax/_src/api.py", "diff": "@@ -2005,14 +2005,18 @@ def linear_transpose(fun: Callable, *primals) -> Callable:\nflat_fun, out_tree = flatten_fun_nokwargs(lu.wrap_init(fun), in_tree)\nin_avals = map(shaped_abstractify, primals_flat...
Python
Apache License 2.0
google/jax
Allow integer inputs/outputs of linear_transpose Refine check to only allow float/complex -> float/complex or int -> int functions (i.e. no mixing float/int inputs/outputs).
260,335
15.04.2021 15:16:29
25,200
9d6263a7436f2006722ad564d105e8931a850af1
support implicit broadcasting in transpose rules
[ { "change_type": "MODIFY", "old_path": "jax/_src/lax/lax.py", "new_path": "jax/_src/lax/lax.py", "diff": "@@ -141,7 +141,7 @@ def nextafter(x1: Array, x2: Array) -> Array:\nFor the smallest usable (i.e. normal) float, use ``tiny`` of ``jnp.finfo``.\n\"\"\"\n- return nextafter_p.bind(_brcast(x1, x2),...
Python
Apache License 2.0
google/jax
support implicit broadcasting in transpose rules
260,638
15.02.2021 20:43:55
-3,600
7091ae5af6652a119e774c98e34ab5d8720d3da6
Add support for padding and cropping to fft
[ { "change_type": "MODIFY", "old_path": "jax/_src/numpy/fft.py", "new_path": "jax/_src/numpy/fft.py", "diff": "# limitations under the License.\n+import operator\nimport numpy as np\nfrom jax import lax\nfrom jax.lib import xla_client\n+from jax._src.util import safe_zip\nfrom .util import _wraps\nfr...
Python
Apache License 2.0
google/jax
Add support for padding and cropping to fft
260,424
15.04.2021 14:21:53
-3,600
fa5e19b630e97bc197645455e7cd7e36db66d6cd
Fix Zero handling in select_jvp.
[ { "change_type": "MODIFY", "old_path": "jax/_src/lax/lax.py", "new_path": "jax/_src/lax/lax.py", "diff": "@@ -3899,6 +3899,9 @@ def _select_jvp(primals, tangents):\n_, on_true_dot, on_false_dot = tangents\nout = select(pred, on_true, on_false)\nif type(on_true_dot) is ad_util.Zero:\n+ if type(on_fal...
Python
Apache License 2.0
google/jax
Fix Zero handling in select_jvp.
260,287
19.04.2021 11:28:12
25,200
f285f0cc4774a2cce769caeca1439ed19f142cd0
Fix pjit resource checks to verify the shapes against the local mesh.
[ { "change_type": "MODIFY", "old_path": "jax/BUILD", "new_path": "jax/BUILD", "diff": "@@ -115,7 +115,10 @@ pytype_library(\nname = \"pjit\",\nsrcs = [\"experimental/pjit.py\"],\nsrcs_version = \"PY3\",\n- deps = [\":jax\"],\n+ deps = [\n+ \":experimental\",\n+ \":jax\",\n+ ],\n)\npytype_library(\n" ...
Python
Apache License 2.0
google/jax
Fix pjit resource checks to verify the shapes against the local mesh. PiperOrigin-RevId: 369263691
260,287
19.04.2021 11:35:12
25,200
d265fd56047bf2e85a24daa6b49339bd03299a7d
Ignore named shape when checking aval equality in AD AD of code with named axes is still WIP, and pmap still doesn't take proper care to handle them, so weaken the check for now.
[ { "change_type": "MODIFY", "old_path": "jax/interpreters/ad.py", "new_path": "jax/interpreters/ad.py", "diff": "@@ -176,8 +176,8 @@ def backward_pass(jaxpr: core.Jaxpr, consts, primals_in, cotangents_in):\nct_env[v] = add_tangents(ct_env[v], ct) if v in ct_env else ct\nif config.jax_enable_checks:\n...
Python
Apache License 2.0
google/jax
Ignore named shape when checking aval equality in AD AD of code with named axes is still WIP, and pmap still doesn't take proper care to handle them, so weaken the check for now. PiperOrigin-RevId: 369265258
260,287
20.04.2021 03:48:07
25,200
93c63d03417d8e89344cf84f387140d52f9576ef
Fix cache misses when re-creating equivalent mesh objects The `Mesh` class was missing `__eq__` and `__hash__` and inherited the (bad) Python defaults of comparison and hashing by identity.
[ { "change_type": "MODIFY", "old_path": "jax/interpreters/pxla.py", "new_path": "jax/interpreters/pxla.py", "diff": "@@ -1254,16 +1254,33 @@ mesh devices ndarray would have to be transposed before flattening and assignmen\nArrayMapping = OrderedDictType[MeshAxisName, int]\nclass Mesh:\n- __slots__ = ...
Python
Apache License 2.0
google/jax
Fix cache misses when re-creating equivalent mesh objects The `Mesh` class was missing `__eq__` and `__hash__` and inherited the (bad) Python defaults of comparison and hashing by identity. PiperOrigin-RevId: 369407380
260,287
20.04.2021 08:32:41
25,200
c09037bd144cc2e271da4d7f1077d281f9d0e7f4
Move vtile to batching.py, make it possible to add new BatchTraces No substantial behavior change right now, but the ability to use subclasses of BatchTrace comes in handy when adding support for nesting xmaps in the SPMD lowering.
[ { "change_type": "MODIFY", "old_path": "jax/_src/custom_derivatives.py", "new_path": "jax/_src/custom_derivatives.py", "diff": "@@ -324,7 +324,7 @@ def _custom_jvp_call_jaxpr_jvp(\nad.primitive_jvps[custom_jvp_call_jaxpr_p] = _custom_jvp_call_jaxpr_jvp\ndef _custom_jvp_call_jaxpr_vmap(\n- args, in_d...
Python
Apache License 2.0
google/jax
Move vtile to batching.py, make it possible to add new BatchTraces No substantial behavior change right now, but the ability to use subclasses of BatchTrace comes in handy when adding support for nesting xmaps in the SPMD lowering. PiperOrigin-RevId: 369445693
260,287
20.04.2021 11:39:33
25,200
09a519828cc807df337ce4d1922df2c165e67aca
Extend pjit error checking to rank errors too Otherwise one gets an inscrutable `IndexError`.
[ { "change_type": "MODIFY", "old_path": "jax/experimental/pjit.py", "new_path": "jax/experimental/pjit.py", "diff": "@@ -118,6 +118,11 @@ def _check_shapes_against_resources(what: str, resource_env, flat_avals, flat_ax\nif aval_axis_resources is None:\ncontinue\nshape = aval.shape\n+ if len(shape) < ...
Python
Apache License 2.0
google/jax
Extend pjit error checking to rank errors too Otherwise one gets an inscrutable `IndexError`. PiperOrigin-RevId: 369485185
260,287
20.04.2021 11:40:32
25,200
828e21060121c637aa5038a848999f7ce04df49f
Add a type checking rule for xmap Also fix the type checking code in core, which incorrectly propagated output avals of custom type checking rules.
[ { "change_type": "MODIFY", "old_path": "jax/core.py", "new_path": "jax/core.py", "diff": "@@ -1846,8 +1846,10 @@ def _check_jaxpr(jaxpr: Jaxpr, in_avals: Sequence[AbstractValue]):\ntypecheck_assert(all(not isinstance(ina, ConcreteArray) for ina in in_avals),\n\"Equation given ConcreteArray type inpu...
Python
Apache License 2.0
google/jax
Add a type checking rule for xmap Also fix the type checking code in core, which incorrectly propagated output avals of custom type checking rules. PiperOrigin-RevId: 369485371
260,287
21.04.2021 04:09:30
25,200
42d2e7620a81db03460738afcde13c9359b33701
Implement nesting of pjits Without this change nesting works only when the inner `pjit`ed functions don't close over any values.
[ { "change_type": "MODIFY", "old_path": "jax/experimental/pjit.py", "new_path": "jax/experimental/pjit.py", "diff": "@@ -27,6 +27,7 @@ from ..api_util import (argnums_partial_except, flatten_axes,\nfrom ..interpreters import ad\nfrom ..interpreters import pxla\nfrom ..interpreters import xla\n+from ....
Python
Apache License 2.0
google/jax
Implement nesting of pjits Without this change nesting works only when the inner `pjit`ed functions don't close over any values. PiperOrigin-RevId: 369626779
260,424
21.04.2021 11:49:21
-3,600
deb2227f4ad61304c2b017c97371f4467d9dd60b
Make sure the out_axes in the HashableFunction closure are hashable. By flattening them before putting them in the closure.
[ { "change_type": "MODIFY", "old_path": "jax/_src/api.py", "new_path": "jax/_src/api.py", "diff": "@@ -1627,9 +1627,14 @@ def pmap(\nlambda: (0,) * out_tree().num_leaves,\nclosure=out_axes)\nelse:\n+ # out_axes_thunk closes over the out_axes, they are flattened here to make\n+ # them hashable.\n+ out...
Python
Apache License 2.0
google/jax
Make sure the out_axes in the HashableFunction closure are hashable. By flattening them before putting them in the closure.
260,287
21.04.2021 11:04:52
25,200
a73fc71e0f8f930420448dad4abcb3e34fb63f8d
Implement JVP for pjit
[ { "change_type": "MODIFY", "old_path": "jax/experimental/pjit.py", "new_path": "jax/experimental/pjit.py", "diff": "@@ -32,7 +32,8 @@ from ..lib import xla_bridge as xb\nfrom ..lib import xla_client as xc\nfrom ..tree_util import tree_flatten, tree_unflatten\nfrom .._src.util import (extend_name_sta...
Python
Apache License 2.0
google/jax
Implement JVP for pjit PiperOrigin-RevId: 369692330
260,411
22.04.2021 08:14:56
-10,800
6ed068637b9f5721b3fd8ae8e78218960be1c715
Extended the comment about disabling the test_multiple_barriers
[ { "change_type": "MODIFY", "old_path": "tests/host_callback_test.py", "new_path": "tests/host_callback_test.py", "diff": "@@ -573,7 +573,9 @@ class HostCallbackIdTapTest(jtu.JaxTestCase):\nhcb.barrier_wait()\nself.assertEqual(received, set(range(count)))\n- # TODO(necula): see comment for test_multi...
Python
Apache License 2.0
google/jax
Extended the comment about disabling the test_multiple_barriers
260,642
08.04.2021 14:08:51
0
eb9d6e4d210c7d69760c94839122a2c959625f24
Pass axis name to _match_axes and add to error message.
[ { "change_type": "MODIFY", "old_path": "jax/interpreters/batching.py", "new_path": "jax/interpreters/batching.py", "diff": "@@ -57,7 +57,8 @@ def batch_subtrace(main, in_dims, *in_vals):\nyield out_vals, out_dims\n@lu.transformation\n-def _match_axes(axis_size, in_dims, out_dims_thunk, out_dim_dests...
Python
Apache License 2.0
google/jax
Pass axis name to _match_axes and add to error message.
260,335
09.04.2021 21:29:42
25,200
ba9233b9b6dd85147b5185f1bd920b93b6861fa8
prune trivial convert_element_types from jaxprs also add a test for not performing H2D transfers while tracing jnp.array
[ { "change_type": "MODIFY", "old_path": "docs/jaxpr.rst", "new_path": "docs/jaxpr.rst", "diff": "@@ -380,10 +380,8 @@ For the example consider the function ``func11`` below\nf = convert_element_type[ new_dtype=float32\nweak_type=False ] b\ng = add f e\n- h = convert_element_type[ new_dtype=float32\n-...
Python
Apache License 2.0
google/jax
prune trivial convert_element_types from jaxprs also add a test for not performing H2D transfers while tracing jnp.array
260,287
22.04.2021 15:30:03
25,200
23f847e0d377f5f7cbf50272838d6324a2114fef
Make the initial-style -> final-style conversion rule based Also, add a rule for pjit to make sure that we can eval jaxprs that contain pjits.
[ { "change_type": "MODIFY", "old_path": "jax/core.py", "new_path": "jax/core.py", "diff": "@@ -328,21 +328,13 @@ def eval_jaxpr(jaxpr: Jaxpr, consts, *args):\nsubfuns = [lu.wrap_init(partial(eval_jaxpr, call_jaxpr, ()))]\nelse:\nsubfuns = []\n- # TODO(apaszke): Make this initial -> final style conver...
Python
Apache License 2.0
google/jax
Make the initial-style -> final-style conversion rule based Also, add a rule for pjit to make sure that we can eval jaxprs that contain pjits. PiperOrigin-RevId: 369964136
260,287
23.04.2021 03:42:14
25,200
973ca07a04488d767242f49d5892f581a216bed0
Add stricter resource overlap checking We've had some checks for coinciding logical axes mapped to the same resources in the existing xmap code, but they were quite lax. This adds a proper type checker and a bunch of tests to verify that we can catch the interesting failure cases.
[ { "change_type": "MODIFY", "old_path": "jax/experimental/maps.py", "new_path": "jax/experimental/maps.py", "diff": "@@ -32,6 +32,7 @@ from .._src.tree_util import _replace_nones\nfrom ..api_util import (flatten_fun_nokwargs, flatten_axes, _ensure_index_tuple,\ndonation_vector)\nfrom .._src import so...
Python
Apache License 2.0
google/jax
Add stricter resource overlap checking We've had some checks for coinciding logical axes mapped to the same resources in the existing xmap code, but they were quite lax. This adds a proper type checker and a bunch of tests to verify that we can catch the interesting failure cases. PiperOrigin-RevId: 370051512
260,424
23.04.2021 14:43:20
-3,600
b244e2b8c8ba0bd6e848814173536b1e765bd602
Add eval_shape to the UnexpectedTracerError too.
[ { "change_type": "MODIFY", "old_path": "jax/_src/api.py", "new_path": "jax/_src/api.py", "diff": "@@ -2391,7 +2391,8 @@ def eval_shape(fun: Callable, *args, **kwargs):\nargs_flat, in_tree = tree_flatten((args, kwargs))\nwrapped_fun, out_tree = flatten_fun(lu.wrap_init(fun), in_tree)\nout = pe.abstra...
Python
Apache License 2.0
google/jax
Add eval_shape to the UnexpectedTracerError too.
260,335
23.04.2021 08:31:11
25,200
4ac8937ebe782385e26af5b652c77be2281f4633
enable fori-to-scan lowering
[ { "change_type": "MODIFY", "old_path": "jax/_src/lax/control_flow.py", "new_path": "jax/_src/lax/control_flow.py", "diff": "@@ -138,8 +138,8 @@ def _fori_body_fun(body_fun):\n@cache()\ndef _fori_scan_body_fun(body_fun):\ndef scanned_fun(loop_carry, _):\n- i, upper, x = loop_carry\n- return (lax.add(...
Python
Apache License 2.0
google/jax
enable fori-to-scan lowering
260,335
24.04.2021 15:18:26
25,200
8f434539e12725068dc3aea5c7702f938da523c5
re-enable a working test
[ { "change_type": "MODIFY", "old_path": "tests/api_test.py", "new_path": "tests/api_test.py", "diff": "@@ -2274,8 +2274,6 @@ class APITest(jtu.JaxTestCase):\nf()\ndef test_xla_computation_zeros_doesnt_device_put(self):\n- raise unittest.SkipTest(\"broken test\") # TODO(mattjj): fix\n-\nwith jtu.count...
Python
Apache License 2.0
google/jax
re-enable a working test
260,287
26.04.2021 07:51:19
25,200
64130daf819a3d96b38d45586c173e9ff54514a9
Fix xmap nesting in SPMD lowering mode Previous implementation has been incorrectly accumulating the SPMD axis partitioning into an existing dimension, while what we really need is to create a separate dimension for every annotation.
[ { "change_type": "MODIFY", "old_path": "jax/experimental/maps.py", "new_path": "jax/experimental/maps.py", "diff": "@@ -24,6 +24,7 @@ from functools import wraps, partial, partialmethod\nfrom .. import numpy as jnp\nfrom .. import core\n+from .. import config\nfrom .. import linear_util as lu\nfrom ...
Python
Apache License 2.0
google/jax
Fix xmap nesting in SPMD lowering mode Previous implementation has been incorrectly accumulating the SPMD axis partitioning into an existing dimension, while what we really need is to create a separate dimension for every annotation. PiperOrigin-RevId: 370455989
260,287
26.04.2021 11:41:26
25,200
d0606463e40c878702e511f5fd8dd1e8a6e208a4
Fix the batching rule for named reductions
[ { "change_type": "MODIFY", "old_path": "jax/_src/lax/parallel.py", "new_path": "jax/_src/lax/parallel.py", "diff": "@@ -558,7 +558,6 @@ def _reduction_with_positional_batcher(prim, vals_in, dims_in, axis_index_groups\nif axis_index_groups is not None:\nraise NotImplementedError(\"axis_index_groups n...
Python
Apache License 2.0
google/jax
Fix the batching rule for named reductions PiperOrigin-RevId: 370505998
260,287
27.04.2021 02:19:18
25,200
9f2ac6e0a3b4ad529490e927180d5b710dfcf4d3
Add resource type-checking rules for pjit and with_sharding_constraint To prevent people from doubly-sharding values that are sharded by outer xmaps.
[ { "change_type": "MODIFY", "old_path": "jax/experimental/maps.py", "new_path": "jax/experimental/maps.py", "diff": "@@ -575,7 +575,8 @@ def make_xmap_callable(fun: lu.WrappedFun,\njaxpr, out_avals, consts = pe.trace_to_jaxpr_final(fun, mapped_in_avals)\nout_axes = out_axes_thunk()\n_check_out_avals_...
Python
Apache License 2.0
google/jax
Add resource type-checking rules for pjit and with_sharding_constraint To prevent people from doubly-sharding values that are sharded by outer xmaps. PiperOrigin-RevId: 370635057
260,411
27.04.2021 15:21:20
-10,800
80cd273de815826ebe65a25e80515f66bc951ed8
[jax2tf] Improve the handling of variable capture for jax2tf.call Fixes:
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/call_tf.py", "new_path": "jax/experimental/jax2tf/call_tf.py", "diff": "@@ -226,7 +226,6 @@ def _call_tf_translation_rule(builder, *args_op, func_tf,\nassert next_var_idx < len(func_tf_concrete.variables)\n# TODO(necula): better checkin...
Python
Apache License 2.0
google/jax
[jax2tf] Improve the handling of variable capture for jax2tf.call Fixes: #6050
260,287
27.04.2021 07:11:41
25,200
aa9d350aaf99482c32164e573879399d272a3ba6
Fix an assertion in xmap code This has gone under the radar because it's valid in tests and it's in code sufficiently new that I don't expect anyone to be using it just yet.
[ { "change_type": "MODIFY", "old_path": "jax/experimental/maps.py", "new_path": "jax/experimental/maps.py", "diff": "@@ -1111,7 +1111,7 @@ def _xmap_translation_rule_spmd(c, axis_env,\nfor dim, dim_extra_axis in enumerate(extra):\nif dim_extra_axis is None: continue\nassert dim_extra_axis not in axes...
Python
Apache License 2.0
google/jax
Fix an assertion in xmap code This has gone under the radar because it's valid in tests and it's in code sufficiently new that I don't expect anyone to be using it just yet. PiperOrigin-RevId: 370672928
260,411
28.04.2021 12:22:32
-10,800
d762ec1d21b9a470c0790e642c85e1f133dd64d5
[host_callback] Minor fix to use the new xla_shape.is_token
[ { "change_type": "MODIFY", "old_path": "CHANGELOG.md", "new_path": "CHANGELOG.md", "diff": "@@ -17,6 +17,7 @@ PLEASE REMEMBER TO CHANGE THE '..master' WITH AN ACTUAL TAG in GITHUB LINK.\n* {func}`jax.nonzero` has a new optional `size` argument that allows it to\nbe used within `jit` ({jax-issue}`#65...
Python
Apache License 2.0
google/jax
[host_callback] Minor fix to use the new xla_shape.is_token
260,287
09.04.2021 12:43:40
0
8df502aeb2cf45934df6722069b875c108c1867c
Use the axis names attached to a primitive when selecting the top trace This is useful e.g. for handling psums of values that are not sharded, but are also not statically known constants that we can fold.
[ { "change_type": "MODIFY", "old_path": "jax/_src/lax/parallel.py", "new_path": "jax/_src/lax/parallel.py", "diff": "@@ -690,7 +690,7 @@ def _psum_transpose_rule(cts, *args, axes, axis_index_groups):\naxis_index_groups=axis_index_groups)\nreturn tree_util.tree_unflatten(treedef, nonzero_in_cts)\n-psu...
Python
Apache License 2.0
google/jax
Use the axis names attached to a primitive when selecting the top trace This is useful e.g. for handling psums of values that are not sharded, but are also not statically known constants that we can fold.
260,335
22.04.2021 20:56:17
25,200
3c400a3e588abf9e2259119c50343cba6f3477f1
add 'inline' option to xla_call for jaxpr inlining
[ { "change_type": "MODIFY", "old_path": "docs/jaxpr.rst", "new_path": "docs/jaxpr.rst", "diff": "@@ -433,6 +433,7 @@ computation should run. For example\nin (g,) }\ndevice=None\ndonated_invars=(False, False)\n+ inline=False\nname=inner ] a b\nd = convert_element_type[ new_dtype=float32\nweak_type=Fal...
Python
Apache License 2.0
google/jax
add 'inline' option to xla_call for jaxpr inlining
260,335
29.04.2021 08:25:26
25,200
e968672740547b83639932366565dc4906c6e9c9
add tanh to jax.nn package
[ { "change_type": "MODIFY", "old_path": "jax/nn/__init__.py", "new_path": "jax/nn/__init__.py", "diff": "# flake8: noqa: F401\n+from jax.numpy import tanh\nfrom . import initializers\nfrom jax._src.nn.functions import (\ncelu,\n" }, { "change_type": "MODIFY", "old_path": "tests/nn_test.py...
Python
Apache License 2.0
google/jax
add tanh to jax.nn package
260,563
01.05.2021 01:05:22
-7,200
e985e86808325c388c17d29ef0fc09ba6905fe41
Implement jnp.r_ and jnp.c_
[ { "change_type": "MODIFY", "old_path": "docs/jax.numpy.rst", "new_path": "docs/jax.numpy.rst", "diff": "@@ -89,6 +89,7 @@ Not every function in NumPy is implemented; contributions are welcome!\nbroadcast_arrays\nbroadcast_shapes\nbroadcast_to\n+ c_\ncan_cast\ncbrt\ncdouble\n@@ -298,6 +299,7 @@ Not e...
Python
Apache License 2.0
google/jax
Implement jnp.r_ and jnp.c_
260,563
01.05.2021 19:32:19
-7,200
f559a9bfb4a836be2f5d7007d02def39997d130e
Fix pip install instruction for forks in contributing section
[ { "change_type": "MODIFY", "old_path": "docs/contributing.md", "new_path": "docs/contributing.md", "diff": "@@ -39,7 +39,7 @@ Follow these steps to contribute code:\n```bash\ngit clone https://github.com/YOUR_USERNAME/jax\ncd jax\n- pip install -r requirements.txt # Installs all testing requirements...
Python
Apache License 2.0
google/jax
Fix pip install instruction for forks in contributing section
260,335
01.05.2021 12:32:44
25,200
fe297e39ca37896b75d7943b9b77c0b53fad13ee
add 'inline' to jit docstring
[ { "change_type": "MODIFY", "old_path": "jax/_src/api.py", "new_path": "jax/_src/api.py", "diff": "@@ -263,7 +263,10 @@ def jit(\nbuffers to reduce the amount of memory needed to perform a computation,\nfor example recycling one of your input buffers to store a result. You\nshould not reuse buffers t...
Python
Apache License 2.0
google/jax
add 'inline' to jit docstring
260,335
01.05.2021 12:41:41
25,200
ff6866c4b3757cde66fe659c2f27d8aeff024e8f
new_sublevel in jax2tf
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/jax2tf.py", "new_path": "jax/experimental/jax2tf/jax2tf.py", "diff": "@@ -310,7 +310,9 @@ def _interpret_fun(fun: lu.WrappedFun,\n) -> Sequence[Tuple[TfVal, core.AbstractValue]]:\nwith core.new_base_main(TensorFlowTrace) as main: # type...
Python
Apache License 2.0
google/jax
new_sublevel in jax2tf
260,545
01.05.2021 19:07:58
25,200
74b67627456c0712a1464b7da79548f38b6ce042
Fixed error in autodidax broadasting batching
[ { "change_type": "MODIFY", "old_path": "docs/autodidax.ipynb", "new_path": "docs/autodidax.ipynb", "diff": "\" if x_bdim != y_bdim:\\n\",\n\" if x_bdim is not_mapped:\\n\",\n\" x = move_batch_axis(axis_size, x_bdim, y_bdim, x)\\n\",\n+ \" x_bdim = y_bdim\\n\",\n\" else:\\n\",\n\" y = move_batch_axis...
Python
Apache License 2.0
google/jax
Fixed error in autodidax broadasting batching
260,631
01.05.2021 22:18:13
25,200
75b00a1235e44898c5613d5c4479061970e1a22b
Copybara import of the project: by Matthew Johnson add 'inline' option to xla_call for jaxpr inlining by Matthew Johnson add 'inline' to jit docstring by Matthew Johnson new_sublevel in jax2tf
[ { "change_type": "MODIFY", "old_path": "docs/jaxpr.rst", "new_path": "docs/jaxpr.rst", "diff": "@@ -433,7 +433,6 @@ computation should run. For example\nin (g,) }\ndevice=None\ndonated_invars=(False, False)\n- inline=False\nname=inner ] a b\nd = convert_element_type[ new_dtype=float32\nweak_type=Fal...
Python
Apache License 2.0
google/jax
Copybara import of the project: -- 3c400a3e588abf9e2259119c50343cba6f3477f1 by Matthew Johnson <mattjj@google.com>: add 'inline' option to xla_call for jaxpr inlining -- fe297e39ca37896b75d7943b9b77c0b53fad13ee by Matthew Johnson <mattjj@google.com>: add 'inline' to jit docstring -- ff6866c4b3757cde66fe659c2f27d8aeff024e8f by Matthew Johnson <mattjj@google.com>: new_sublevel in jax2tf PiperOrigin-RevId: 371542778
260,335
03.05.2021 21:40:50
25,200
7ec0b40173dd51ff5c611a7b589ceaf879dfe3b0
Roll-forward of which broke internal tests.
[ { "change_type": "MODIFY", "old_path": "docs/jaxpr.rst", "new_path": "docs/jaxpr.rst", "diff": "@@ -433,6 +433,7 @@ computation should run. For example\nin (g,) }\ndevice=None\ndonated_invars=(False, False)\n+ inline=False\nname=inner ] a b\nd = convert_element_type[ new_dtype=float32\nweak_type=Fal...
Python
Apache License 2.0
google/jax
Roll-forward of #6584, which broke internal tests. PiperOrigin-RevId: 371839298
260,287
05.05.2021 06:07:16
25,200
049383c553612c4b9a5c8822d56b7786fb9115b6
Make sure that we don't pass unhashable axis_resources to cached pjit implementation
[ { "change_type": "MODIFY", "old_path": "jax/experimental/pjit.py", "new_path": "jax/experimental/pjit.py", "diff": "@@ -100,8 +100,8 @@ def pjit(fun: Callable,\nlocal_in_avals = tuple(core.raise_to_shaped(core.get_aval(a)) for a in args_flat)\njaxpr, in_axis_resources_flat, out_axis_resources_flat =...
Python
Apache License 2.0
google/jax
Make sure that we don't pass unhashable axis_resources to cached pjit implementation PiperOrigin-RevId: 372110726
260,287
05.05.2021 06:43:47
25,200
31f35373269803f6653b0e21b4f65686c1e03f2f
Make sure that no errors are raised for no-op PartitionSpecs
[ { "change_type": "MODIFY", "old_path": "jax/experimental/pjit.py", "new_path": "jax/experimental/pjit.py", "diff": "@@ -202,6 +202,7 @@ def _check_unique_resources(axis_resources, arg_name):\nfor arg_axis_resources in axis_resources:\nif not arg_axis_resources: continue\nresource_counts = Counter(it...
Python
Apache License 2.0
google/jax
Make sure that no errors are raised for no-op PartitionSpecs PiperOrigin-RevId: 372115524
260,347
06.05.2021 10:16:42
0
17cc9502b76c5584e45533f9c6be4ff91b335c42
Make jnp.unravel_index rank promotion explicit. There was a silent rank promotion in jax.numpy.unravel_index which caused the code to fail with jax_numpy_rank_promotion='raise'. This commit makes the rank promotion explicit. Furthermore, it fixes an outdated comment in the same function.
[ { "change_type": "MODIFY", "old_path": "jax/_src/numpy/lax_numpy.py", "new_path": "jax/_src/numpy/lax_numpy.py", "diff": "@@ -1403,8 +1403,9 @@ def unravel_index(indices, shape):\ntotal_size = cumulative_sizes[0]\n# Clip so raveling and unraveling an oob index will not change the behavior\nclipped_i...
Python
Apache License 2.0
google/jax
Make jnp.unravel_index rank promotion explicit. There was a silent rank promotion in jax.numpy.unravel_index which caused the code to fail with jax_numpy_rank_promotion='raise'. This commit makes the rank promotion explicit. Furthermore, it fixes an outdated comment in the same function.
260,287
06.05.2021 04:18:47
25,200
e481da90567d444e14efb8d4afd25791404c1099
Raise a clear error when pjit is used with an empty mesh Technically we could make pjit behave in the same way as jit in that case, but this felt easier. We can always relax it in the future.
[ { "change_type": "MODIFY", "old_path": "jax/experimental/pjit.py", "new_path": "jax/experimental/pjit.py", "diff": "@@ -81,6 +81,9 @@ def pjit(fun: Callable,\n# Putting this outside of wrapped would make resources lexically scoped\nresource_env = maps.thread_resources.env\nmesh = resource_env.physic...
Python
Apache License 2.0
google/jax
Raise a clear error when pjit is used with an empty mesh Technically we could make pjit behave in the same way as jit in that case, but this felt easier. We can always relax it in the future. PiperOrigin-RevId: 372314531
260,631
06.05.2021 04:44:41
25,200
4c0d3671c8f55908c96c42654700d247b9467f8d
Remove non-determinism in axis ordering during vectorization.
[ { "change_type": "MODIFY", "old_path": "jax/experimental/maps.py", "new_path": "jax/experimental/maps.py", "diff": "@@ -639,7 +639,7 @@ class EvaluationPlan(NamedTuple):\naxis_resource_count = _get_axis_resource_count(axis_resources, resource_env)\naxis_subst_dict = dict(axis_resources)\naxis_vmap_s...
Python
Apache License 2.0
google/jax
Remove non-determinism in axis ordering during vectorization. PiperOrigin-RevId: 372317377
260,287
06.05.2021 12:00:18
25,200
ee93ee221c24c8a23d4ebc73cc6dc9c08f18de87
Add a docstring for pjit
[ { "change_type": "ADD", "old_path": null, "new_path": "docs/jax.experimental.pjit.rst", "diff": "+jax.experimental.pjit module\n+============================\n+\n+.. automodule:: jax.experimental.pjit\n+\n+API\n+---\n+\n+.. autofunction:: pjit\n" }, { "change_type": "MODIFY", "old_path":...
Python
Apache License 2.0
google/jax
Add a docstring for pjit Co-authored-by: Skye Wanderman-Milne <skyewm@google.com> PiperOrigin-RevId: 372393461
260,287
06.05.2021 12:34:15
25,200
c3553d22d6917bf4448f6e043726507cdfa40637
Implement vmap for pjit
[ { "change_type": "MODIFY", "old_path": "jax/experimental/pjit.py", "new_path": "jax/experimental/pjit.py", "diff": "# See the License for the specific language governing permissions and\n# limitations under the License.\n+from enum import IntEnum\nimport numpy as np\nfrom collections import OrderedD...
Python
Apache License 2.0
google/jax
Implement vmap for pjit PiperOrigin-RevId: 372401914
260,335
05.05.2021 12:44:49
25,200
d88acd8b8c52245d349f4853d0e0c1663b55b1bc
autodidax: delete while_loop for now
[ { "change_type": "MODIFY", "old_path": "docs/autodidax.ipynb", "new_path": "docs/autodidax.ipynb", "diff": "\"\\n\",\n\"There are actually two rules to write: one for trace-time partial evaluation,\\n\",\n\"which we'll call `xla_call_partial_eval`, and one for partial evaluation of\\n\",\n- \"jaxprs...
Python
Apache License 2.0
google/jax
autodidax: delete while_loop for now
260,335
06.05.2021 17:55:47
25,200
d21e8c06571bf80fa79da1b760585027ed27a14d
handle case where trace debug_info is None
[ { "change_type": "MODIFY", "old_path": "jax/_src/errors.py", "new_path": "jax/_src/errors.py", "diff": "@@ -23,7 +23,7 @@ class _JAXErrorMixin:\nerror_page = self._error_page\nmodule_name = self._module_name\nclass_name = self.__class__.__name__\n- error_msg = f'{message}See {error_page}#{module_nam...
Python
Apache License 2.0
google/jax
handle case where trace debug_info is None
260,335
06.05.2021 20:23:34
25,200
173cd062855d38ce8bcf4e5265946332e1f85d2f
fix debug info handling (again)
[ { "change_type": "MODIFY", "old_path": "jax/interpreters/partial_eval.py", "new_path": "jax/interpreters/partial_eval.py", "diff": "@@ -907,6 +907,8 @@ class DynamicJaxprTracer(core.Tracer):\ndef _origin_msg(self):\ninvar_pos, progenitor_eqns = self._trace.frame.find_progenitors(self)\ndbg = self._t...
Python
Apache License 2.0
google/jax
fix debug info handling (again)
260,411
08.05.2021 07:38:50
-10,800
244fc7b11cb91faec6275e3466dae4b8ab11a821
[jax2tf] Expand shape polymorphism support for some instances of lax.gather
[ { "change_type": "MODIFY", "old_path": "jax/_src/numpy/lax_numpy.py", "new_path": "jax/_src/numpy/lax_numpy.py", "diff": "@@ -4829,11 +4829,11 @@ def _gather(arr, treedef, static_idx, dynamic_idx):\n# Avoid calling gather if the slice shape is empty, both as a fast path and to\n# handle cases like z...
Python
Apache License 2.0
google/jax
[jax2tf] Expand shape polymorphism support for some instances of lax.gather
260,690
06.05.2021 16:36:52
25,200
8d54a53b538a3e14a8425b5ec4273b2eb17a8667
Update convolutions.md Update convolutions.ipynb Update convolutions.ipynb Update convolutions.md
[ { "change_type": "MODIFY", "old_path": "docs/notebooks/convolutions.ipynb", "new_path": "docs/notebooks/convolutions.ipynb", "diff": "],\n\"source\": [\n\"# 2D kernel - HWIO layout\\n\",\n- \"kernel = jnp.zeros((3, 3, 3, 3), dtype=np.float32)\\n\",\n+ \"kernel = jnp.zeros((3, 3, 3, 3), dtype=jnp.flo...
Python
Apache License 2.0
google/jax
Update convolutions.md Update convolutions.ipynb Update convolutions.ipynb Update convolutions.md
260,654
07.05.2021 15:04:40
-3,600
db39a67ccabce689da12e9ddc59746150c4d53af
Update documentation for custom_jvp handling of nondiff_argnums as arguments of _fwd and _bwd rules.
[ { "change_type": "MODIFY", "old_path": "docs/notebooks/Custom_derivative_rules_for_Python_code.ipynb", "new_path": "docs/notebooks/Custom_derivative_rules_for_Python_code.ipynb", "diff": "\"id\": \"0u0jn4aWC8k1\"\n},\n\"source\": [\n- \"A similar option exists for `jax.custom_vjp`, and similarly the...
Python
Apache License 2.0
google/jax
Update documentation for custom_jvp handling of nondiff_argnums as arguments of _fwd and _bwd rules.
260,411
13.05.2021 09:10:24
-10,800
e7568c7ae62838c588a1f112222ef33c4d62613b
Add additional message to the error when we cannot convert
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/jax2tf.py", "new_path": "jax/experimental/jax2tf/jax2tf.py", "diff": "@@ -114,11 +114,13 @@ tf_impl_with_avals: Dict[core.Primitive,\n# requiring a TFXLA operation, an exception is thrown instead.\n_enable_xla = True\n-def _xla_path_dis...
Python
Apache License 2.0
google/jax
Add additional message to the error when we cannot convert
260,296
13.05.2021 12:20:31
25,200
bc84c9fe8fd6ce55b6376d479a544897cc0b0c5a
Add `lax.conv_general_dilated_local`
[ { "change_type": "MODIFY", "old_path": "docs/jax.lax.rst", "new_path": "docs/jax.lax.rst", "diff": "@@ -53,6 +53,7 @@ Operators\nconv\nconvert_element_type\nconv_general_dilated\n+ conv_general_dilated_local\nconv_general_dilated_patches\nconv_with_general_padding\nconv_transpose\n" }, { "ch...
Python
Apache License 2.0
google/jax
Add `lax.conv_general_dilated_local`
260,424
14.05.2021 15:46:27
-3,600
73e9302fc34cfeaf575a5a34edf1119b36dc3ce8
Fix `jsp.linalg.lu` translation rule to pass backend arg to `lower_fun`. If it doesn't, trying to run `lu` with a custom CPU backend when a GPU is present results in a `Unable to resolve runtime symbol: `cuda_lu_pivots_to_permutation'` fatal error.
[ { "change_type": "MODIFY", "old_path": "jax/_src/lax/linalg.py", "new_path": "jax/_src/lax/linalg.py", "diff": "@@ -1002,7 +1002,7 @@ def _lu_batching_rule(batched_args, batch_dims):\nx = batching.moveaxis(x, bd, 0)\nreturn lu_p.bind(x), (0, 0, 0)\n-def _lu_cpu_gpu_translation_rule(getrf_impl, c, op...
Python
Apache License 2.0
google/jax
Fix `jsp.linalg.lu` translation rule to pass backend arg to `lower_fun`. If it doesn't, trying to run `lu` with a custom CPU backend when a GPU is present results in a `Unable to resolve runtime symbol: `cuda_lu_pivots_to_permutation'` fatal error.
260,527
01.05.2021 12:09:03
0
81903e894bf142c71c35fea5c570994426c6df6c
Add JVP rules COO sparse ops. Updated coo_matvec jvp rule. Make flake8 happy.
[ { "change_type": "MODIFY", "old_path": "jax/experimental/sparse_ops.py", "new_path": "jax/experimental/sparse_ops.py", "diff": "@@ -39,16 +39,17 @@ from jax import api\nfrom jax import core\nfrom jax import jit\nfrom jax import tree_util\n+from jax import lax\nfrom jax.interpreters import xla\nfrom ...
Python
Apache License 2.0
google/jax
Add JVP rules COO sparse ops. Updated coo_matvec jvp rule. Make flake8 happy.
260,411
17.05.2021 10:01:13
-10,800
a08cdb30ff60e93a8e13b34cc745238676fdcac4
[jax2tf] Update the limitations for unsupported primitives Also update the documentation.
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/README.md", "new_path": "jax/experimental/jax2tf/README.md", "diff": "@@ -645,3 +645,29 @@ To run jax2tf on GPU, both jaxlib and TensorFlow must be installed with support\nfor CUDA. One must be mindful to install a version of CUDA that ...
Python
Apache License 2.0
google/jax
[jax2tf] Update the limitations for unsupported primitives Also update the documentation.
260,372
18.05.2021 15:22:06
21,600
bcd5deb2697387054153205f39898be39b861b7f
Prevent nans in scale_and_translate fixes
[ { "change_type": "MODIFY", "old_path": "jax/_src/image/scale.py", "new_path": "jax/_src/image/scale.py", "diff": "@@ -66,7 +66,8 @@ def compute_weight_mat(input_size: int, output_size: int, scale,\nwith np.errstate(invalid='ignore', divide='ignore'):\nweights = jnp.where(\njnp.abs(total_weight_sum) ...
Python
Apache License 2.0
google/jax
Prevent nans in scale_and_translate fixes #6780
260,411
18.05.2021 17:13:09
-10,800
a27109d1bd38efae9ec01cc391e000ce4c2071ed
[jax2tf] Added documentation explaining how to handle undefined TF ops Added a test case showing how to mix compileable and non-compileable code.
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/README.md", "new_path": "jax/experimental/jax2tf/README.md", "diff": "@@ -77,6 +77,14 @@ The Autograph feature of `tf.function` cannot be expected to work on\nfunctions converted from JAX as above, so it is recommended to\nset `autograp...
Python
Apache License 2.0
google/jax
[jax2tf] Added documentation explaining how to handle undefined TF ops Added a test case showing how to mix compileable and non-compileable code.
260,678
18.05.2021 10:18:58
18,000
bbcaec4a3aaaca7f8d33223d86a539d9253d37bf
Initial implementation of jax.numpy.poly This is an initial jax.numpy.poly implementation. It is tested by testPoly in the tests/lax_numpy_test.py test file.
[ { "change_type": "MODIFY", "old_path": "docs/jax.numpy.rst", "new_path": "docs/jax.numpy.rst", "diff": "@@ -286,6 +286,7 @@ Not every function in NumPy is implemented; contributions are welcome!\npad\npercentile\npiecewise\n+ poly\npolyadd\npolyder\npolyint\n" }, { "change_type": "MODIFY", ...
Python
Apache License 2.0
google/jax
Initial implementation of jax.numpy.poly This is an initial jax.numpy.poly implementation. It is tested by testPoly in the tests/lax_numpy_test.py test file.
260,372
20.05.2021 21:49:41
21,600
9f61f41e4730664ce2d0977383ac1661c027c461
prevent nans in _fill_lanczos_kernel
[ { "change_type": "MODIFY", "old_path": "jax/_src/image/scale.py", "new_path": "jax/_src/image/scale.py", "diff": "@@ -24,9 +24,8 @@ import numpy as np\ndef _fill_lanczos_kernel(radius, x):\ny = radius * jnp.sin(np.pi * x) * jnp.sin(np.pi * x / radius)\n- with np.errstate(divide='ignore', invalid='ig...
Python
Apache License 2.0
google/jax
prevent nans in _fill_lanczos_kernel
260,411
21.05.2021 11:16:00
-10,800
eabe631fa45160a3984c81734b74f69b4a6e6241
Update jax2tf.py Minor typo fix
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/jax2tf.py", "new_path": "jax/experimental/jax2tf/jax2tf.py", "diff": "@@ -171,7 +171,7 @@ def convert(fun: Callable, *,\nsequence of TF ops for any non-zero values of the dimension variables.\npolymorphic_shapes are only supported for p...
Python
Apache License 2.0
google/jax
Update jax2tf.py Minor typo fix
260,411
11.05.2021 11:11:37
-10,800
2ad9c0c34c096b67a6ea40963f6a5b20bde65355
[jax2tf] Fix the scoping of the enable_xla conversion parameter Previously, the global enable_xla flag was set upon entry to `jax.convert`. It should instead be set only for the duration of the just-in-time conversion, which may happen later when the converted function is invoked.
[ { "change_type": "MODIFY", "old_path": "CHANGELOG.md", "new_path": "CHANGELOG.md", "diff": "@@ -10,6 +10,16 @@ PLEASE REMEMBER TO CHANGE THE '..master' WITH AN ACTUAL TAG in GITHUB LINK.\n## jax 0.2.14 (unreleased)\n* [GitHub commits](https://github.com/google/jax/compare/jax-v0.2.13...master).\n+* ...
Python
Apache License 2.0
google/jax
[jax2tf] Fix the scoping of the enable_xla conversion parameter Previously, the global enable_xla flag was set upon entry to `jax.convert`. It should instead be set only for the duration of the just-in-time conversion, which may happen later when the converted function is invoked.
260,411
21.05.2021 11:42:44
-10,800
b3411cc92ab79ed073dc80919b493daac7de2db7
Update jax2tf_test.py Fix the type of np.zeros to be float32. Also, replaced `jnp.zeros` with `np.zeros` because technically the argument to `f_tf` should be a TF value, not a JAX value.
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/tests/jax2tf_test.py", "new_path": "jax/experimental/jax2tf/tests/jax2tf_test.py", "diff": "@@ -522,7 +522,7 @@ class Jax2TfTest(tf_test_util.JaxToTfTestCase):\nreturn jnp.sum(x)\nf_tf = jax2tf.convert(f_jax)\nself.assertAllClose(\n- f_...
Python
Apache License 2.0
google/jax
Update jax2tf_test.py Fix the type of np.zeros to be float32. Also, replaced `jnp.zeros` with `np.zeros` because technically the argument to `f_tf` should be a TF value, not a JAX value.
260,411
21.05.2021 11:10:41
-10,800
e7766838db1f447ade8870b22255de0b7dd97e0a
Minor cleanup of the translation rules for cumred primitives
[ { "change_type": "MODIFY", "old_path": "jax/_src/lax/control_flow.py", "new_path": "jax/_src/lax/control_flow.py", "diff": "@@ -2574,39 +2574,27 @@ def _cumred_dtype_rule(name, operand, *args, **kw):\n\"of number.\".format(name, np.dtype(operand.dtype).name))\nreturn dtypes.canonicalize_dtype(operan...
Python
Apache License 2.0
google/jax
Minor cleanup of the translation rules for cumred primitives
260,411
23.05.2021 19:32:45
-10,800
70f0110b3296eb26b04d33758c07724a78ae74d3
Fix dtypes.issubdtype when called with "bfloat16" (as string) Fixes:
[ { "change_type": "MODIFY", "old_path": "jax/_src/dtypes.py", "new_path": "jax/_src/dtypes.py", "diff": "@@ -198,6 +198,8 @@ def _issubclass(a, b):\nreturn False\ndef issubdtype(a, b):\n+ if a == \"bfloat16\":\n+ a = bfloat16\nif a == bfloat16:\nif isinstance(b, np.dtype):\nreturn b == _bfloat16_dtyp...
Python
Apache License 2.0
google/jax
Fix dtypes.issubdtype when called with "bfloat16" (as string) Fixes: #6813
260,595
24.05.2021 21:13:43
-3,600
dc81610bb682b1c97e9e9fe6801d4440d1bb01af
updated to official bazel.4.1.0
[ { "change_type": "MODIFY", "old_path": "build/build.py", "new_path": "build/build.py", "diff": "@@ -118,10 +118,10 @@ bazel_packages = {\n\"80c82e93a12ba30021692b11c78007807e82383a673be1602573b944beb359ab\"),\n(\"Darwin\", \"arm64\"):\nBazelPackage(\n- base_uri=\"https://releases.bazel.build/4.1.0/r...
Python
Apache License 2.0
google/jax
updated to official bazel.4.1.0
260,374
24.05.2021 17:31:20
0
369ca134fc6e7b80ed14c64862bb0746115b2130
add fingerprint to debugging log
[ { "change_type": "MODIFY", "old_path": "jax/interpreters/pxla.py", "new_path": "jax/interpreters/pxla.py", "diff": "@@ -59,6 +59,10 @@ from . import partial_eval as pe\nfrom . import xla\nfrom . import ad\n+# Built in Python lists don't support weak refs but subclasses of lists do.\n+class WeakRefLi...
Python
Apache License 2.0
google/jax
add fingerprint to debugging log
260,563
25.05.2021 18:00:46
-7,200
0308527f5514dd6dadddcad8aebe061b0892b886
Add auxiliary data support in custom_linear_solve
[ { "change_type": "MODIFY", "old_path": "jax/_src/lax/control_flow.py", "new_path": "jax/_src/lax/control_flow.py", "diff": "@@ -49,7 +49,7 @@ from jax._src.util import (partial, unzip2, unzip3, safe_map, safe_zip,\nsplit_list, cache, extend_name_stack)\nfrom jax.tree_util import (tree_flatten, tree_...
Python
Apache License 2.0
google/jax
Add auxiliary data support in custom_linear_solve
260,300
25.05.2021 16:07:09
25,200
00e030247ea3326d3d634e26e9a70baebc64184a
Enable custom gradients SavedModel option in test
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/tests/savedmodel_test.py", "new_path": "jax/experimental/jax2tf/tests/savedmodel_test.py", "diff": "@@ -35,7 +35,10 @@ class SavedModelTest(tf_test_util.JaxToTfTestCase):\ndef save_and_load_model(self, model: tf.Module) -> tf.Module:\n#...
Python
Apache License 2.0
google/jax
Enable custom gradients SavedModel option in test
260,364
13.05.2021 12:08:06
-3,600
03a1ee926974a8ac88c1863156f4a5e67ebb4d88
Update Jax linesearch to behave more like Scipy
[ { "change_type": "MODIFY", "old_path": "jax/_src/scipy/optimize/bfgs.py", "new_path": "jax/_src/scipy/optimize/bfgs.py", "diff": "@@ -55,6 +55,7 @@ class _BFGSResults(NamedTuple):\nf_k: jnp.ndarray\ng_k: jnp.ndarray\nH_k: jnp.ndarray\n+ old_old_fval: jnp.ndarray\nstatus: Union[int, jnp.ndarray]\nlin...
Python
Apache License 2.0
google/jax
Update Jax linesearch to behave more like Scipy
260,374
26.05.2021 19:03:28
0
b68f2c99fd61f518e853ea94868f888e60aeecfa
fixed fingerprint debugging message to be compatible with current min jaxlib version
[ { "change_type": "MODIFY", "old_path": "jax/interpreters/pxla.py", "new_path": "jax/interpreters/pxla.py", "diff": "@@ -898,7 +898,8 @@ def parallel_callable(fun: lu.WrappedFun,\ncompiled = xla.backend_compile(backend, built, compile_options)\nhandle_args = partial(shard_args, compiled.local_devices...
Python
Apache License 2.0
google/jax
fixed fingerprint debugging message to be compatible with current min jaxlib version
260,563
27.05.2021 18:25:18
-7,200
8226dfc8a4974b4c8031ee267fa5327e778140ee
Handle negative values for list-like sections in jnp.split
[ { "change_type": "MODIFY", "old_path": "jax/_src/numpy/lax_numpy.py", "new_path": "jax/_src/numpy/lax_numpy.py", "diff": "@@ -1805,7 +1805,8 @@ def _split(op, ary, indices_or_sections, axis=0):\nsize = ary.shape[axis]\nif isinstance(indices_or_sections, (tuple, list) + _arraylike_types):\nindices_or...
Python
Apache License 2.0
google/jax
Handle negative values for list-like sections in jnp.split
260,631
27.05.2021 20:33:18
25,200
44b1791b7aed840073d868e251359adefc3c0dd5
Copybara import of the project: by Nicholas Junge Handle negative values for list-like sections in jnp.split
[ { "change_type": "MODIFY", "old_path": "jax/_src/numpy/lax_numpy.py", "new_path": "jax/_src/numpy/lax_numpy.py", "diff": "@@ -1805,8 +1805,7 @@ def _split(op, ary, indices_or_sections, axis=0):\nsize = ary.shape[axis]\nif isinstance(indices_or_sections, (tuple, list) + _arraylike_types):\nindices_or...
Python
Apache License 2.0
google/jax
Copybara import of the project: -- 8226dfc8a4974b4c8031ee267fa5327e778140ee by Nicholas Junge <nicholas.junge@web.de>: Handle negative values for list-like sections in jnp.split PiperOrigin-RevId: 376302305
260,462
28.05.2021 17:45:30
-3,600
f30a36dbbcc86e9a4706aab1d23c85d6453d8bb9
Typo. Update Common_Gotchas_in_JAX.md
[ { "change_type": "MODIFY", "old_path": "docs/notebooks/Common_Gotchas_in_JAX.ipynb", "new_path": "docs/notebooks/Common_Gotchas_in_JAX.ipynb", "diff": "\"source\": [\n\"JAX transformation and compilation are designed to work only on Python functions that are functionally pure: all the input data is ...
Python
Apache License 2.0
google/jax
Typo. Update Common_Gotchas_in_JAX.md
260,435
31.05.2021 22:18:24
-32,400
160c3e935780d717cad5eee28a40925722237d6b
rename v to vh
[ { "change_type": "MODIFY", "old_path": "jax/_src/numpy/linalg.py", "new_path": "jax/_src/numpy/linalg.py", "diff": "@@ -329,12 +329,12 @@ def pinv(a, rcond=None):\nmax_rows_cols = max(a.shape[-2:])\nrcond = 10. * max_rows_cols * jnp.finfo(a.dtype).eps\nrcond = jnp.asarray(rcond)\n- u, s, v = svd(a, ...
Python
Apache License 2.0
google/jax
rename v to vh
260,411
01.06.2021 14:32:59
-10,800
973171bb6d76e1ae25a00b47e2eed3e1e0f204ca
[jax2tf] Add support for pjit.
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/jax2tf.py", "new_path": "jax/experimental/jax2tf/jax2tf.py", "diff": "@@ -31,6 +31,8 @@ from jax._src.lax import lax\nfrom jax._src.lax import linalg as lax_linalg\nimport jax._src.random\nfrom jax.api_util import flatten_fun\n+from jax...
Python
Apache License 2.0
google/jax
[jax2tf] Add support for pjit.
260,372
01.06.2021 15:28:25
21,600
c28b472ae2eff54fda8c69a4322016bc414cb4ea
Add test to ensure gradient is finite
[ { "change_type": "MODIFY", "old_path": "tests/image_test.py", "new_path": "tests/image_test.py", "diff": "@@ -324,6 +324,41 @@ class ImageTest(jtu.JaxTestCase):\noutput = jax.jit(jit_fn)(x, scale_a, translation_a)\nself.assertAllClose(output, expected, atol=2e-03)\n+ @parameterized.named_parameters(...
Python
Apache License 2.0
google/jax
Add test to ensure gradient is finite
260,623
01.06.2021 19:14:24
25,200
dca61aa78ff98101ac30737148f6dce03f85c412
address comments and update documentation
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/README.md", "new_path": "jax/experimental/jax2tf/README.md", "diff": "@@ -137,8 +137,12 @@ custom gradients and wrap instead the converted function with\n``tf.raw_ops.PreventGradient`` to generated an error in case a gradient\ncomputati...
Python
Apache License 2.0
google/jax
address comments and update documentation
260,456
02.06.2021 04:03:06
0
012da545f734663c98a0c62d1790ff6266cd4c63
add gpu to the rocsolver backend
[ { "change_type": "MODIFY", "old_path": "jax/_src/lax/linalg.py", "new_path": "jax/_src/lax/linalg.py", "diff": "@@ -1029,7 +1029,7 @@ if cusolver is not None:\nif rocsolver is not None:\nxla.backend_specific_translations['gpu'][lu_p] = partial(\n- _lu_cpu_gpu_translation_rule, rocsolver.getrf)\n+ _l...
Python
Apache License 2.0
google/jax
add gpu to the rocsolver backend
260,411
02.06.2021 12:17:03
-10,800
ccc1ba701a3f30eaf9028b1adac386fe5a8421d6
Update savedmodel_test.py I added a cast to `float32`. This is needed because of an obscure bug in JAX
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/tests/savedmodel_test.py", "new_path": "jax/experimental/jax2tf/tests/savedmodel_test.py", "diff": "@@ -83,7 +83,7 @@ class SavedModelTest(tf_test_util.JaxToTfTestCase):\nx, = primals\nx_dot, = tangents\nprimal_out = f_jax(x)\n- tangent...
Python
Apache License 2.0
google/jax
Update savedmodel_test.py I added a cast to `float32`. This is needed because of an obscure bug in JAX (#6874).
260,411
02.06.2021 13:28:48
-10,800
6219b1947e31652c3f3f29034cb2b233804e9b54
Update savedmodel_test.py Minor change, just to trigger a copybara re-import.
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/tests/savedmodel_test.py", "new_path": "jax/experimental/jax2tf/tests/savedmodel_test.py", "diff": "@@ -36,8 +36,7 @@ class SavedModelTest(tf_test_util.JaxToTfTestCase):\n# Roundtrip through saved model on disk.\nmodel_dir = os.path.joi...
Python
Apache License 2.0
google/jax
Update savedmodel_test.py Minor change, just to trigger a copybara re-import.
260,447
02.06.2021 11:37:37
25,200
a02bf592331503258134613ccd7fb297c3f25a1d
Adds associated Legendre functions of the first kind.
[ { "change_type": "MODIFY", "old_path": "docs/jax.scipy.rst", "new_path": "docs/jax.scipy.rst", "diff": "@@ -99,6 +99,7 @@ jax.scipy.special\nlog_ndtr\nlogit\nlogsumexp\n+ lpmn\nmultigammaln\nndtr\nndtri\n" }, { "change_type": "MODIFY", "old_path": "jax/scipy/special.py", "new_path": ...
Python
Apache License 2.0
google/jax
Adds associated Legendre functions of the first kind. Co-authored-by: Jake VanderPlas <jakevdp@google.com>
260,287
02.06.2021 14:02:47
25,200
ed96e5305f94788d440ab58f7485b9eb5dd40023
Fix incorrect handling of axis_index_groups in parallel primitive fallbacks
[ { "change_type": "MODIFY", "old_path": "jax/_src/lax/parallel.py", "new_path": "jax/_src/lax/parallel.py", "diff": "@@ -817,11 +817,18 @@ def _foldaxis(axis, x):\nnew_shape[axis:axis+2] = [x.shape[axis] * x.shape[axis + 1]]\nreturn x.reshape(new_shape)\n+def _index_in_group(axis_name, axis_index_gro...
Python
Apache License 2.0
google/jax
Fix incorrect handling of axis_index_groups in parallel primitive fallbacks PiperOrigin-RevId: 377139424
260,411
02.06.2021 12:08:15
-10,800
3e9b13b011c6853037c607de570e980253a94c6d
Expanded the type promotion documentation with a confusing case I filed this as a bug, but I am assuming that it is not easy to fix, so I also change the documentation.
[ { "change_type": "MODIFY", "old_path": "docs/type_promotion.rst", "new_path": "docs/type_promotion.rst", "diff": "@@ -130,11 +130,14 @@ Jax's type promotion rules differ from those of NumPy, as given by\n:func:`numpy.promote_types`, in those cells highlighted with a green background\nin the table ab...
Python
Apache License 2.0
google/jax
Expanded the type promotion documentation with a confusing case I filed this as a bug, but I am assuming that it is not easy to fix, so I also change the documentation. Bug: 6874
260,287
03.06.2021 04:13:02
25,200
bca3d61b3b0dc917aff1e5e70346dc60f1119a47
Insert xmap SPMD axes into pjit sharding annotations This should let us emit good XLA annotations for `xmap(pjit)`. Previously we might have been overestimating the set of replicated mesh dimensions.
[ { "change_type": "MODIFY", "old_path": "jax/experimental/maps.py", "new_path": "jax/experimental/maps.py", "diff": "@@ -24,7 +24,6 @@ from functools import wraps, partial, partialmethod\nfrom .. import numpy as jnp\nfrom .. import core\n-from .. import config\nfrom .. import linear_util as lu\nfrom ...
Python
Apache License 2.0
google/jax
Insert xmap SPMD axes into pjit sharding annotations This should let us emit good XLA annotations for `xmap(pjit)`. Previously we might have been overestimating the set of replicated mesh dimensions. PiperOrigin-RevId: 377259226
260,287
03.06.2021 06:36:57
25,200
c7a98b3b6216b4fa764e94998cad09f14d488094
Fix a typo in shape checks for associative_scan Fixes
[ { "change_type": "MODIFY", "old_path": "jax/_src/lax/control_flow.py", "new_path": "jax/_src/lax/control_flow.py", "diff": "@@ -2510,7 +2510,7 @@ def associative_scan(fn: Callable, elems, reverse: bool = False, axis: int = 0):\nif not all(int(elem.shape[axis]) == num_elems for elem in elems_flat[1:]...
Python
Apache License 2.0
google/jax
Fix a typo in shape checks for associative_scan Fixes #6884. PiperOrigin-RevId: 377276183
260,411
04.06.2021 15:29:54
-10,800
ede457f1a5d93136c8a87c0afbf6ab5ee757e677
[jax2tf] Fix bug with max_int for uint64
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/jax2tf.py", "new_path": "jax/experimental/jax2tf/jax2tf.py", "diff": "@@ -76,8 +76,10 @@ PrecisionType = int # Enum xla_data.PrecisionConfig.Precision\ndef _is_tfval(v: TfVal) -> bool:\n+ if isinstance(v, (tf.Tensor, tf.Variable)):\n+ r...
Python
Apache License 2.0
google/jax
[jax2tf] Fix bug with max_int for uint64
260,411
04.06.2021 11:02:50
-10,800
d243258b86fc5e33405925acc794801b65fa0137
[jax2tf] Implement inequalities and friends for complex numbers. This requires re-using JAX's lowering rule for comparisons of complex numbers to use lexicographic comparison.
[ { "change_type": "MODIFY", "old_path": "CHANGELOG.md", "new_path": "CHANGELOG.md", "diff": "@@ -16,6 +16,9 @@ PLEASE REMEMBER TO CHANGE THE '..master' WITH AN ACTUAL TAG in GITHUB LINK.\ntracebacks.\n* A new traceback filtering mode using `__tracebackhide__` is now enabled by\ndefault in sufficientl...
Python
Apache License 2.0
google/jax
[jax2tf] Implement inequalities and friends for complex numbers. This requires re-using JAX's lowering rule for comparisons of complex numbers to use lexicographic comparison.
260,697
04.06.2021 17:26:27
25,200
f0e55e3ce241f997b097bd9381504c08a9d35cf8
move #endif so that Windows doesn't have GetXCR0EAX defined twice (erroring out)
[ { "change_type": "MODIFY", "old_path": "jaxlib/cpu_feature_guard.c", "new_path": "jaxlib/cpu_feature_guard.c", "diff": "@@ -58,7 +58,6 @@ static int GetXCR0EAX() { return _xgetbv(0); }\n\"xchg %%rdi, %%rbx\\n\" \\\n: \"=a\"(a), \"=D\"(b), \"=c\"(c), \"=d\"(d) \\\n: \"a\"(a_inp), \"2\"(c_inp))\n-#end...
Python
Apache License 2.0
google/jax
move #endif so that Windows doesn't have GetXCR0EAX defined twice (erroring out)
260,374
01.06.2021 18:51:44
0
75cc734c8e335d26f6baa4f79dcb05849dd91bf9
completed FilesystemCache class with corresponding unit tests
[ { "change_type": "ADD", "old_path": null, "new_path": "jax/experimental/compilation_cache/file_system_cache.py", "diff": "+# Copyright 2021 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+# Yo...
Python
Apache License 2.0
google/jax
completed FilesystemCache class with corresponding unit tests
260,287
07.06.2021 12:43:36
0
490f9778c8b465b79e518feac5b9e6326982b439
Raise a friendlier error message when using loop axes in collectives
[ { "change_type": "MODIFY", "old_path": "jax/_src/lax/parallel.py", "new_path": "jax/_src/lax/parallel.py", "diff": "@@ -543,7 +543,7 @@ def pgather(src, idx, axes: Union[int, AxisName]):\n### parallel primitives\ndef _subst_all_names_in_param(\n- pname: str, params: core.ParamDict, subst: core.AxisS...
Python
Apache License 2.0
google/jax
Raise a friendlier error message when using loop axes in collectives
260,287
15.03.2021 13:30:44
0
54ba051631c0f89d607d4f6f8896b845e620387c
Always run xmap over all mesh axes Automatic mesh slicing might be surprising, and can always be performed manually.
[ { "change_type": "MODIFY", "old_path": "jax/experimental/maps.py", "new_path": "jax/experimental/maps.py", "diff": "@@ -635,12 +635,11 @@ def make_xmap_callable(fun: lu.WrappedFun,\nused_mesh_axes = used_resources & resource_env.physical_resource_axes\nif used_mesh_axes:\nassert spmd_in_axes is None...
Python
Apache License 2.0
google/jax
Always run xmap over all mesh axes Automatic mesh slicing might be surprising, and can always be performed manually.
260,411
10.06.2021 17:01:22
-7,200
1994f6df4a7a7a53af1fb9e21800973bf541632f
[jax2tf] Fix the round-trip call_tf(convert) Also cleaned the handling of global state in jax2tf.
[ { "change_type": "MODIFY", "old_path": "CHANGELOG.md", "new_path": "CHANGELOG.md", "diff": "@@ -8,9 +8,15 @@ Remember 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 GITHUB LINK.\n-->\n-## jax 0.2.14 (unreleased)...
Python
Apache License 2.0
google/jax
[jax2tf] Fix the round-trip call_tf(convert) Also cleaned the handling of global state in jax2tf.
260,411
12.06.2021 11:42:15
-10,800
edd96884c7309adef01d9242dbe3b71a8e983039
[jax2tf] Extend shape polymorphism to handle add_transpose with broadcasting
[ { "change_type": "MODIFY", "old_path": "jax/_src/lax/lax.py", "new_path": "jax/_src/lax/lax.py", "diff": "@@ -656,7 +656,7 @@ def dot(lhs: Array, rhs: Array, precision: PrecisionLike = None,\nReturns:\nAn array containing the product.\n\"\"\"\n- if 1 <= lhs.ndim <= 2 and 1 <= rhs.ndim <= 2 and lhs.s...
Python
Apache License 2.0
google/jax
[jax2tf] Extend shape polymorphism to handle add_transpose with broadcasting
260,411
10.06.2021 12:42:40
-7,200
dd8ab851214b855dd5073c1b68b1f3e89ab3cb85
[jax2tf] Support inequality and min/max for booleans. For inequalities we add casts to int8. For min/max we rewrite to logical operations and/or.
[ { "change_type": "MODIFY", "old_path": "CHANGELOG.md", "new_path": "CHANGELOG.md", "diff": "@@ -11,6 +11,8 @@ PLEASE REMEMBER TO CHANGE THE '..master' WITH AN ACTUAL TAG in GITHUB LINK.\n## jax 0.2.15 (unreleased)\n* [GitHub commits](https://github.com/google/jax/compare/jax-v0.2.14...master).\n* Ne...
Python
Apache License 2.0
google/jax
[jax2tf] Support inequality and min/max for booleans. For inequalities we add casts to int8. For min/max we rewrite to logical operations and/or.
260,411
13.06.2021 17:58:22
-10,800
07cc58122dc0a1b063d2b66618846a63d2163d6e
[jax2tf] Change the InconclusiveDimensionOperation error to include link to documentation
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/README.md", "new_path": "jax/experimental/jax2tf/README.md", "diff": "@@ -300,8 +300,8 @@ A few examples of shape specifications and uses:\n### Computing with dimension variables\n-JAX keeps track of the shape of all intermediate result...
Python
Apache License 2.0
google/jax
[jax2tf] Change the InconclusiveDimensionOperation error to include link to documentation
260,447
14.06.2021 14:51:37
25,200
095e6507b9f8c69c3a5f4c0fcd68033dae428bd6
Support value computation of associated Legendre functions.
[ { "change_type": "MODIFY", "old_path": "docs/jax.scipy.rst", "new_path": "docs/jax.scipy.rst", "diff": "@@ -100,6 +100,7 @@ jax.scipy.special\nlogit\nlogsumexp\nlpmn\n+ lpmn_values\nmultigammaln\nndtr\nndtri\n" }, { "change_type": "MODIFY", "old_path": "jax/scipy/special.py", "new_pa...
Python
Apache License 2.0
google/jax
Support value computation of associated Legendre functions. Co-authored-by: Jake VanderPlas <jakevdp@google.com>
260,411
15.06.2021 11:13:20
-10,800
2888e7ca81a90e989e79af0d640def549c8bd779
[jax2tf] Add more documentation about saving models with custom gradients
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/README.md", "new_path": "jax/experimental/jax2tf/README.md", "diff": "@@ -95,7 +95,8 @@ is trivial:\nmy_model = tf.Module()\n# Save a function that can take scalar inputs.\nmy_model.f = tf.function(jax2tf.convert(f_jax), input_signature...
Python
Apache License 2.0
google/jax
[jax2tf] Add more documentation about saving models with custom gradients
260,411
15.06.2021 15:23:27
-10,800
bfb3bbf859e74c172e1d6e25d1e6e24f2ab58024
[jax2tf] Fix bug in gradient for functions with integer results The bug is really a peculiarity of tf.function, which uses None for the cotangents corresponding to non-float results. This does not happen in TF eager. Fixes:
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/jax2tf.py", "new_path": "jax/experimental/jax2tf/jax2tf.py", "diff": "@@ -300,6 +300,18 @@ def convert(fun: Callable,\ndef converted_grad_fn(*out_cts_flat: TfVal,\n_out_cts_avals: Sequence[core.AbstractValue],\nvariables=None):\n+ # If ...
Python
Apache License 2.0
google/jax
[jax2tf] Fix bug in gradient for functions with integer results The bug is really a peculiarity of tf.function, which uses None for the cotangents corresponding to non-float results. This does not happen in TF eager. Fixes: #6975
260,287
08.06.2021 11:53:23
0
1e288b2f422299129e10a1cde9f9c9bdd043d995
Add support for anonymous serial loops
[ { "change_type": "MODIFY", "old_path": "jax/experimental/maps.py", "new_path": "jax/experimental/maps.py", "diff": "@@ -82,13 +82,13 @@ AxisName = core.AxisName\nResourceAxisName = AxisName # Different name just for documentation purposes\nMesh = pxla.Mesh\n-class Loop(NamedTuple):\n+class _Loop(Nam...
Python
Apache License 2.0
google/jax
Add support for anonymous serial loops
260,411
16.06.2021 10:20:24
-10,800
9831371e389f9226f34a53fc91790b1fc4384226
[jax2tf] Fix the custom gradients for call_tf The previous implementation did not work correctly when the call_tf function was embedded in a larger JAX function for which we take the gradient. Also fixed handling of non-float args/results.
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/call_tf.py", "new_path": "jax/experimental/jax2tf/call_tf.py", "diff": "@@ -115,40 +115,42 @@ def call_tf(func_tf: Callable) -> Callable:\n# Define the fwd and bwd custom_vjp functions\ndef make_call_vjp_fwd(*args_jax):\n- # Return the ...
Python
Apache License 2.0
google/jax
[jax2tf] Fix the custom gradients for call_tf The previous implementation did not work correctly when the call_tf function was embedded in a larger JAX function for which we take the gradient. Also fixed handling of non-float args/results.