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
260,578
13.08.2021 09:03:08
25,200
3a99d18df845566af1b41a24dd65551610ddcfed
Add nocuda jaxlib wheels to the index as well
[ { "change_type": "MODIFY", "old_path": "build/generate_release_indexes.py", "new_path": "build/generate_release_indexes.py", "diff": "@@ -41,7 +41,7 @@ def get_entries(gcs_uri, whl_filter=\".whl\"):\nls_output = subprocess.check_output([\"gsutil\", \"ls\", gcs_uri])\nfor line in ls_output.decode(\"u...
Python
Apache License 2.0
google/jax
Add nocuda jaxlib wheels to the index as well
260,578
13.08.2021 11:11:50
25,200
2ab264dc8543d2d734e406228811f1f6b2302d6c
Update workspace org_tensorflow commit
[ { "change_type": "MODIFY", "old_path": "WORKSPACE", "new_path": "WORKSPACE", "diff": "@@ -7,10 +7,10 @@ load(\"@bazel_tools//tools/build_defs/repo:http.bzl\", \"http_archive\")\n# and update the sha256 with the result.\nhttp_archive(\nname = \"org_tensorflow\",\n- sha256 = \"0f13410284b9186e436350e9...
Python
Apache License 2.0
google/jax
Update workspace org_tensorflow commit
260,335
13.08.2021 14:47:45
25,200
2e6a30a595b358c06afcd183621ce5af23dda625
always use same object for vmap temp axis name
[ { "change_type": "MODIFY", "old_path": "jax/_src/api.py", "new_path": "jax/_src/api.py", "diff": "@@ -1256,8 +1256,6 @@ def vmap(fun: F, in_axes=0, out_axes=0, axis_name=None) -> F:\ndocstr += \"\\n\\nOriginal documentation:\\n\\n\"\ndocstr += fun.__doc__\n- axis_name = core._TempAxisName(fun) if ax...
Python
Apache License 2.0
google/jax
always use same object for vmap temp axis name
260,456
16.08.2021 04:37:18
0
16a110e4ff7342aaac8c93d5c5204576c6e794f7
fix custom_call_status for rocm
[ { "change_type": "MODIFY", "old_path": "jaxlib/rocm_gpu_kernel_helpers.cc", "new_path": "jaxlib/rocm_gpu_kernel_helpers.cc", "diff": "@@ -42,6 +42,6 @@ absl::StatusOr<std::unique_ptr<void* []>> MakeBatchPointers(\nJAX_RETURN_IF_ERROR(\nAsStatus(hipMemcpyAsync(dev_ptrs, host_ptrs.get(), sizeof(void*)...
Python
Apache License 2.0
google/jax
fix custom_call_status for rocm
260,411
17.08.2021 14:37:54
-7,200
663ba16694b5f2f1b3541bcc1b2b3055351d4d45
[jax2tf] Update conversion of scatter for shape polymorphism It turns out that the XlaScatter inference rule is enough, we don't need the `.set_shape` workaround.
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/jax2tf.py", "new_path": "jax/experimental/jax2tf/jax2tf.py", "diff": "@@ -2519,8 +2519,6 @@ def _scatter(operand, scatter_indices, updates, *, update_jaxpr, update_consts,\nxla_update_computation,\nproto,\nindices_are_sorted=indices_are...
Python
Apache License 2.0
google/jax
[jax2tf] Update conversion of scatter for shape polymorphism It turns out that the XlaScatter inference rule is enough, we don't need the `.set_shape` workaround.
260,285
17.08.2021 15:20:13
-7,200
6d83027b69b21f8c3df6ec49feaa3ed9017ada9b
Support scipy.fft.dct/dctn type=2
[ { "change_type": "MODIFY", "old_path": "docs/jax.scipy.rst", "new_path": "docs/jax.scipy.rst", "diff": "jax.scipy package\n=================\n+jax.scipy.fft\n+-------------\n+\n+.. automodule:: jax.scipy.fft\n+\n+.. autosummary::\n+ :toctree: _autosummary\n+\n+ dct\n+ dctn\n+\njax.scipy.linalg\n----...
Python
Apache License 2.0
google/jax
Support scipy.fft.dct/dctn type=2
260,456
17.08.2021 20:42:02
0
f454f6b7b8db328f7fd6a23f693bbbddf434ca85
fix rocm_amdgpu_targets for rocm
[ { "change_type": "MODIFY", "old_path": ".bazelrc", "new_path": ".bazelrc", "diff": "@@ -46,7 +46,7 @@ build:native_arch_posix --host_copt=-march=native\nbuild:mkl_open_source_only --define=tensorflow_mkldnn_contraction_kernel=1\nbuild:cuda --repo_env TF_NEED_CUDA=1\n-build:cuda --action_env=TF_CUDA_...
Python
Apache License 2.0
google/jax
fix rocm_amdgpu_targets for rocm
260,335
17.08.2021 16:18:57
25,200
b90daf9cdad16c13bf31b9947d762e1430912162
custom_vjp: automatically handle float0 cotangents
[ { "change_type": "MODIFY", "old_path": "jax/_src/custom_derivatives.py", "new_path": "jax/_src/custom_derivatives.py", "diff": "@@ -553,8 +553,10 @@ def _flatten_bwd(in_tree, in_avals, out_trees, *args):\n\"number of arguments to the primal function, but got VJP output \"\n\"structure {} for primal ...
Python
Apache License 2.0
google/jax
custom_vjp: automatically handle float0 cotangents
260,335
17.08.2021 17:06:19
25,200
83f95a5dae19124b0f396ba7608c7701a617ccbd
custom_jvp/vjp tweaks and fixes
[ { "change_type": "MODIFY", "old_path": "jax/_src/custom_derivatives.py", "new_path": "jax/_src/custom_derivatives.py", "diff": "@@ -242,6 +242,7 @@ def _flatten_jvp(in_tree, *args):\nmsg = (\"Custom JVP rule must produce primal and tangent outputs with equal \"\n\"container (pytree) structures, but ...
Python
Apache License 2.0
google/jax
custom_jvp/vjp tweaks and fixes
260,296
17.08.2021 20:41:02
25,200
b65f39ca7ade1d50bf8fa22f4ec282ab5163c1ca
Default to `jnp.float_` type in `nn.initializers`.
[ { "change_type": "MODIFY", "old_path": "jax/_src/nn/initializers.py", "new_path": "jax/_src/nn/initializers.py", "diff": "@@ -28,17 +28,20 @@ from jax import ops\nfrom jax import random\nfrom jax import core\nfrom jax._src.util import prod\n+from jax import dtypes\n-def zeros(key, shape, dtype=jnp.f...
Python
Apache License 2.0
google/jax
Default to `jnp.float_` type in `nn.initializers`.
260,411
18.08.2021 16:31:09
-7,200
bc1e41438b5a5ea72ba819684c5b67f29cbedc13
[jax2tf] Added support for conversion of padding with enable_xla=False
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/jax2tf.py", "new_path": "jax/experimental/jax2tf/jax2tf.py", "diff": "@@ -230,9 +230,12 @@ def convert(fun: Callable,\nin a SavedModel, the custom gradients are currently lost and an error will\nbe raised if a gradient computation is at...
Python
Apache License 2.0
google/jax
[jax2tf] Added support for conversion of padding with enable_xla=False
260,635
31.07.2021 19:26:53
-7,200
5138743e8e15d2a565f7cecb23a91e5a0dd11d93
Implement variance scaling initializers with complex dtype
[ { "change_type": "MODIFY", "old_path": "jax/_src/nn/initializers.py", "new_path": "jax/_src/nn/initializers.py", "diff": "@@ -48,7 +48,59 @@ def _compute_fans(shape: core.NamedShape, in_axis=-2, out_axis=-1):\nfan_out = shape[out_axis] * receptive_field_size\nreturn fan_in, fan_out\n+def _complex_un...
Python
Apache License 2.0
google/jax
Implement variance scaling initializers with complex dtype
260,688
19.08.2021 16:54:53
25,200
6b00b4480727d7d4877f279d6cf2d58cfad42b94
Move all Abseil dependencies out of jaxlib CUDA libraries These were breaking the build with CUDA 10.2
[ { "change_type": "MODIFY", "old_path": "jaxlib/BUILD", "new_path": "jaxlib/BUILD", "diff": "@@ -247,10 +247,6 @@ cuda_library(\nsrcs = [\"cuda_lu_pivot_kernels.cu.cc\"],\nhdrs = [\"cuda_lu_pivot_kernels.h\"],\ndeps = [\n- \":cuda_gpu_kernel_helpers\",\n- \":kernel_helpers\",\n- \"@org_tensorflow//te...
Python
Apache License 2.0
google/jax
Move all Abseil dependencies out of jaxlib CUDA libraries These were breaking the build with CUDA 10.2 PiperOrigin-RevId: 391875083
260,335
20.08.2021 13:43:38
25,200
e416e8730165bb2308dd0a69fbeeb5a32ca09eb8
inline jit-decorated jax.random calls
[ { "change_type": "MODIFY", "old_path": "jax/_src/prng.py", "new_path": "jax/_src/prng.py", "diff": "@@ -371,7 +371,7 @@ if cuda_prng:\n_threefry2x32_gpu_translation_rule\n-@jit\n+@partial(jit, inline=True)\ndef threefry_2x32(keypair, count):\n\"\"\"Apply the Threefry 2x32 hash.\n@@ -408,7 +408,7 @@ ...
Python
Apache License 2.0
google/jax
inline jit-decorated jax.random calls
260,282
04.07.2021 19:55:45
-28,800
85367804552c149c96b09986ccfd9a96c20a5735
Add documentation for jax.device_get
[ { "change_type": "MODIFY", "old_path": "docs/jax.rst", "new_path": "docs/jax.rst", "diff": "@@ -36,6 +36,7 @@ Just-in-time compilation (:code:`jit`)\ndevice_put\ndevice_put_replicated\ndevice_put_sharded\n+ device_get\ndefault_backend\nnamed_call\n@@ -91,6 +92,7 @@ Parallelization (:code:`pmap`)\n.....
Python
Apache License 2.0
google/jax
Add documentation for jax.device_get
260,628
21.08.2021 16:03:43
14,400
fec72e1852cc09ddf499b3b3d388300e87d4f116
add support for scipy.special.{expn,expi,exp1}
[ { "change_type": "MODIFY", "old_path": "docs/jax.scipy.rst", "new_path": "docs/jax.scipy.rst", "diff": "@@ -99,7 +99,10 @@ jax.scipy.special\nerf\nerfc\nerfinv\n+ exp1\n+ expi\nexpit\n+ expn\ngammainc\ngammaincc\ngammaln\n" }, { "change_type": "MODIFY", "old_path": "jax/_src/scipy/specia...
Python
Apache License 2.0
google/jax
add support for scipy.special.{expn,expi,exp1}
260,510
24.08.2021 17:46:34
25,200
0fa70084a395ddd5ec2bd18250a73f9e3f7a6188
Pass `x` into transpose in autodidax
[ { "change_type": "MODIFY", "old_path": "docs/autodidax.ipynb", "new_path": "docs/autodidax.ipynb", "diff": "\"def reduce_sum(x, axis=None): return bind1(reduce_sum_p, x, axis=axis)\\n\",\n\"def greater(x, y): return bind1(greater_p, x, y)\\n\",\n\"def less(x, y): return bind1(less_p, x, y)\\n\",\n- ...
Python
Apache License 2.0
google/jax
Pass `x` into transpose in autodidax
260,296
25.08.2021 12:50:28
25,200
00435bd9002e8840064640c9641150c0714635fc
Make `input_shape` explicitly integer-typed in `_reduce_jvp` to avoid accidental float shapes when primal shape is `()`. Example where this fails: `grad(jnp.prod)(jnp.ones(()))`. Add a test.
[ { "change_type": "MODIFY", "old_path": "jax/_src/lax/lax.py", "new_path": "jax/_src/lax/lax.py", "diff": "@@ -5353,7 +5353,7 @@ def _reduction_computation(c, jaxpr, consts, init_values, singleton=True):\nreturn subc.build(out_nodes)\ndef _reduce_jvp(reducer, init_values, primals, tangents, axes):\n-...
Python
Apache License 2.0
google/jax
Make `input_shape` explicitly integer-typed in `_reduce_jvp` to avoid accidental float shapes when primal shape is `()`. Example where this fails: `grad(jnp.prod)(jnp.ones(()))`. Add a test.
260,335
25.08.2021 20:46:11
25,200
2d28951ba4705f41d0f18d9657dbc82b759240f8
address comments form
[ { "change_type": "MODIFY", "old_path": "jax/__init__.py", "new_path": "jax/__init__.py", "diff": "@@ -45,6 +45,7 @@ from ._src.config import (\nfrom ._src.api import (\nad, # TODO(phawkins): update users to avoid this.\ncheckpoint,\n+ checkpoint_policies,\nclosure_convert,\ncurry, # TODO(phawkins): ...
Python
Apache License 2.0
google/jax
address comments form @apaszke
260,335
27.08.2021 16:59:54
25,200
a56a1679ab88dd2340429a5bcd0d91bd2186e6aa
fix constants problem
[ { "change_type": "MODIFY", "old_path": "jax/_src/traceback_util.py", "new_path": "jax/_src/traceback_util.py", "diff": "@@ -23,6 +23,7 @@ from jax._src import util\n_exclude_paths = [__file__, util.__file__]\ndef register_exclusion(path):\n+ return\n_exclude_paths.append(path)\n_jax_message_append =...
Python
Apache License 2.0
google/jax
fix constants problem
260,335
27.08.2021 17:42:42
25,200
9f919c69e43d6b967a040b16678814bd8dd78c32
fix custom_vjp issue?
[ { "change_type": "MODIFY", "old_path": "jax/_src/custom_derivatives.py", "new_path": "jax/_src/custom_derivatives.py", "diff": "@@ -681,6 +681,9 @@ xla.initial_style_translations[custom_vjp_call_jaxpr_p] = \\\nbatching.primitive_batchers[ad.custom_lin_p] = ad._raise_custom_vjp_error_on_jvp\nxla.tran...
Python
Apache License 2.0
google/jax
fix custom_vjp issue?
260,335
30.08.2021 19:22:43
25,200
d69f57e1ed3a8fdd05bd2e844524cc612cff0a53
raise NotImplementedError for pjit partial eval
[ { "change_type": "MODIFY", "old_path": "jax/experimental/pjit.py", "new_path": "jax/experimental/pjit.py", "diff": "@@ -519,6 +519,13 @@ def _pjit_jvp(primals_in, tangents_in,\nad.primitive_jvps[pjit_p] = _pjit_jvp\n+def _pjit_partial_eval(trace, *in_tracers, jaxpr, in_axis_resources,\n+ out_axis_re...
Python
Apache License 2.0
google/jax
raise NotImplementedError for pjit partial eval
260,289
29.08.2021 16:36:45
-19,080
c4d300c4caa860c6c0ee61842e60194c49959880
Implement np.polyfit
[ { "change_type": "MODIFY", "old_path": "docs/jax.numpy.rst", "new_path": "docs/jax.numpy.rst", "diff": "@@ -294,6 +294,7 @@ Not every function in NumPy is implemented; contributions are welcome!\npoly\npolyadd\npolyder\n+ polyfit\npolyint\npolymul\npolysub\n" }, { "change_type": "MODIFY", ...
Python
Apache License 2.0
google/jax
Implement np.polyfit
260,578
01.09.2021 07:14:48
25,200
67bea579abcaf222739788a3bb4827d82323f693
Update org_tensorflow commit for jaxlib release
[ { "change_type": "MODIFY", "old_path": "WORKSPACE", "new_path": "WORKSPACE", "diff": "@@ -7,10 +7,10 @@ load(\"@bazel_tools//tools/build_defs/repo:http.bzl\", \"http_archive\")\n# and update the sha256 with the result.\nhttp_archive(\nname = \"org_tensorflow\",\n- sha256 = \"8e1b21a456b8d1f245910e56...
Python
Apache License 2.0
google/jax
Update org_tensorflow commit for jaxlib release
260,578
01.09.2021 10:43:20
25,200
be824a792edeb1908f21264b5d50dd3dbfefe5b3
Update files after new jaxlib release 0.1.71
[ { "change_type": "MODIFY", "old_path": "CHANGELOG.md", "new_path": "CHANGELOG.md", "diff": "@@ -16,6 +16,11 @@ PLEASE REMEMBER TO CHANGE THE '..main' WITH AN ACTUAL TAG in GITHUB LINK.\n* `jnp.unique` and other set-like operations now require array-like inputs\n({jax-issue}`#7662`)\n+## jaxlib 0.1.7...
Python
Apache License 2.0
google/jax
Update files after new jaxlib release 0.1.71
260,578
01.09.2021 10:56:54
25,200
84edde2f9b611d4323257faddd284e3692a20d36
Add new features section
[ { "change_type": "MODIFY", "old_path": "CHANGELOG.md", "new_path": "CHANGELOG.md", "diff": "@@ -20,6 +20,8 @@ PLEASE REMEMBER TO CHANGE THE '..main' WITH AN ACTUAL TAG in GITHUB LINK.\n* Breaking changes:\n* Support for CUDA 11.0 and CUDA 10.1 has been dropped. Jaxlib now supports\nCUDA 10.2 and CUD...
Python
Apache License 2.0
google/jax
Add new features section
260,578
01.09.2021 11:26:41
25,200
14a02c68800f67e8d0762767a6e6152054353b8a
Remove new features
[ { "change_type": "MODIFY", "old_path": "CHANGELOG.md", "new_path": "CHANGELOG.md", "diff": "@@ -20,8 +20,6 @@ PLEASE REMEMBER TO CHANGE THE '..main' WITH AN ACTUAL TAG in GITHUB LINK.\n* Breaking changes:\n* Support for CUDA 11.0 and CUDA 10.1 has been dropped. Jaxlib now supports\nCUDA 10.2 and CUD...
Python
Apache License 2.0
google/jax
Remove new features
260,439
01.09.2021 13:08:10
25,200
f38d3e8735bb3b0afb5f49d3cc49672cedc4d554
Allow axis index groups to have different sizes for AllReduce.
[ { "change_type": "MODIFY", "old_path": "jax/_src/lax/parallel.py", "new_path": "jax/_src/lax/parallel.py", "diff": "@@ -57,7 +57,7 @@ def psum(x, axis_name, *, axis_index_groups=None):\naxis_index_groups: optional list of lists containing axis indices (e.g. for\nan axis of size 4, [[0, 1], [2, 3]] w...
Python
Apache License 2.0
google/jax
Allow axis index groups to have different sizes for AllReduce. PiperOrigin-RevId: 394297426
260,335
01.09.2021 22:38:17
25,200
ffa4ec05007ba2bf03bc36ce961603898647e803
[remat] fix two unit bugs, add test for one
[ { "change_type": "MODIFY", "old_path": "jax/interpreters/partial_eval.py", "new_path": "jax/interpreters/partial_eval.py", "diff": "@@ -798,8 +798,9 @@ def _remat_partial_eval(trace, _, f, tracers, params):\nout_consts = core.eval_jaxpr(jaxpr_known, (), *in_consts)\nout_consts_ = iter(out_consts)\n#...
Python
Apache License 2.0
google/jax
[remat] fix two unit bugs, add test for one
260,335
02.09.2021 15:41:39
25,200
9955e44653a734affc4f2e4818be621486bba8b5
fix dce logic for nullary primitives
[ { "change_type": "MODIFY", "old_path": "jax/interpreters/partial_eval.py", "new_path": "jax/interpreters/partial_eval.py", "diff": "@@ -1015,9 +1015,11 @@ def dce_jaxpr(jaxpr: Jaxpr, used_outputs: List[bool]\nif rule:\nused_ins, new_eqn = rule(used_outs, eqn)\nif any(used_ins): new_eqns.append(new_e...
Python
Apache License 2.0
google/jax
fix dce logic for nullary primitives
260,510
07.07.2021 11:03:59
25,200
d693324dabc1af2b323bd489078e23b360a885b8
change while loop batching fixed point condition Fixes
[ { "change_type": "MODIFY", "old_path": "jax/_src/lax/control_flow.py", "new_path": "jax/_src/lax/control_flow.py", "diff": "@@ -373,7 +373,7 @@ def _pred_bcast_select(c, pred, x, y, x_y_aval: core.AbstractValue):\nelif x_y_aval is core.abstract_token:\nreturn xops.AfterAll(c, [x, y])\nelse:\n- asser...
Python
Apache License 2.0
google/jax
change while loop batching fixed point condition Fixes #7063 Co-authored-by: Sharad Vikram <sharad.vikram@gmail.com> Co-authored-by: Adam Paszke <apaszke@google.com>
260,528
06.09.2021 17:55:22
-7,200
cdbbefa00af1af3623f525c1f78d50e976f6f9d2
Fix argument names for jnp.ndarray.clip and deprecate the old ones
[ { "change_type": "MODIFY", "old_path": "jax/_src/numpy/lax_numpy.py", "new_path": "jax/_src/numpy/lax_numpy.py", "diff": "@@ -5946,6 +5946,20 @@ def _nbytes(arr):\nreturn size(arr) * _dtype(arr).itemsize\n+def _clip(number, min=None, max=None, out=None, *, a_min=None, a_max=None):\n+ # ndarray.clip ...
Python
Apache License 2.0
google/jax
Fix argument names for jnp.ndarray.clip and deprecate the old ones
260,287
07.09.2021 03:25:54
25,200
0636f490f34d4a3c90cd6f2e3d83973643676934
Ensure that named axes consistently refer to global axis sizes in xmap Fixes
[ { "change_type": "MODIFY", "old_path": "jax/core.py", "new_path": "jax/core.py", "diff": "@@ -1654,10 +1654,12 @@ def _map_unit(size: int, axis: int, aval: AbstractUnit) -> AbstractUnit:\ndef _map_shaped_array(size: int, axis: int, aval: ShapedArray) -> ShapedArray:\nassert aval.shape[axis] == size\...
Python
Apache License 2.0
google/jax
Ensure that named axes consistently refer to global axis sizes in xmap Fixes #6959. PiperOrigin-RevId: 395210686
260,528
07.09.2021 15:37:04
-7,200
5ed619afbb47c382bc3edca2862abf7b2d92bb8a
Implement 'reflect' and 'mirror' padding modes for scipy.ndimage.map_coordinates
[ { "change_type": "MODIFY", "old_path": "jax/_src/scipy/ndimage.py", "new_path": "jax/_src/scipy/ndimage.py", "diff": "@@ -30,10 +30,22 @@ from jax._src.util import safe_zip as zip\n_nonempty_prod = functools.partial(functools.reduce, operator.mul)\n_nonempty_sum = functools.partial(functools.reduce,...
Python
Apache License 2.0
google/jax
Implement 'reflect' and 'mirror' padding modes for scipy.ndimage.map_coordinates
260,287
07.09.2021 07:53:42
25,200
0c03e9804688c4fd3b27e851b6928c4563b7842b
Don't cast out_axis_resources to a tuple automatically It's confusing and makes it impossible to specify non-trivial pytrees of out_axis_resources for functions that return lists. Also extend the error messages to be less confusing and hint at potential fixes.
[ { "change_type": "MODIFY", "old_path": "jax/api_util.py", "new_path": "jax/api_util.py", "diff": "@@ -255,12 +255,13 @@ def wrap_hashably(arg):\nelse:\nreturn Hashable(arg)\n-def flatten_axes(name, treedef, axis_tree, *, kws=False):\n+def flatten_axes(name, treedef, axis_tree, *, kws=False, tupled_a...
Python
Apache License 2.0
google/jax
Don't cast out_axis_resources to a tuple automatically It's confusing and makes it impossible to specify non-trivial pytrees of out_axis_resources for functions that return lists. Also extend the error messages to be less confusing and hint at potential fixes. PiperOrigin-RevId: 395246450
260,287
08.09.2021 01:41:38
25,200
1158530faa390bf6f4cddb8655f26b0503e938a6
Remove axis name from named_shape when unmapping avals Even though `vmap` and `pmap` don't use avals with names, the batching infrastructure is used to implement xmap and pjit. So while we keep the introduction of names carefully scoped, forgetting to remove them at the right points leads to extremely confusing errors.
[ { "change_type": "MODIFY", "old_path": "jax/_src/api.py", "new_path": "jax/_src/api.py", "diff": "@@ -2460,7 +2460,7 @@ def device_put_replicated(x: Any, devices: Sequence[xc.Device]):\nraise ValueError(\"`devices` argument to `device_put_replicated must be \"\n\"a non-empty sequence.\")\ndef _devic...
Python
Apache License 2.0
google/jax
Remove axis name from named_shape when unmapping avals Even though `vmap` and `pmap` don't use avals with names, the batching infrastructure is used to implement xmap and pjit. So while we keep the introduction of names carefully scoped, forgetting to remove them at the right points leads to extremely confusing errors. PiperOrigin-RevId: 395423006
260,287
08.09.2021 10:54:09
25,200
5b4757d234ad52f56a32caad450dc0ae2ea02bf6
New lowering APIs for pjit This is the first in a series of refactoring patches that add the new AOT APIs to all JIT-like transforms in JAX. I'm sending this early, because I expect that it will come in handy when adding reverse-mode AD support for pjit.
[ { "change_type": "MODIFY", "old_path": "jax/experimental/maps.py", "new_path": "jax/experimental/maps.py", "diff": "@@ -560,7 +560,7 @@ def xmap(fun: Callable,\nhas_input_rank_assertions = any(spec.expected_rank is not None for spec in in_axes_entries)\nhas_output_rank_assertions = any(spec.expected...
Python
Apache License 2.0
google/jax
New lowering APIs for pjit This is the first in a series of refactoring patches that add the new AOT APIs to all JIT-like transforms in JAX. I'm sending this early, because I expect that it will come in handy when adding reverse-mode AD support for pjit. PiperOrigin-RevId: 395510449
260,682
08.09.2021 23:29:20
0
add830037246983e6bd911d02e52f911a7d69a59
added images for pjit tutorial
[ { "change_type": "ADD", "old_path": "docs/_static/mesh.jpg", "new_path": "docs/_static/mesh.jpg", "diff": "Binary files /dev/null and b/docs/_static/mesh.jpg differ\n" }, { "change_type": "ADD", "old_path": "docs/_static/multi_host.jpg", "new_path": "docs/_static/multi_host.jpg", ...
Python
Apache License 2.0
google/jax
added images for pjit tutorial
260,287
09.09.2021 13:50:41
25,200
00528a42e530a7a04e82bdffd6e6f98d3558b5d3
Add new lowering APIs for jit
[ { "change_type": "MODIFY", "old_path": "jax/interpreters/xla.py", "new_path": "jax/interpreters/xla.py", "diff": "@@ -689,6 +689,9 @@ def _xla_consts(c, consts):\n@lu.cache\ndef _xla_callable(fun: lu.WrappedFun, device, backend, name, donated_invars, *arg_specs):\n+ return lower_xla_callable(fun, de...
Python
Apache License 2.0
google/jax
Add new lowering APIs for jit PiperOrigin-RevId: 395779305
260,682
09.09.2021 15:24:04
25,200
e7e8183a40e46377abd4e696887c942c2c579fdc
not exclude 08-pjit.md
[ { "change_type": "MODIFY", "old_path": "docs/conf.py", "new_path": "docs/conf.py", "diff": "@@ -115,7 +115,15 @@ exclude_patterns = [\n# Ignore markdown source for notebooks; myst-nb builds from the ipynb\n# These are kept in sync using the jupytext pre-commit hook.\n'notebooks/*.md',\n- 'jax-101/*....
Python
Apache License 2.0
google/jax
not exclude 08-pjit.md
260,287
10.09.2021 07:09:26
25,200
c845d15b3ad2fd7526c044eca62de8f0a94875ef
Cache used_axis_names calls for jaxprs All control-flow primitives are `AxisPrimitive`s now, which means that we're doing lots of those used-names traversals during dispatch and they can be expensive! This adds caching to try to lower the cost of that change.
[ { "change_type": "MODIFY", "old_path": "jax/core.py", "new_path": "jax/core.py", "diff": "@@ -38,7 +38,7 @@ from . import linear_util as lu\nfrom jax._src import source_info_util\nfrom ._src.util import (safe_zip, safe_map, partial, curry, prod, partialmethod,\n- tuple_insert, tuple_delete, as_hasha...
Python
Apache License 2.0
google/jax
Cache used_axis_names calls for jaxprs All control-flow primitives are `AxisPrimitive`s now, which means that we're doing lots of those used-names traversals during dispatch and they can be expensive! This adds caching to try to lower the cost of that change. PiperOrigin-RevId: 395921887
260,287
10.09.2021 10:02:07
25,200
cc2efbf03240e89609289daf963b2f0f4e187179
Add new lowering APIs for xmap
[ { "change_type": "MODIFY", "old_path": "jax/experimental/maps.py", "new_path": "jax/experimental/maps.py", "diff": "@@ -636,10 +636,22 @@ def xmap(fun: Callable,\nout_flat = xmap_p.bind(fun_flat, *args_flat, **params)\nreturn verify_outputs(out_flat, out_tree, params)\n- # Decorate fun_mapped\n+ def...
Python
Apache License 2.0
google/jax
Add new lowering APIs for xmap PiperOrigin-RevId: 395950859
260,411
12.09.2021 18:54:32
-10,800
b59db5b181c374291f5a72cf9e3ca518fd61b50d
[jax2tf] Fix stale comments
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/examples/saved_model_lib.py", "new_path": "jax/experimental/jax2tf/examples/saved_model_lib.py", "diff": "@@ -43,8 +43,9 @@ def convert_and_save_model(\nsaved_model_options: Optional[tf.saved_model.SaveOptions] = None):\n\"\"\"Convert a...
Python
Apache License 2.0
google/jax
[jax2tf] Fix stale comments
260,510
13.09.2021 15:39:02
25,200
ebd8d95847e90ee673d476a1b26947dd7ba340a1
Add precision param for pdot
[ { "change_type": "MODIFY", "old_path": "jax/_src/lax/lax.py", "new_path": "jax/_src/lax/lax.py", "diff": "@@ -633,7 +633,7 @@ def conv_general_dilated(\nfeature_group_count=feature_group_count,\nbatch_group_count=batch_group_count,\nlhs_shape=lhs.shape, rhs_shape=rhs.shape,\n- precision=_canonicaliz...
Python
Apache License 2.0
google/jax
Add precision param for pdot
260,335
03.09.2021 16:43:57
25,200
36154fac346be78ba0ee6298c19d04fa1e4fc0d6
handle linear custom_jvp functions
[ { "change_type": "MODIFY", "old_path": "jax/_src/custom_derivatives.py", "new_path": "jax/_src/custom_derivatives.py", "diff": "@@ -381,16 +381,24 @@ ad.reducing_transposes[custom_jvp_call_jaxpr_p] = _custom_jvp_call_jaxpr_transpo\ndef custom_jvp_jaxpr_custom_partial_eval_rule(\nsaveable: Callable[....
Python
Apache License 2.0
google/jax
handle linear custom_jvp functions
260,632
12.09.2021 09:00:11
-28,800
507cc99c8476fb841db7bab0162023d5c23bbc0b
Add a note about Google Colab setup User must run `jax.tools.colab_tpu.setup_tpu()`
[ { "change_type": "MODIFY", "old_path": "README.md", "new_path": "README.md", "diff": "@@ -466,6 +466,11 @@ the following in your cloud TPU VM:\npip install --upgrade pip\npip install \"jax[tpu]>=0.2.16\" -f https://storage.googleapis.com/jax-releases/libtpu_releases.html\n```\n+Colab TPU runtimes co...
Python
Apache License 2.0
google/jax
Add a note about Google Colab setup User must run `jax.tools.colab_tpu.setup_tpu()`
260,411
14.09.2021 14:15:40
-7,200
805e359f43f65a9368bed566b04ea6f70b16bf3c
[call_tf] Fix bug when calling TF functions with variables in a jit context. The code was assuming that the `.captured_inputs` of a function are variables in the same order as `.variables`. This seems to not be true. Lookup variables now by handle.
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/call_tf.py", "new_path": "jax/experimental/jax2tf/call_tf.py", "diff": "@@ -302,14 +302,11 @@ def _code_generator_and_avals(\n\"See https://github.com/google/jax/blob/main/jax/experimental/jax2tf/README.md#limitations-of-call-tf for a d...
Python
Apache License 2.0
google/jax
[call_tf] Fix bug when calling TF functions with variables in a jit context. The code was assuming that the `.captured_inputs` of a function are variables in the same order as `.variables`. This seems to not be true. Lookup variables now by handle.
260,411
15.09.2021 09:19:55
-7,200
d417e31d91cc7c8cc852d80c62d1713fa5a673e6
[jax2tf] Document an expected error when saving converted functions If the JAX function is not reverse-mode differentiable (e.g., lax.while_loop) then it should be converted without gradients, or saved without gradients.
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/README.md", "new_path": "jax/experimental/jax2tf/README.md", "diff": "@@ -119,6 +119,16 @@ tf.saved_model.save(my_model, '/some/directory',\noptions=tf.saved_model.SaveOptions(experimental_custom_gradients=True))\n```\n+Note that if the...
Python
Apache License 2.0
google/jax
[jax2tf] Document an expected error when saving converted functions If the JAX function is not reverse-mode differentiable (e.g., lax.while_loop) then it should be converted without gradients, or saved without gradients.
260,437
09.09.2021 18:18:13
14,400
af22aae1e02308b7bc5120b4913537f4a314c104
fix html formatting of markdown
[ { "change_type": "MODIFY", "old_path": "docs/notebooks/Common_Gotchas_in_JAX.ipynb", "new_path": "docs/notebooks/Common_Gotchas_in_JAX.ipynb", "diff": "\"\\\\hline\\n\",\n\"\\\\end{array}\\n\",\n\"$$\\n\",\n- \"<center>$\\\\ast$ = argument-__value__-independent loop condition - unrolls the loop </ce...
Python
Apache License 2.0
google/jax
fix html formatting of markdown
260,510
16.09.2021 12:28:39
25,200
8ae58be90ae53e06fac657a4ac409c1b774a6f0c
Fix for singleton axis name in `axis_index_translation_rule`
[ { "change_type": "MODIFY", "old_path": "jax/_src/lax/parallel.py", "new_path": "jax/_src/lax/parallel.py", "diff": "@@ -1297,6 +1297,12 @@ def psum_scatter(x, axis_name, *, scatter_dimension=0, axis_index_groups=None, t\ndef _axis_index_translation_rule(c, *, axis_name, axis_env, platform):\n+ if is...
Python
Apache License 2.0
google/jax
Fix for singleton axis name in `axis_index_translation_rule`
260,303
16.09.2021 21:49:57
25,200
2cee42cf6fd6b3e57baf84a3fe9fa17a1d8ad8bd
Check presence of __jax_array__ in _arraylike before calling it in isscalar
[ { "change_type": "MODIFY", "old_path": "jax/_src/numpy/lax_numpy.py", "new_path": "jax/_src/numpy/lax_numpy.py", "diff": "@@ -301,7 +301,7 @@ def _result_dtype(op, *args):\ndef _arraylike(x):\n- return isinstance(x, ndarray) or isscalar(x) or hasattr(x, '__jax_array__')\n+ return isinstance(x, ndarr...
Python
Apache License 2.0
google/jax
Check presence of __jax_array__ in _arraylike before calling it in isscalar
260,411
17.09.2021 16:42:31
-7,200
d172ba7b7e152c644c0b39231523091a45380d0e
[host_callback] Fix an assertion failure for grad(remat(host_callback)) Fixes:
[ { "change_type": "MODIFY", "old_path": "jax/experimental/host_callback.py", "new_path": "jax/experimental/host_callback.py", "diff": "@@ -1215,28 +1215,31 @@ ad.primitive_jvps[outside_call_p] = _outside_call_jvp_rule\ndef _outside_call_partial_eval_rule(trace, *args, **params):\n- # The args have be...
Python
Apache License 2.0
google/jax
[host_callback] Fix an assertion failure for grad(remat(host_callback)) Fixes: #5878
260,578
17.09.2021 08:43:25
25,200
d0acd9f3435a7eff924c9be937dc2abb1716e79a
Add flags to configure the cuda_compute_capability and rocm_amd_targets
[ { "change_type": "MODIFY", "old_path": "build/build.py", "new_path": "build/build.py", "diff": "@@ -214,7 +214,8 @@ def check_bazel_version(bazel_path):\ndef write_bazelrc(python_bin_path=None, remote_build=None,\ncuda_toolkit_path=None, cudnn_install_path=None,\ncuda_version=None, cudnn_version=Non...
Python
Apache License 2.0
google/jax
Add flags to configure the cuda_compute_capability and rocm_amd_targets
260,664
18.09.2021 14:40:29
-32,400
8b489134c818364577f630ada6aa63beefd7376a
Reflected the opinions from the reviewer.
[ { "change_type": "MODIFY", "old_path": "docs/developer.md", "new_path": "docs/developer.md", "diff": "@@ -94,7 +94,7 @@ for more details. Install the following packages:\n```\npacman -S patch coreutils\n```\n-`realpath` command should be found in shell command line.\n+Once coreutils is installed, th...
Python
Apache License 2.0
google/jax
Reflected the opinions from the reviewer.
260,287
20.09.2021 08:58:09
25,200
45c1a1b0600773f089a112c939cd4b838c8e608a
Make sure that xmap raises a clear error when using an undefined resource name
[ { "change_type": "MODIFY", "old_path": "jax/experimental/maps.py", "new_path": "jax/experimental/maps.py", "diff": "@@ -565,7 +565,7 @@ def xmap(fun: Callable,\nresource_env = thread_resources.env\navailable_resources = set(resource_env.shape.keys())\n- if necessary_resources > available_resources:\...
Python
Apache License 2.0
google/jax
Make sure that xmap raises a clear error when using an undefined resource name PiperOrigin-RevId: 397762568
260,411
22.09.2021 11:18:43
-10,800
60d0b0829166c1d3f800ea85b10bf029f6f19c05
[jax2tf] Minor updates to the jax2tf examples documentation
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/examples/README.md", "new_path": "jax/experimental/jax2tf/examples/README.md", "diff": "@@ -11,6 +11,10 @@ This directory contains a number of examples of using the\nin a larger TensorFlow Keras model.\n* use Flax models with TensorFlow...
Python
Apache License 2.0
google/jax
[jax2tf] Minor updates to the jax2tf examples documentation
260,411
23.09.2021 11:14:19
-10,800
f03baaec517da39de7fbcc3c423cf74cbcbaf9a0
Expand the docstring for jax.checkpoint. I found the old docstring to be a bit inaccurate, but mostly not explanatory enough. I think that this feature deserves a tutorial, in addition to an expanded docstring.
[ { "change_type": "MODIFY", "old_path": "jax/_src/api.py", "new_path": "jax/_src/api.py", "diff": "@@ -2708,15 +2708,33 @@ def checkpoint(fun: Callable, concrete: bool = False, prevent_cse: bool = True,\n... z = jnp.sin(y)\n... return z\n...\n- >>> jax.grad(g)(2.0)\n- DeviceArray(-0.25563914, dtype=f...
Python
Apache License 2.0
google/jax
Expand the docstring for jax.checkpoint. I found the old docstring to be a bit inaccurate, but mostly not explanatory enough. I think that this feature deserves a tutorial, in addition to an expanded docstring.
260,411
23.09.2021 17:45:57
-10,800
b1ad9e32a78eec55467f2ec0cfb3158653bddac1
[jax2tf] Ensure that shared constants in JAX are shared also in the converted function Fixes:
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/jax2tf.py", "new_path": "jax/experimental/jax2tf/jax2tf.py", "diff": "@@ -148,6 +148,13 @@ class _ThreadLocalState(threading.local):\n# Whether to actually include XLA op metadata in the generated TF ops\nself.include_xla_op_metadata = ...
Python
Apache License 2.0
google/jax
[jax2tf] Ensure that shared constants in JAX are shared also in the converted function Fixes: #7992
260,411
24.09.2021 07:21:05
-10,800
afe57bee69b24e493943bf8780b9c67bb259c39c
[jax2tf] Fix workaround for get_compiler_ir error
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/call_tf.py", "new_path": "jax/experimental/jax2tf/call_tf.py", "diff": "@@ -318,9 +318,12 @@ def _code_generator_and_avals(\nfunc_tf_hlo = function_flat_tf.experimental_get_compiler_ir(*args_tf_flat)(\nstage=\"hlo_serialized\", device_n...
Python
Apache License 2.0
google/jax
[jax2tf] Fix workaround for get_compiler_ir error
260,424
24.09.2021 12:13:24
-3,600
327e00a668de2475bdb7e9ebdccb95885a2c0820
PRNGKeys can have dtype float0 if they are a tangent.
[ { "change_type": "MODIFY", "old_path": "jax/_src/prng.py", "new_path": "jax/_src/prng.py", "diff": "@@ -24,6 +24,7 @@ from jax import core\nfrom jax import numpy as jnp\nfrom jax import tree_util\nfrom jax.config import config\n+from jax.dtypes import float0\nfrom jax.interpreters import batching\nf...
Python
Apache License 2.0
google/jax
PRNGKeys can have dtype float0 if they are a tangent.
260,287
27.09.2021 05:31:48
25,200
484c7e0915522377320ec17aa57d910a90b00302
Use shaped_abstractify to allow abstract arguments in .lower methods
[ { "change_type": "MODIFY", "old_path": "jax/experimental/maps.py", "new_path": "jax/experimental/maps.py", "diff": "@@ -30,7 +30,8 @@ from ..tree_util import (tree_flatten, tree_unflatten, all_leaves, tree_map,\ntree_leaves)\nfrom .._src.tree_util import _replace_nones\nfrom .._src.api_util import (...
Python
Apache License 2.0
google/jax
Use shaped_abstractify to allow abstract arguments in .lower methods PiperOrigin-RevId: 399166736
260,563
28.09.2021 18:42:44
-7,200
8a68f7761ba89c19ad33c8969f1f0f7deee3f519
Change absl flag documentation to mention `re.search` The absl help texts for test target discovery mention that targets will be discovered by `re.match` use. However, in the subsequent implementation, actually `re.search` is used. This commit changes the help texts for the `test_targets` and `exclude_test_targets` flags to correctly mention `re.search` as the discovery algorithm.
[ { "change_type": "MODIFY", "old_path": "jax/_src/test_util.py", "new_path": "jax/_src/test_util.py", "diff": "@@ -71,12 +71,12 @@ flags.DEFINE_bool(\nflags.DEFINE_string(\n'test_targets', '',\n- 'Regular expression specifying which tests to run, called via re.match on '\n+ 'Regular expression specif...
Python
Apache License 2.0
google/jax
Change absl flag documentation to mention `re.search` The absl help texts for test target discovery mention that targets will be discovered by `re.match` use. However, in the subsequent implementation, actually `re.search` is used. This commit changes the help texts for the `test_targets` and `exclude_test_targets` flags to correctly mention `re.search` as the discovery algorithm.
260,397
28.09.2021 20:34:35
-7,200
f9a246ac19c3e1d399b355b6cd739fd01170347e
schur lapack wrapper
[ { "change_type": "MODIFY", "old_path": "jax/_src/lax/linalg.py", "new_path": "jax/_src/lax/linalg.py", "diff": "@@ -42,6 +42,7 @@ from jax._src.lib import rocsolver\nfrom jax._src.lib import xla_client\nfrom jax._src.lib import xla_bridge as xb\n+from jax._src.lib import version as jaxlib_version\nx...
Python
Apache License 2.0
google/jax
schur lapack wrapper
260,287
29.09.2021 07:40:13
25,200
2a4648b08c382cfad69d022f98eb70c81e7cdc13
Improve the error message raised when local mesh is not a continuous subcube We need this restriction for now to make the stacking of per-device chunks into process-local SDAs to make sense. Once we have GSDA we can revisit this restriction.
[ { "change_type": "MODIFY", "old_path": "jax/interpreters/pxla.py", "new_path": "jax/interpreters/pxla.py", "diff": "@@ -1452,7 +1452,9 @@ class Mesh:\n# subcube of the full array. This is because we were biased towards taking a\n# \"hull\" spanned by the devices, and in case the local devices don't ...
Python
Apache License 2.0
google/jax
Improve the error message raised when local mesh is not a continuous subcube We need this restriction for now to make the stacking of per-device chunks into process-local SDAs to make sense. Once we have GSDA we can revisit this restriction. PiperOrigin-RevId: 399683607
260,287
30.09.2021 07:26:46
25,200
004d8558d39113a207e401074f4a67d3ff0c5af8
Raise a clear error message when pjit is used with CPU devices
[ { "change_type": "MODIFY", "old_path": "jax/experimental/pjit.py", "new_path": "jax/experimental/pjit.py", "diff": "@@ -193,6 +193,8 @@ def pjit(fun: Callable,\nif mesh.empty:\nraise RuntimeError(\"pjit requires a non-empty mesh! Are you sure that \"\n\"it's defined at the call site?\")\n+ if any(d....
Python
Apache License 2.0
google/jax
Raise a clear error message when pjit is used with CPU devices PiperOrigin-RevId: 399927116
260,287
01.10.2021 11:12:14
0
08685efb22c8b6651dc4896526c9c09382f01e42
Keep axis_env initialized during jaxpr_subcomp ``jaxpr_subcomp`` likes to lower control-flow primitives by tracing them again as JAX callables, but they're all axis primitives now and so they do require a properly initialized axis env.
[ { "change_type": "MODIFY", "old_path": "tests/api_test.py", "new_path": "tests/api_test.py", "diff": "@@ -1860,6 +1860,16 @@ class APITest(jtu.JaxTestCase):\naxis_env = [(axis_name, api.local_device_count())]\n_ = api.xla_computation(fn, axis_env=axis_env, backend='cpu')(input_x)\n+ def test_xla_com...
Python
Apache License 2.0
google/jax
Keep axis_env initialized during jaxpr_subcomp ``jaxpr_subcomp`` likes to lower control-flow primitives by tracing them again as JAX callables, but they're all axis primitives now and so they do require a properly initialized axis env.
260,287
01.10.2021 11:24:39
0
6c36c3e11c6ec6ff0ca2b6fa6552b80490a764e7
Fix xmap error message when checking axis specs The error message helpfully suggested using dictionaries, but it swapped the types for keys and values, for added confusion :)
[ { "change_type": "MODIFY", "old_path": "jax/experimental/maps.py", "new_path": "jax/experimental/maps.py", "diff": "@@ -306,7 +306,7 @@ def _parse_entry(arg_name, entry):\nraise TypeError(f\"\"\"\\\nValue mapping specification in xmap {arg_name} pytree can be either:\n- lists of axis names (possibly...
Python
Apache License 2.0
google/jax
Fix xmap error message when checking axis specs The error message helpfully suggested using dictionaries, but it swapped the types for keys and values, for added confusion :)
260,287
04.10.2021 03:24:50
25,200
490d049751fec8c832c7b2d2e3af30d9436337ba
Strengthen resource type checking to verify physical mesh consistency This lets us verify that the users are not modifying physical mesh inside pjits.
[ { "change_type": "MODIFY", "old_path": "jax/experimental/maps.py", "new_path": "jax/experimental/maps.py", "diff": "@@ -692,8 +692,9 @@ def make_xmap_callable(fun: lu.WrappedFun,\n_resource_typing_xmap([], dict(axis_resources=axis_resources,\nout_axes=out_axes,\ncall_jaxpr=jaxpr,\n+ resource_env=res...
Python
Apache License 2.0
google/jax
Strengthen resource type checking to verify physical mesh consistency This lets us verify that the users are not modifying physical mesh inside pjits. PiperOrigin-RevId: 400675775
260,392
04.10.2021 21:32:57
14,400
5d675220c0a437ea422180a65128a24d0ac27561
Add float0 support to equality and closeness check
[ { "change_type": "MODIFY", "old_path": "jax/_src/test_util.py", "new_path": "jax/_src/test_util.py", "diff": "@@ -101,6 +101,7 @@ def is_sequence(x):\nreturn True\n_default_tolerance = {\n+ _dtypes.float0: 0,\nnp.dtype(np.bool_): 0,\nnp.dtype(np.int8): 0,\nnp.dtype(np.int16): 0,\n@@ -136,6 +137,11 @...
Python
Apache License 2.0
google/jax
Add float0 support to equality and closeness check
260,335
04.10.2021 21:47:12
25,200
7ec797b794a57c211ca7ee89a7165451ff3bf738
lower rng_bit_generator using a BitcastConvertType
[ { "change_type": "MODIFY", "old_path": "jax/_src/lax/lax.py", "new_path": "jax/_src/lax/lax.py", "diff": "@@ -55,6 +55,7 @@ import jax._src.lib\nfrom jax._src.lib import pytree\nfrom jax._src.lib import xla_bridge\nfrom jax._src.lib import xla_client\n+from jax._src.lib import version as jaxlib_vers...
Python
Apache License 2.0
google/jax
lower rng_bit_generator using a BitcastConvertType
260,645
05.10.2021 09:19:23
-3,600
8bb8c8e994ba4cfb183a8abf60c05b4d0a93d3d2
Update doc for eigh. Mention `eigenvectors` before `eigenvalues` in doc to match the order of returned values.
[ { "change_type": "MODIFY", "old_path": "jax/_src/lax/linalg.py", "new_path": "jax/_src/lax/linalg.py", "diff": "@@ -89,7 +89,7 @@ def eig(x, compute_left_eigenvectors=True, compute_right_eigenvectors=True):\ndef eigh(x, lower: bool = True, symmetrize_input: bool = True):\n\"\"\"Eigendecomposition of...
Python
Apache License 2.0
google/jax
Update doc for eigh. Mention `eigenvectors` before `eigenvalues` in doc to match the order of returned values.
260,392
05.10.2021 11:30:19
14,400
ece532556bb79f3f0135b388f4b7ff691951b480
Simplify shape comparison with numpy assert
[ { "change_type": "MODIFY", "old_path": "jax/_src/test_util.py", "new_path": "jax/_src/test_util.py", "diff": "@@ -138,9 +138,7 @@ default_gradient_tolerance = {\ndef _assert_numpy_allclose(a, b, atol=None, rtol=None, err_msg=''):\nif a.dtype == b.dtype == _dtypes.float0:\n- if a.shape != b.shape:\n-...
Python
Apache License 2.0
google/jax
Simplify shape comparison with numpy assert
260,335
05.10.2021 13:46:57
25,200
8d641a1d1be49ef56e7c30b404f20d6dc48490b8
fix rng_bit_generator translation rule Fix two issues: 1. bit packing was incorrect 2. output key had different shape from input key
[ { "change_type": "MODIFY", "old_path": "jax/_src/lax/lax.py", "new_path": "jax/_src/lax/lax.py", "diff": "@@ -6745,12 +6745,10 @@ def _rng_bit_generator_shape_rule(key, *, shape, dtype, algorithm):\ndel dtype, algorithm\nreturn (key.shape, tuple(shape))\n-\ndef _rng_bit_generator_dtype_rule(key, *, ...
Python
Apache License 2.0
google/jax
fix rng_bit_generator translation rule Fix two issues: 1. bit packing was incorrect 2. output key had different shape from input key Co-authored-by: Lena Martens <lenamartens@google.com>
260,424
05.10.2021 10:57:21
-3,600
342948dcc446b882cb7ec779db1b55bb77cc015b
Add batching rule for rng_bit_generator.
[ { "change_type": "MODIFY", "old_path": "jax/_src/lax/lax.py", "new_path": "jax/_src/lax/lax.py", "diff": "@@ -6826,6 +6826,25 @@ def _convert_2xU64_to_4xU32_without_bitcast(c, key):\ndef _rng_bit_generator_named_shape_rule(key, *, shape, dtype, algorithm):\nreturn [key.named_shape, key.named_shape]\...
Python
Apache License 2.0
google/jax
Add batching rule for rng_bit_generator.
260,335
07.10.2021 19:14:13
25,200
ef710ec1f6b5f227db5f4d623df6656611b8650b
update test of dlpack error message
[ { "change_type": "MODIFY", "old_path": "tests/array_interoperability_test.py", "new_path": "tests/array_interoperability_test.py", "diff": "@@ -167,7 +167,7 @@ class DLPackTest(jtu.JaxTestCase):\nbackend = xla_bridge.get_backend()\nclient = getattr(backend, \"client\", backend)\n- regex_str = (r'Uni...
Python
Apache License 2.0
google/jax
update test of dlpack error message
260,335
07.10.2021 21:19:06
25,200
022cb8c0fc035c6e2bb3c95354f80e2b5c831371
rbg_split and rbg_fold_in: use vmap for fewer HLOs
[ { "change_type": "MODIFY", "old_path": "jax/_src/prng.py", "new_path": "jax/_src/prng.py", "diff": "@@ -505,12 +505,10 @@ def _rbg_seed(seed: int) -> jnp.ndarray:\nreturn jnp.concatenate([halfkey, halfkey])\ndef _rbg_split(key: jnp.ndarray, num: int) -> jnp.ndarray:\n- return jnp.concatenate([_three...
Python
Apache License 2.0
google/jax
rbg_split and rbg_fold_in: use vmap for fewer HLOs
260,335
07.10.2021 22:04:16
25,200
482e41d796aa3fe10b8755d9763557c3892466be
remove ShapedArray.__len__ It was confusing to overload, since we sometimes think of avals like shapes paired with dtypes, and in that case len(aval) should perhaps be like len(aval.shape). The only place where this behavior was relied on was sparse/ops.py.
[ { "change_type": "MODIFY", "old_path": "jax/core.py", "new_path": "jax/core.py", "diff": "@@ -1159,15 +1159,12 @@ class ShapedArray(UnshapedArray):\ndef strip_named_shape(self):\nreturn self.update(named_shape={})\n- def __len__(self):\n+ def _len(self, ignored_tracer):\ntry:\nreturn self.shape[0]\n...
Python
Apache License 2.0
google/jax
remove ShapedArray.__len__ It was confusing to overload, since we sometimes think of avals like shapes paired with dtypes, and in that case len(aval) should perhaps be like len(aval.shape). The only place where this behavior was relied on was sparse/ops.py.
260,411
08.10.2021 10:11:31
-7,200
39380182289c44380437ecdf8df9ab77bfbc68cd
Applied review suggestsions
[ { "change_type": "MODIFY", "old_path": "jax/_src/api.py", "new_path": "jax/_src/api.py", "diff": "@@ -2721,20 +2721,12 @@ def checkpoint(fun: Callable, concrete: bool = False, prevent_cse: bool = True,\nAt that time, the value ``jnp.sin(2.0)`` is recomputed, along with the values\n``jnp.cos(2.0)`` a...
Python
Apache License 2.0
google/jax
Applied review suggestsions
260,287
11.10.2021 10:47:43
0
dad29d343d94fedeee7e42889ec1c1b59a42e536
Disallow jax2tf translation of multi-process pjits I'm pretty sure it doesn't handle the local/global shape boundary correctly which likely leads to very confusing errors on the TF side.
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/jax2tf.py", "new_path": "jax/experimental/jax2tf/jax2tf.py", "diff": "@@ -3042,6 +3042,9 @@ def _pjit(*args: TfVal,\n_in_avals: Sequence[core.ShapedArray],\n_out_aval: core.ShapedArray) -> TfVal:\ndel donated_invars\n+ if resource_env.p...
Python
Apache License 2.0
google/jax
Disallow jax2tf translation of multi-process pjits I'm pretty sure it doesn't handle the local/global shape boundary correctly which likely leads to very confusing errors on the TF side.
260,287
11.10.2021 10:33:21
0
b8dc8ca0041c51e129d078569ff79ef08debe09c
Fix while loop batching rule for loops with constant bodies The previous implementation failed to reach a fixpoint when the body was ignoring the carry and was returning an unbatched constant. Ensuring that the result carry is at least as batched as the input carry should fix the issue.
[ { "change_type": "MODIFY", "old_path": "jax/_src/lax/control_flow.py", "new_path": "jax/_src/lax/control_flow.py", "diff": "@@ -393,7 +393,7 @@ def _while_loop_batching_rule(axis_size, axis_name, main_type, args, dims,\n# reach a fixpoint.\nfor _ in range(1 + len(carry_bat)):\n_, carry_bat_out = bat...
Python
Apache License 2.0
google/jax
Fix while loop batching rule for loops with constant bodies The previous implementation failed to reach a fixpoint when the body was ignoring the carry and was returning an unbatched constant. Ensuring that the result carry is at least as batched as the input carry should fix the issue.
260,544
11.10.2021 18:56:15
-7,200
bec943cee0234178a9143a0447b224a5faa9fbdc
fix fori_loop and scan when trivial and with disable_jit
[ { "change_type": "MODIFY", "old_path": "jax/_src/lax/control_flow.py", "new_path": "jax/_src/lax/control_flow.py", "diff": "@@ -210,6 +210,10 @@ def fori_loop(lower, upper, body_fun, init_val):\nuse_scan = False\nif use_scan:\n+ if config.jax_disable_jit and upper_ == lower_:\n+ # non-jit implementa...
Python
Apache License 2.0
google/jax
fix fori_loop and scan when trivial and with disable_jit
260,631
12.10.2021 13:57:03
25,200
9c1198167be0ba07dac5289a23115eb082889b59
Rollback: breaks when layouts of variadic tuple elements differ
[ { "change_type": "MODIFY", "old_path": "jax/_src/lax/lax.py", "new_path": "jax/_src/lax/lax.py", "diff": "@@ -5807,8 +5807,7 @@ argmin_p = standard_primitive(_argminmax_shape_rule, _argminmax_dtype_rule,\nweak_type_rule=_strip_weak_type)\nbatching.defreducer(argmin_p)\nad.defjvp_zero(argmin_p)\n-if ...
Python
Apache License 2.0
google/jax
Rollback: breaks when layouts of variadic tuple elements differ PiperOrigin-RevId: 402648033
260,335
12.10.2021 20:06:38
25,200
a310a8173c167d0faf46033657c6f0de6f6cb7f9
rewrite remat, leave old implementation for now
[ { "change_type": "ADD", "old_path": null, "new_path": "jax/_src/ad_checkpoint.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+# You may obtain a copy of the Lic...
Python
Apache License 2.0
google/jax
rewrite remat, leave old implementation for now
260,285
11.10.2021 12:00:43
21,600
63898b6ca6dc09e24fb61bfd081a6059411398c7
Allow random.choice and random.permutation on multidimensional arrays
[ { "change_type": "MODIFY", "old_path": "CHANGELOG.md", "new_path": "CHANGELOG.md", "diff": "@@ -39,6 +39,8 @@ PLEASE REMEMBER TO CHANGE THE '..main' WITH AN ACTUAL TAG in GITHUB LINK.\ndefault when using jaxlib 0.1.72 or newer. The feature can be disabled using\nthe `--experimental_cpp_pmap` flag (o...
Python
Apache License 2.0
google/jax
Allow random.choice and random.permutation on multidimensional arrays
260,335
13.10.2021 11:06:17
25,200
82d28899c7b9754609aa1e90eb89295014a3e957
add more grad-of-jit/pmap caching tests
[ { "change_type": "MODIFY", "old_path": "tests/api_test.py", "new_path": "tests/api_test.py", "diff": "@@ -2265,6 +2265,21 @@ class APITest(jtu.JaxTestCase):\nself.assertAllClose(ans1, np.cos(2.), check_dtypes=False)\nself.assertAllClose(ans2, np.cos(3.), check_dtypes=False)\n+ def test_grad_of_jit_c...
Python
Apache License 2.0
google/jax
add more grad-of-jit/pmap caching tests
260,285
13.10.2021 11:57:57
21,600
1934fd6e65d9f3781935fab70b3279ae3491956b
Cleanup random.permutation
[ { "change_type": "MODIFY", "old_path": "jax/_src/random.py", "new_path": "jax/_src/random.py", "diff": "@@ -374,8 +374,7 @@ def shuffle(key: KeyArray, x: Array, axis: int = 0) -> jnp.ndarray:\ndef permutation(key: KeyArray,\nx: Union[int, Array],\naxis: int = 0) -> jnp.ndarray:\n- \"\"\"\n- Return a...
Python
Apache License 2.0
google/jax
Cleanup random.permutation
260,335
13.10.2021 18:21:20
25,200
6741da632fa2623c1a986d9981747e5a9223565a
checkpoint_name for checkpoint policies by name Also add the jax.ad_checkpoint.print_saved_residuals utility. All this is experimental, and undocumented for now...
[ { "change_type": "MODIFY", "old_path": "jax/_src/ad_checkpoint.py", "new_path": "jax/_src/ad_checkpoint.py", "diff": "# limitations under the License.\nfrom functools import partial\n-from typing import Callable, Optional\n+from typing import Callable, Optional, List, Tuple\nimport types\nimport jax...
Python
Apache License 2.0
google/jax
checkpoint_name for checkpoint policies by name Also add the jax.ad_checkpoint.print_saved_residuals utility. All this is experimental, and undocumented for now... Co-authored-by: Qiao Zhang <zhangqiaorjc@google.com>
260,335
14.10.2021 07:09:06
25,200
725fe3abd4afb6a09b4d34ae3f3183325930c8e8
don't automatically use new checkpoint implementation There's a bug we're struggling to repro. To use the new checkpoint, just use ```python from jax.ad_checkpoint import checkpoint ``` rather than `from jax import checkpoint.
[ { "change_type": "MODIFY", "old_path": "jax/_src/api.py", "new_path": "jax/_src/api.py", "diff": "@@ -76,7 +76,7 @@ from ..interpreters import invertible_ad as iad\nfrom ..interpreters.invertible_ad import custom_ivjp\nfrom ..custom_derivatives import (closure_convert, custom_gradient, custom_jvp,\n...
Python
Apache License 2.0
google/jax
don't automatically use new checkpoint implementation There's a bug we're struggling to repro. To use the new checkpoint, just use ```python from jax.ad_checkpoint import checkpoint ``` rather than `from jax import checkpoint.
260,410
14.10.2021 09:40:41
25,200
de777c8d37fed5504e5b161a40f4fbdc3bb31026
Set dtypes of constant coefficients appropriately based on the state dtypes. This avoids unexpected casting of the output states.
[ { "change_type": "MODIFY", "old_path": "jax/experimental/ode.py", "new_path": "jax/experimental/ode.py", "diff": "@@ -58,7 +58,7 @@ def interp_fit_dopri(y0, y1, k, dt):\ndps_c_mid = jnp.array([\n6025192743 / 30085553152 / 2, 0, 51252292925 / 65400821598 / 2,\n-2691868925 / 45128329728 / 2, 187940372...
Python
Apache License 2.0
google/jax
Set dtypes of constant coefficients appropriately based on the state dtypes. This avoids unexpected casting of the output states. PiperOrigin-RevId: 403116057
260,424
14.10.2021 17:37:50
-3,600
f5d8d4bc4e26c991b3302e646f3e34d6d1c1f247
Replace loop with map in RBG batching_rule.
[ { "change_type": "MODIFY", "old_path": "jax/_src/lax/control_flow.py", "new_path": "jax/_src/lax/control_flow.py", "diff": "@@ -1979,6 +1979,19 @@ def _memcpy(axis, num, src, dst, offset):\nmasking.masking_rules[lax.concatenate_p] = _concat_masking_rule # type: ignore\n+def _rng_bit_generator_batchi...
Python
Apache License 2.0
google/jax
Replace loop with map in RBG batching_rule.
260,335
14.10.2021 11:32:09
25,200
297f79c1de3d5ada6952fe177a8939c16a040b6d
make saved_residuals utility work w/ literals
[ { "change_type": "MODIFY", "old_path": "jax/_src/ad_checkpoint.py", "new_path": "jax/_src/ad_checkpoint.py", "diff": "@@ -231,31 +231,35 @@ def saved_residuals(f, *args, **kwargs) -> List[Tuple[core.AbstractValue, str]]:\nreturn f(*args, **kwargs)\njaxpr = jax.make_jaxpr(lambda *args: jax.linearize(...
Python
Apache License 2.0
google/jax
make saved_residuals utility work w/ literals
260,335
14.10.2021 13:09:24
25,200
584aa13360887f5ea739027a880525f1e3ae9fba
document `axis_name` in the `vmap` docstring fixes
[ { "change_type": "MODIFY", "old_path": "jax/_src/api.py", "new_path": "jax/_src/api.py", "diff": "@@ -1329,6 +1329,8 @@ def vmap(fun: F, in_axes=0, out_axes=0, axis_name=None) -> F:\n(axes) of the array returned by the :func:`vmap`-ed function, which is one\nmore than the number of dimensions (axes)...
Python
Apache License 2.0
google/jax
document `axis_name` in the `vmap` docstring fixes #8220
260,335
14.10.2021 20:41:29
25,200
8bfa5eec8cbcef84e41006a03f67d71760d738ab
fix dce logic
[ { "change_type": "MODIFY", "old_path": "jax/interpreters/partial_eval.py", "new_path": "jax/interpreters/partial_eval.py", "diff": "@@ -1020,13 +1020,18 @@ def dce_jaxpr(jaxpr: Jaxpr, used_outputs: List[bool]\nmap(write, jaxpr.outvars, used_outputs)\nfor eqn in jaxpr.eqns[::-1]:\nused_outs = map(rea...
Python
Apache License 2.0
google/jax
fix dce logic
260,287
15.10.2021 14:37:38
0
49d9affce0161fb49bd99b6aa3a51ef2991ed6a5
Enable batcher and batched collective rules for tiled all gathers Fixes
[ { "change_type": "MODIFY", "old_path": "jax/_src/lax/parallel.py", "new_path": "jax/_src/lax/parallel.py", "diff": "@@ -1089,12 +1089,10 @@ def _all_gather_transpose_rule(cts, x, *, all_gather_dimension, axis_name, axis_\naxis=concat_axis),)\ndef _all_gather_batcher(vals_in, dims_in, *, all_gather_d...
Python
Apache License 2.0
google/jax
Enable batcher and batched collective rules for tiled all gathers Fixes #8221.
260,631
19.10.2021 08:09:16
25,200
e96f3632425e9b6f1076b8ff4083df73d8567682
[XLA:TPU] Adding approximate nearest neighbor search on TPU feature to JAX. The JAX primitive would call the XLA python interface for ApproxTopK on TPU, and fallbacked to sort-and-slice XLA implementation on other platforms. Auto differntiation have two possible implementations and will be submitted in seprated CLs.
[ { "change_type": "ADD", "old_path": null, "new_path": "jax/experimental/ann.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+# You may obtain a copy of the Licen...
Python
Apache License 2.0
google/jax
[XLA:TPU] Adding approximate nearest neighbor search on TPU feature to JAX. The JAX primitive would call the XLA python interface for ApproxTopK on TPU, and fallbacked to sort-and-slice XLA implementation on other platforms. Auto differntiation have two possible implementations and will be submitted in seprated CLs. PiperOrigin-RevId: 404263763
260,400
20.10.2021 03:21:55
25,200
142be1348b025837a5dbade80ae007422b80a5e5
Update jax2tf documentation with leading underscore when setting tf.Module() variables, e.g. `m._variables = ...`. Also added a test for this.
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/README.md", "new_path": "jax/experimental/jax2tf/README.md", "diff": "@@ -151,7 +151,7 @@ prediction_tf = lambda inputs: jax2tf.convert(model_jax)(params_vars, inputs)\nmy_model = tf.Module()\n# Tell the model saver what are the variabl...
Python
Apache License 2.0
google/jax
Update jax2tf documentation with leading underscore when setting tf.Module() variables, e.g. `m._variables = ...`. Also added a test for this. PiperOrigin-RevId: 404483990
260,631
20.10.2021 12:07:16
25,200
06b595321f4d62fb20d8bbf0116636aef7e7c7bb
[XLA:TPU] Support jvp/vjp in approx_top_k Copies the jvp implementation lax.sort uses. Left some comments for future optimizations
[ { "change_type": "MODIFY", "old_path": "jax/experimental/ann.py", "new_path": "jax/experimental/ann.py", "diff": "@@ -19,9 +19,9 @@ import numpy as np\nfrom jax import lax, core\nfrom jax._src.lib import xla_bridge as xb\nfrom jax._src.lib import xla_client as xc\n-from jax._src import dtypes\n+from...
Python
Apache License 2.0
google/jax
[XLA:TPU] Support jvp/vjp in approx_top_k Copies the jvp implementation lax.sort uses. Left some comments for future optimizations PiperOrigin-RevId: 404608289
260,631
20.10.2021 12:38:59
25,200
09c2c9a24b5ac34ec8be6cae60405e7368865377
[JAX] Export ann documentation.
[ { "change_type": "ADD", "old_path": null, "new_path": "docs/jax.experimental.ann.rst", "diff": "+jax.experimental.ann module\n+===========================\n+\n+.. automodule:: jax.experimental.ann\n+\n+API\n+---\n+\n+.. autofunction:: approx_max_k\n+.. autofunction:: approx_min_k\n" }, { "ch...
Python
Apache License 2.0
google/jax
[JAX] Export ann documentation. PiperOrigin-RevId: 404615254
260,317
20.10.2021 13:01:26
25,200
6cfbb890e683efab9b8a2cc05f86ba404deb8a0a
Fix broken link to experimental/README.md
[ { "change_type": "MODIFY", "old_path": "README.md", "new_path": "README.md", "diff": "@@ -98,7 +98,7 @@ For a deeper dive into JAX:\nnotebooks](https://github.com/google/jax/tree/main/docs/notebooks).\nYou can also take a look at [the mini-libraries in\n-`jax.example_libraries`](https://github.com/g...
Python
Apache License 2.0
google/jax
Fix broken link to experimental/README.md
260,362
20.10.2021 12:46:50
-3,600
67dc16fc24fbaae5a09fb1617e767433315b0ece
add fft normalisation
[ { "change_type": "MODIFY", "old_path": "jax/_src/numpy/fft.py", "new_path": "jax/_src/numpy/fft.py", "diff": "@@ -23,6 +23,17 @@ from .util import _wraps\nfrom . import lax_numpy as jnp\n+def _fft_norm(s, func_name, norm):\n+ if norm is None or norm == \"backward\":\n+ return 1\n+ elif norm == \"ort...
Python
Apache License 2.0
google/jax
add fft normalisation