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
260,447
04.08.2022 14:29:34
25,200
07da50232371837a80c27a7b69f801290d6e99f4
[sparse] Enable batch mode of COO matmat from cusparse kernels.
[ { "change_type": "MODIFY", "old_path": "jaxlib/cuda/cusparse.cc", "new_path": "jaxlib/cuda/cusparse.cc", "diff": "@@ -487,18 +487,38 @@ std::pair<size_t, py::bytes> BuildCooMatmatDescriptor(\ncusparseDnMatDescr_t mat_b = 0;\ncusparseDnMatDescr_t mat_c = 0;\n+ // All three matrices A, B, and C must h...
Python
Apache License 2.0
google/jax
[sparse] Enable batch mode of COO matmat from cusparse kernels. PiperOrigin-RevId: 465405490
260,335
05.08.2022 07:37:55
25,200
fbf6aa2a16d06caab6285b810fd09750a86cd068
small tweaks for bint ad
[ { "change_type": "MODIFY", "old_path": "jax/_src/lax/lax.py", "new_path": "jax/_src/lax/lax.py", "diff": "@@ -1461,7 +1461,7 @@ def unop(result_dtype, accepted_dtypes, name):\nprim = standard_primitive(_attrgetter('shape'), dtype_rule, name,\nweak_type_rule=weak_type_rule)\nbatching.defvectorized(pr...
Python
Apache License 2.0
google/jax
small tweaks for bint ad
260,335
05.08.2022 09:12:32
25,200
f159cdfec96b77d29108204f9aa87b9535b972a3
add trivial convert_element_type padding rule
[ { "change_type": "MODIFY", "old_path": "jax/_src/lax/lax.py", "new_path": "jax/_src/lax/lax.py", "diff": "@@ -2343,6 +2343,7 @@ ad.primitive_transposes[convert_element_type_p] = _convert_element_type_transpos\nbatching.defvectorized(convert_element_type_p)\npe.const_fold_rules[convert_element_type_p...
Python
Apache License 2.0
google/jax
add trivial convert_element_type padding rule
260,510
05.08.2022 10:58:48
25,200
89150eef1d114ff3015e238b138d4687c6523427
Enable debug callbacks in checkpoint
[ { "change_type": "MODIFY", "old_path": "jax/_src/ad_checkpoint.py", "new_path": "jax/_src/ad_checkpoint.py", "diff": "from functools import partial\nimport operator as op\n-from typing import Callable, Optional, List, Tuple, Sequence, Union, Any\n+from typing import Callable, Optional, List, Tuple, ...
Python
Apache License 2.0
google/jax
Enable debug callbacks in checkpoint PiperOrigin-RevId: 465601044
260,287
05.08.2022 18:27:11
0
bfb0814ce864f6d15c54ed70689ce5a95fbf149a
Remove experimental warning from xmap It doesn't have bugs, or at least not noticeably more than the rest of our code :)
[ { "change_type": "MODIFY", "old_path": "jax/experimental/maps.py", "new_path": "jax/experimental/maps.py", "diff": "@@ -457,7 +457,6 @@ def xmap(fun: Callable,\nspecified. This is in line with the current multi-host :py:func:`pmap`\nprogramming model.\n\"\"\"\n- warn(\"xmap is an experimental featur...
Python
Apache License 2.0
google/jax
Remove experimental warning from xmap It doesn't have bugs, or at least not noticeably more than the rest of our code :)
260,342
05.08.2022 10:15:26
25,200
736bed0344dd75b73b9ba588fe7dcfdbe0beddff
Fix jax numpy norm bug where passing ord=inf always returns one
[ { "change_type": "MODIFY", "old_path": "jax/_src/numpy/linalg.py", "new_path": "jax/_src/numpy/linalg.py", "diff": "@@ -469,6 +469,13 @@ def norm(x, ord=None, axis : Union[None, Tuple[int, ...], int] = None,\n# code has slightly different type promotion semantics, so we need a\n# special case too.\n...
Python
Apache License 2.0
google/jax
Fix jax numpy norm bug where passing ord=inf always returns one
260,510
05.08.2022 16:45:56
25,200
7c49f9a6035293436ab83c1aacc74515ed46a3be
Error in debug print if unused arguments are provided
[ { "change_type": "MODIFY", "old_path": "jax/_src/debugging.py", "new_path": "jax/_src/debugging.py", "diff": "\"\"\"Module for JAX debugging primitives and related functionality.\"\"\"\nimport enum\nimport functools\n+import string\nimport sys\nfrom typing import Callable, Any\n@@ -188,6 +189,22 @@ ...
Python
Apache License 2.0
google/jax
Error in debug print if unused arguments are provided
260,335
05.08.2022 20:55:27
25,200
e2c1df6c5175b7cf29eaadc5a020f256a160e9a7
add api_boundary, fix flax failure with new-remat-by-default
[ { "change_type": "MODIFY", "old_path": "jax/_src/ad_checkpoint.py", "new_path": "jax/_src/ad_checkpoint.py", "diff": "@@ -96,6 +96,7 @@ checkpoint_policies = types.SimpleNamespace(\n### Main API\n+@api_boundary\ndef checkpoint(fun: Callable, *, prevent_cse: bool = True,\npolicy: Optional[Callable[.....
Python
Apache License 2.0
google/jax
add api_boundary, fix flax failure with new-remat-by-default
260,335
05.08.2022 22:18:53
25,200
81b6263ed0a9e0f5d97c97a93d58be0aea1f0dd3
Rolling forward after test failures caused roll-back (from use of np.empty).
[ { "change_type": "MODIFY", "old_path": "jax/_src/api.py", "new_path": "jax/_src/api.py", "diff": "@@ -2938,7 +2938,7 @@ def _valid_jaxtype(arg):\ntry:\nxla.abstractify(arg) # faster than core.get_aval\nexcept TypeError:\n- return False\n+ return core.valid_jaxtype(arg)\nelse:\nreturn True\n" }, ...
Python
Apache License 2.0
google/jax
Rolling forward #11768 after test failures caused roll-back (from use of np.empty). PiperOrigin-RevId: 465712458
260,631
08.08.2022 08:39:20
25,200
8e2c68cbe40c3ac7db204a19e32d0479df641212
MHLO CompareOp pretty printing
[ { "change_type": "MODIFY", "old_path": "tests/filecheck/math.filecheck.py", "new_path": "tests/filecheck/math.filecheck.py", "diff": "@@ -197,30 +197,26 @@ def main(_):\nprint_ir(np.float32(1), np.float32(2))(lax.div)\n# CHECK-LABEL: TEST: eq float32[] float32[]\n- # CHECK: mhlo.compare\n- # CHECK-S...
Python
Apache License 2.0
google/jax
MHLO CompareOp pretty printing PiperOrigin-RevId: 466051458
260,510
08.08.2022 11:41:46
25,200
2d72dc899b5609ea9dd65fec4328590ac9061b5e
Enable receives in TPU callbacks and add tests
[ { "change_type": "MODIFY", "old_path": "jax/_src/debugging.py", "new_path": "jax/_src/debugging.py", "diff": "@@ -50,17 +50,15 @@ debug_callback_p.multiple_results = True\nmap, unsafe_map = util.safe_map, map\n@debug_callback_p.def_impl\n-def debug_callback_impl(*flat_args, callback: Callable[..., A...
Python
Apache License 2.0
google/jax
Enable receives in TPU callbacks and add tests PiperOrigin-RevId: 466103306
260,631
08.08.2022 16:56:14
25,200
c4192dca655aae3556450abaac722570aa6ac010
Ensure PartitionSpec is picklable. PartitionSpec instances currently don't roundtrip correctly through pickle, wrapping any value in an extra tuple, e.g. ``` > pickle.loads(pickle.dumps(PartitionSpec('model', 'batch'))) > PartitionSpec(('model', 'batch'),) ```
[ { "change_type": "MODIFY", "old_path": "jax/interpreters/pxla.py", "new_path": "jax/interpreters/pxla.py", "diff": "@@ -2332,6 +2332,9 @@ class PartitionSpec(tuple):\ndef __repr__(self):\nreturn \"PartitionSpec%s\" % tuple.__repr__(self)\n+ def __reduce__(self):\n+ return (PartitionSpec, tuple(self)...
Python
Apache License 2.0
google/jax
Ensure PartitionSpec is picklable. PartitionSpec instances currently don't roundtrip correctly through pickle, wrapping any value in an extra tuple, e.g. ``` > pickle.loads(pickle.dumps(PartitionSpec('model', 'batch'))) > PartitionSpec(('model', 'batch'),) ``` PiperOrigin-RevId: 466188998
260,510
08.08.2022 15:36:26
25,200
c34aa3933f8224134908c7e0addfb7bfd1439a67
Various debugger improvements disables globals can opt out of filtering frames can limit number of frames
[ { "change_type": "MODIFY", "old_path": "jax/_src/debugger/core.py", "new_path": "jax/_src/debugger/core.py", "diff": "@@ -80,7 +80,7 @@ class DebuggerFrame:\nreturn DebuggerFrame(\nfilename=frame_info.filename,\nlocals=frame_info.frame.f_locals,\n- globals=frame_info.frame.f_globals,\n+ globals={},\...
Python
Apache License 2.0
google/jax
Various debugger improvements - disables globals - can opt out of filtering frames - can limit number of frames
260,335
09.08.2022 12:22:10
25,200
666f4f838f011bb61958d7430732a3458809b22b
fix diffrax with new remat, do for cond what did for while_loop
[ { "change_type": "MODIFY", "old_path": "jax/_src/lax/control_flow/conditionals.py", "new_path": "jax/_src/lax/control_flow/conditionals.py", "diff": "@@ -438,9 +438,20 @@ def _cond_partial_eval(trace, *tracers, branches, linear):\n# TODO(mattjj): de-duplicate with _cond_partial_eval\ndef _cond_parti...
Python
Apache License 2.0
google/jax
fix diffrax with new remat, do for cond what #11773 did for while_loop
260,510
08.08.2022 22:18:24
25,200
18f164ff1cf8b7e04812390047c01c8b75af296a
Try flattening the locals/globals dict in the debugger and have a fallback if it fails
[ { "change_type": "MODIFY", "old_path": "jax/_src/debugger/core.py", "new_path": "jax/_src/debugger/core.py", "diff": "from __future__ import annotations\nimport dataclasses\n-import functools\nimport inspect\nimport threading\n-from typing import Any, Dict, List, Optional, Tuple\n+from typing import...
Python
Apache License 2.0
google/jax
Try flattening the locals/globals dict in the debugger and have a fallback if it fails
260,335
09.08.2022 12:42:50
25,200
580e7f39d5a3a911637ffa118bc9fbc827ff89ad
fix traceback exclusion on new checkpoint
[ { "change_type": "MODIFY", "old_path": "jax/_src/ad_checkpoint.py", "new_path": "jax/_src/ad_checkpoint.py", "diff": "@@ -27,12 +27,14 @@ from jax.interpreters import mlir\nfrom jax.tree_util import tree_flatten, tree_unflatten\nfrom jax._src import ad_util\nfrom jax._src import source_info_util\n+f...
Python
Apache License 2.0
google/jax
fix traceback exclusion on new checkpoint
260,510
09.08.2022 14:34:30
25,200
3ec1b1b9878bd6ac48f07ec66160a52346c3e14e
Check if XLA Executable has `execute_with_token` before using it
[ { "change_type": "MODIFY", "old_path": "jax/_src/dispatch.py", "new_path": "jax/_src/dispatch.py", "diff": "@@ -48,9 +48,9 @@ from jax._src import traceback_util\nfrom jax._src.abstract_arrays import array_types\nfrom jax._src.config import config, flags\nfrom jax._src.lib.mlir import ir\n+from jax....
Python
Apache License 2.0
google/jax
Check if XLA Executable has `execute_with_token` before using it PiperOrigin-RevId: 466470801
260,335
09.08.2022 19:57:40
25,200
d76754e40eaf8b6ce2ddd8e2922323f2189a0218
fix type annotation on remat
[ { "change_type": "MODIFY", "old_path": "jax/_src/api.py", "new_path": "jax/_src/api.py", "diff": "@@ -3051,7 +3051,7 @@ def eval_shape(fun: Callable, *args, **kwargs):\ndef checkpoint(fun: Callable, *,\nconcrete: bool = False,\nprevent_cse: bool = True,\n- static_argnums: Union[int, Tuple[int, ...],...
Python
Apache License 2.0
google/jax
fix type annotation on remat
260,335
10.08.2022 08:26:47
25,200
be6f6bfe9fc2eb19573bccb74ac27930838e39bb
set new jax.remat / jax.checkpoint to be on-by-default
[ { "change_type": "MODIFY", "old_path": "CHANGELOG.md", "new_path": "CHANGELOG.md", "diff": "@@ -28,6 +28,9 @@ PLEASE REMEMBER TO CHANGE THE '..main' WITH AN ACTUAL TAG in GITHUB LINK.\n{mod}`jax.example_libraries.stax`.\n* Removed `jax.experimental.optimizers`; it has long been a deprecated alias of...
Python
Apache License 2.0
google/jax
set new jax.remat / jax.checkpoint to be on-by-default
260,408
11.08.2022 15:22:02
-3,600
4ce88ec71f5de75a1c81c6bf5ec3380ee4cf3c74
Fix bug with integer typed-Gelu Certain lines in gelu() would round down constants if called with integer types (sqrt_2_over_pi = np.sqrt(2 / np.pi).astype(x.dtype)) Cast the input array to the nearest float-like type to avoid this, as done for trigonometic functions.
[ { "change_type": "MODIFY", "old_path": "jax/_src/nn/functions.py", "new_path": "jax/_src/nn/functions.py", "diff": "@@ -252,6 +252,10 @@ def gelu(x: Array, approximate: bool = True) -> Array:\nx : input array\napproximate: whether to use the approximate or exact formulation.\n\"\"\"\n+\n+ # Promote ...
Python
Apache License 2.0
google/jax
Fix bug with integer typed-Gelu Certain lines in gelu() would round down constants if called with integer types (sqrt_2_over_pi = np.sqrt(2 / np.pi).astype(x.dtype)) Cast the input array to the nearest float-like type to avoid this, as done for trigonometic functions.
260,408
11.08.2022 15:41:07
-3,600
32b2a8ff006ba7057c9e8922f48fa772a076951d
Update nn_test.py Add test for fixed integer-type Gelu behaviour.
[ { "change_type": "MODIFY", "old_path": "tests/nn_test.py", "new_path": "tests/nn_test.py", "diff": "@@ -84,6 +84,11 @@ class NNFunctionsTest(jtu.JaxTestCase):\nval = nn.glu(jnp.array([1.0, 0.0]), axis=0)\nself.assertAllClose(val, jnp.array([0.5]))\n+ def testGeluIntType(self):\n+ val_float = nn.gelu...
Python
Apache License 2.0
google/jax
Update nn_test.py Add test for fixed integer-type Gelu behaviour.
260,408
11.08.2022 15:46:34
-3,600
5026a810a4e331f38025c66e87500318839d2e0a
Update nn_test.py Add parameter to sweep over `approximate` kwarg.
[ { "change_type": "MODIFY", "old_path": "tests/nn_test.py", "new_path": "tests/nn_test.py", "diff": "@@ -84,9 +84,10 @@ class NNFunctionsTest(jtu.JaxTestCase):\nval = nn.glu(jnp.array([1.0, 0.0]), axis=0)\nself.assertAllClose(val, jnp.array([0.5]))\n- def testGeluIntType(self):\n- val_float = nn.gelu...
Python
Apache License 2.0
google/jax
Update nn_test.py Add parameter to sweep over `approximate` kwarg.
260,510
11.08.2022 15:51:41
25,200
b0309dc33a34473584e723c6d0819d70e3838e69
Bump libtpu nightly version
[ { "change_type": "MODIFY", "old_path": "setup.py", "new_path": "setup.py", "diff": "@@ -26,7 +26,7 @@ _available_cuda_versions = ['11']\n_default_cuda_version = '11'\n_available_cudnn_versions = ['82', '805']\n_default_cudnn_version = '82'\n-_libtpu_version = '0.1.dev20220722'\n+_libtpu_version = '0...
Python
Apache License 2.0
google/jax
Bump libtpu nightly version
260,510
12.08.2022 12:39:22
25,200
88f2b5e86d36fbc9e0dcfba3eab5b56975148c20
Add functionality for "pure" callbacks Also avoids using CPP dispatch path when host callbacks are involved
[ { "change_type": "MODIFY", "old_path": "jax/_src/api.py", "new_path": "jax/_src/api.py", "diff": "@@ -548,6 +548,8 @@ def _cpp_jit(\n# No effects in computation\nnot execute.args[5] and\nnot execute.args[6] and\n+ # Has no host callbacks\n+ not execute.args[8] and\n# Not supported: ShardedDeviceArra...
Python
Apache License 2.0
google/jax
Add functionality for "pure" callbacks Also avoids using CPP dispatch path when host callbacks are involved PiperOrigin-RevId: 467270949
260,631
12.08.2022 14:42:03
25,200
56440c35a2c7d7cdc61e3dbff18f26d558902207
Pass host callbacks to `backend.compile_replicated`
[ { "change_type": "MODIFY", "old_path": "jax/interpreters/pxla.py", "new_path": "jax/interpreters/pxla.py", "diff": "@@ -1337,8 +1337,8 @@ class PmapExecutable(stages.XlaExecutable):\nif hasattr(pci.backend, \"compile_replicated\"):\nexecute_fun = pci.backend.compile_replicated(\n- xla_computation, c...
Python
Apache License 2.0
google/jax
Pass host callbacks to `backend.compile_replicated` PiperOrigin-RevId: 467299712
260,510
12.08.2022 19:48:40
25,200
7cd81ca1a8118a821036c5daad073066da94515d
Allow debug prints in staged out custom derivative functions
[ { "change_type": "MODIFY", "old_path": "jax/_src/custom_derivatives.py", "new_path": "jax/_src/custom_derivatives.py", "diff": "from functools import update_wrapper, reduce, partial\nimport inspect\nimport operator as op\n-from typing import (Callable, Generic, Optional, Sequence, Tuple, List, TypeV...
Python
Apache License 2.0
google/jax
Allow debug prints in staged out custom derivative functions PiperOrigin-RevId: 467344265
260,335
11.08.2022 19:39:50
25,200
42dd7cac43fd5f971ed314c0b9d8817dfbc2273b
simplify slicing jaxprs a little
[ { "change_type": "MODIFY", "old_path": "benchmarks/api_benchmark.py", "new_path": "benchmarks/api_benchmark.py", "diff": "@@ -539,5 +539,13 @@ def bench_remat_eager_retracing_overheads_static_argnums(state):\ny.block_until_ready()\n+@google_benchmark.register\n+@google_benchmark.option.unit(google_b...
Python
Apache License 2.0
google/jax
simplify slicing jaxprs a little Co-authored-by: Sharad Vikram <sharad.vikram@gmail.com>
260,335
03.08.2022 16:02:29
25,200
5310515c80f4d786d7067d0c5cafb2c96cd502f9
Initial implementation of eager pmap
[ { "change_type": "MODIFY", "old_path": "jax/_src/api.py", "new_path": "jax/_src/api.py", "diff": "@@ -113,12 +113,12 @@ zip, unsafe_zip = safe_zip, zip\nFLAGS = flags.FLAGS\nflags.DEFINE_bool(\n- \"experimental_cpp_jit\", bool_env(\"JAX_CPP_JIT\", True),\n+ \"experimental_cpp_jit\", bool_env(\"JAX_C...
Python
Apache License 2.0
google/jax
Initial implementation of eager pmap Co-authored-by: Sharad Vikram <sharad.vikram@gmail.com> Co-authored-by: Matthew Johnson <mattjj@google.com>
260,335
10.08.2022 18:07:56
25,200
a7f760d9eda2c4f41b0a529b87854fcc689436b0
Working multihost eager pmap
[ { "change_type": "MODIFY", "old_path": "jax/core.py", "new_path": "jax/core.py", "diff": "@@ -914,10 +914,11 @@ def new_main(trace_type: Type[Trace],\nif leaked_tracers: raise leaked_tracer_error(\"trace\", t(), leaked_tracers)\n@contextmanager\n-def new_base_main(trace_type: Type[Trace]) -> Generat...
Python
Apache License 2.0
google/jax
Working multihost eager pmap Co-authored-by: Sharad Vikram <sharad.vikram@gmail.com> Co-authored-by: Skye Wanderman-Milne <skyewm@google.com>
260,510
12.08.2022 10:59:20
25,200
72dbe311722e7dfb14fd3894875ad8e57409b92c
Initial transpose implementation
[ { "change_type": "MODIFY", "old_path": "jax/_src/lax/control_flow/for_loop.py", "new_path": "jax/_src/lax/control_flow/for_loop.py", "diff": "@@ -34,6 +34,7 @@ from jax._src import state\nfrom jax._src.util import (partition_list, merge_lists, safe_map, safe_zip,\nsplit_list)\nimport jax.numpy as jn...
Python
Apache License 2.0
google/jax
Initial transpose implementation Co-authored-by: Matthew Johnson <mattjj@google.com>
260,510
11.08.2022 10:49:56
25,200
fe040cc01e76616bf347afd2b2179a4477f29819
Cleaning up eager pmap implementation
[ { "change_type": "MODIFY", "old_path": "jax/_src/api.py", "new_path": "jax/_src/api.py", "diff": "@@ -113,12 +113,12 @@ zip, unsafe_zip = safe_zip, zip\nFLAGS = flags.FLAGS\nflags.DEFINE_bool(\n- \"experimental_cpp_jit\", bool_env(\"JAX_CPP_JIT\", False),\n+ \"experimental_cpp_jit\", bool_env(\"JAX_...
Python
Apache License 2.0
google/jax
Cleaning up eager pmap implementation Co-authored-by: Matthew Johnson <mattjj@google.com>
260,335
15.08.2022 12:26:55
25,200
68e3f58041182adb6c9d0f83bacee0f39e2a68f5
un-skip polar/qdwh decomp tests skipped on gpu in On an A100 machine, these tests seem to run fine now. See
[ { "change_type": "MODIFY", "old_path": "tests/lax_scipy_test.py", "new_path": "tests/lax_scipy_test.py", "diff": "@@ -506,7 +506,6 @@ class LaxBackedScipyTests(jtu.JaxTestCase):\nfor nonzero_condition_number in nonzero_condition_numbers\nfor dtype in jtu.dtypes.inexact\nfor seed in seeds))\n- @jtu.s...
Python
Apache License 2.0
google/jax
un-skip polar/qdwh decomp tests skipped on gpu in ad6ce74 On an A100 machine, these tests seem to run fine now. See https://github.com/google/jax/issues/8628#issuecomment-1215651697.
260,287
16.08.2022 04:53:41
25,200
ffd34d5ad79331030f706046ef16ec9844875f2a
Allow collectives in manually sharded computations ... at least when the manual sharding applies to the whole mesh, because that's all that XLA can support right now. This is especially important when computing gradients of xmapped functions (when manual lowering is enabled), since AD often introduces many `psum`s.
[ { "change_type": "MODIFY", "old_path": "jax/_src/lax/parallel.py", "new_path": "jax/_src/lax/parallel.py", "diff": "@@ -714,10 +714,21 @@ def _allreduce_lowering(prim, pos_fn, ctx, *args, axes, axis_index_groups):\nreplica_groups = _replica_groups_mhlo(\n_replica_groups(ctx.module_context.axis_env, ...
Python
Apache License 2.0
google/jax
Allow collectives in manually sharded computations ... at least when the manual sharding applies to the whole mesh, because that's all that XLA can support right now. This is especially important when computing gradients of xmapped functions (when manual lowering is enabled), since AD often introduces many `psum`s. PiperOrigin-RevId: 467895089
260,287
16.08.2022 09:13:30
25,200
2aea07827cc6171e9f7e66f21c153b7ec463dedb
Fix XLA fallback to avoid checking the mesh conditions The warning about not using the full mesh manually is mainly to improve error messages (otherwise an XLA error is generated). But the MLIR lowering fallback uses axis_env unconditionally, so we have to go around that check.
[ { "change_type": "MODIFY", "old_path": "jax/interpreters/mlir.py", "new_path": "jax/interpreters/mlir.py", "diff": "@@ -364,6 +364,10 @@ class SPMDAxisContext:\n\"Collectives in manually partitioned computations are only supported \"\n\"when all mesh axes are partitioned manually (no partial automat...
Python
Apache License 2.0
google/jax
Fix XLA fallback to avoid checking the mesh conditions The warning about not using the full mesh manually is mainly to improve error messages (otherwise an XLA error is generated). But the MLIR lowering fallback uses axis_env unconditionally, so we have to go around that check. PiperOrigin-RevId: 467941551
260,557
14.08.2022 08:26:27
-7,200
99c5e91874a973b0dc7f39bba040bbcf133d82ce
add dtype arg to jnp.stack and friends Since the np.stack group is getting a dtype argument in numpy 1.24, they should also have it in JAX. Because they are just wrappers of np.concatenate, the changes are small.
[ { "change_type": "MODIFY", "old_path": "jax/_src/numpy/lax_numpy.py", "new_path": "jax/_src/numpy/lax_numpy.py", "diff": "@@ -1616,14 +1616,14 @@ def pad(array, pad_width, mode=\"constant\", **kwargs):\n@_wraps(np.stack, skip_params=['out'])\n-def stack(arrays, axis: int = 0, out=None):\n+def stack(...
Python
Apache License 2.0
google/jax
add dtype arg to jnp.stack and friends Since the np.stack group is getting a dtype argument in numpy 1.24, they should also have it in JAX. Because they are just wrappers of np.concatenate, the changes are small.
260,510
16.08.2022 10:45:17
25,200
6ae46c3d696444c840763c78b38788306227a182
Bump pmap_test size to handle new eager tests
[ { "change_type": "MODIFY", "old_path": "tests/BUILD", "new_path": "tests/BUILD", "diff": "@@ -526,9 +526,9 @@ jax_test(\nname = \"pmap_test\",\nsrcs = [\"pmap_test.py\"],\nshard_count = {\n- \"cpu\": 5,\n+ \"cpu\": 10,\n\"gpu\": 10,\n- \"tpu\": 5,\n+ \"tpu\": 10,\n},\ntags = [\"multiaccelerator\"],\...
Python
Apache License 2.0
google/jax
Bump pmap_test size to handle new eager tests PiperOrigin-RevId: 467967021
260,419
16.08.2022 14:12:20
18,000
475d4d149755907fd0aa0bbc8d3e93a0dd598a70
Increase jax.distributed timeout to 5 min
[ { "change_type": "MODIFY", "old_path": "jax/_src/distributed.py", "new_path": "jax/_src/distributed.py", "diff": "@@ -72,8 +72,10 @@ class State:\nif self.client is not None:\nraise RuntimeError('distributed.initialize should only be called once.')\n+ # Set init_timeout to 5 min to leave time for al...
Python
Apache License 2.0
google/jax
Increase jax.distributed timeout to 5 min
260,631
16.08.2022 14:25:10
25,200
0abbdd064846817e97ca751d1c7ee5f4a933789c
Add a backend field to `mlir.ModuleContext` so that host callback lowering can use the correct backend
[ { "change_type": "MODIFY", "old_path": "jax/_src/api.py", "new_path": "jax/_src/api.py", "diff": "@@ -852,7 +852,7 @@ def xla_computation(fun: Callable,\nfun_name = getattr(fun, \"__name__\", \"unknown\")\n- backend = backend if backend is not None else xb.get_backend().platform\n+ platform = backen...
Python
Apache License 2.0
google/jax
Add a backend field to `mlir.ModuleContext` so that host callback lowering can use the correct backend PiperOrigin-RevId: 468024979
260,335
16.08.2022 12:07:27
25,200
d19e34fa4a1d89c02a72dc49896013ba6fc99127
delete old remat implementation moved lowering rule logic from remat_impl.py (now deleted) to ad_checkpoint.py
[ { "change_type": "MODIFY", "old_path": "CHANGELOG.md", "new_path": "CHANGELOG.md", "diff": "@@ -10,6 +10,10 @@ PLEASE REMEMBER TO CHANGE THE '..main' WITH AN ACTUAL TAG in GITHUB LINK.\n## jax 0.3.17 (Unreleased)\n* [GitHub commits](https://github.com/google/jax/compare/jax-v0.3.16...main).\n+* Brea...
Python
Apache License 2.0
google/jax
delete old remat implementation moved lowering rule logic from remat_impl.py (now deleted) to ad_checkpoint.py
260,287
16.08.2022 11:57:48
0
6ed2d22566df9093994a3a844e2899e7e238a388
Update WORKSPACE to include the addition of use_global_device_ids in AllReduceOp.
[ { "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 = \"5b0ebc0eae9fbc746721dd75...
Python
Apache License 2.0
google/jax
Update WORKSPACE to include the addition of use_global_device_ids in AllReduceOp.
260,510
17.08.2022 10:45:47
25,200
8068e4638c924cfbe3dafb080b26c9024f9fea96
Re-bump shard count for `pmap_test`
[ { "change_type": "MODIFY", "old_path": "tests/BUILD", "new_path": "tests/BUILD", "diff": "@@ -526,9 +526,9 @@ jax_test(\nname = \"pmap_test\",\nsrcs = [\"pmap_test.py\"],\nshard_count = {\n- \"cpu\": 10,\n- \"gpu\": 10,\n- \"tpu\": 10,\n+ \"cpu\": 15,\n+ \"gpu\": 30,\n+ \"tpu\": 15,\n},\ntags = [\"m...
Python
Apache License 2.0
google/jax
Re-bump shard count for `pmap_test` PiperOrigin-RevId: 468239588
260,510
15.08.2022 17:05:27
25,200
393bca122d1731122803ea33d1105047298d48de
Expose pure callback and enable rank polymorphic callbacks
[ { "change_type": "MODIFY", "old_path": "CHANGELOG.md", "new_path": "CHANGELOG.md", "diff": "@@ -14,6 +14,8 @@ PLEASE REMEMBER TO CHANGE THE '..main' WITH AN ACTUAL TAG in GITHUB LINK.\n* {func}`jax.checkpoint`, also known as {func}`jax.remat`, no longer supports\nthe `concrete` option, following the...
Python
Apache License 2.0
google/jax
Expose pure callback and enable rank polymorphic callbacks
260,510
17.08.2022 12:29:14
25,200
87e3898a9f0e0b949d610112199d5d1949ed3512
Set `jax_eager_pmap` to True
[ { "change_type": "MODIFY", "old_path": "jax/_src/config.py", "new_path": "jax/_src/config.py", "diff": "@@ -893,7 +893,7 @@ config.define_bool_state(\n# TODO(sharadmv,mattjj): set default to True, then remove\nconfig.define_bool_state(\nname='jax_eager_pmap',\n- default=False,\n+ default=True,\nupgr...
Python
Apache License 2.0
google/jax
Set `jax_eager_pmap` to True PiperOrigin-RevId: 468265661
260,631
17.08.2022 15:04:18
25,200
ef95e2ecef11fb68529c855f7ec5b30a4935f92a
Replace JAX client side implementation of RoundNearestEven with higher efficiency op RoundNearestEvenOp
[ { "change_type": "MODIFY", "old_path": "jax/_src/lax/lax.py", "new_path": "jax/_src/lax/lax.py", "diff": "@@ -1689,19 +1689,6 @@ ceil_p = standard_unop(_float, 'ceil')\nad.defjvp_zero(ceil_p)\nmlir.register_lowering(ceil_p, partial(_nary_lower_mhlo, mhlo.CeilOp))\n-def _round_to_nearest_even(x):\n- ...
Python
Apache License 2.0
google/jax
Replace JAX client side implementation of RoundNearestEven with higher efficiency op RoundNearestEvenOp PiperOrigin-RevId: 468302607
260,335
07.01.2022 20:21:30
28,800
887b7ce2cb3d6d8aedac5cc273e137f1c876e3c7
remove custom_jvp_call_jaxpr_p and its rules They were superfluous! Instead use the "new" mechanism for converting from jaxpr params to bind params (in This change languished until we could land / and friends. But now we can!
[ { "change_type": "MODIFY", "old_path": "jax/_src/custom_derivatives.py", "new_path": "jax/_src/custom_derivatives.py", "diff": "from functools import update_wrapper, reduce, partial\nimport inspect\nimport operator as op\n-from typing import (Any, Callable, Generic, List, Optional, Sequence, Set,\n-...
Python
Apache License 2.0
google/jax
remove custom_jvp_call_jaxpr_p and its rules They were superfluous! Instead use the "new" mechanism for converting from jaxpr params to bind params (in #9136). This change languished until we could land #11830 / #11950 and friends. But now we can!
260,631
17.08.2022 22:40:14
25,200
fe665b3a641e06e5ebd7067f2204a3696235f844
Copybara import of the project: by Matthew Johnson remove custom_jvp_call_jaxpr_p and its rules They were superfluous! Instead use the "new" mechanism for converting from jaxpr params to bind params (in This change languished until we could land / and friends. But now we can!
[ { "change_type": "MODIFY", "old_path": "jax/_src/custom_derivatives.py", "new_path": "jax/_src/custom_derivatives.py", "diff": "from functools import update_wrapper, reduce, partial\nimport inspect\nimport operator as op\n-from typing import (Callable, Generic, Optional, Sequence, Tuple, TypeVar, Se...
Python
Apache License 2.0
google/jax
Copybara import of the project: -- 887b7ce2cb3d6d8aedac5cc273e137f1c876e3c7 by Matthew Johnson <mattjj@google.com>: remove custom_jvp_call_jaxpr_p and its rules They were superfluous! Instead use the "new" mechanism for converting from jaxpr params to bind params (in #9136). This change languished until we could land #11830 / #11950 and friends. But now we can! PiperOrigin-RevId: 468373797
260,507
18.08.2022 08:45:41
25,200
d5de596d1739051b88dffa2c0a310aa832c047b9
Sparse direct solver via QR factorization CUDA implementation.
[ { "change_type": "MODIFY", "old_path": "jax/experimental/sparse/linalg.py", "new_path": "jax/experimental/sparse/linalg.py", "diff": "@@ -20,6 +20,12 @@ import functools\nimport jax\nimport jax.numpy as jnp\n+from jax import core\n+from jax.interpreters import mlir\n+from jax.interpreters import xla...
Python
Apache License 2.0
google/jax
Sparse direct solver via QR factorization CUDA implementation. PiperOrigin-RevId: 468467698
260,335
18.08.2022 12:07:26
25,200
d2647cf0fa0a878e3fa862aa9a6fc923afee5004
clean up some dead code, from bad merge
[ { "change_type": "MODIFY", "old_path": "jax/_src/custom_derivatives.py", "new_path": "jax/_src/custom_derivatives.py", "diff": "@@ -331,23 +331,6 @@ def process_env_traces(primitive, level: int, jvp_was_run: bool, *args):\ntodo.append(cur_todo)\nyield outs, tuple(todo) # Ensure the aux output is imm...
Python
Apache License 2.0
google/jax
clean up some dead code, from bad merge
260,510
12.08.2022 16:52:49
25,200
49b7729f6b2334d8dbeda8f8d2859ed8471d47d0
More tests for transpose
[ { "change_type": "MODIFY", "old_path": "jax/_src/lax/control_flow/for_loop.py", "new_path": "jax/_src/lax/control_flow/for_loop.py", "diff": "@@ -307,8 +307,9 @@ def _for_partial_eval(trace: pe.JaxprTrace, *tracers: pe.JaxprTracer,\nconstvars=[])\nfor _ in range(num_inputs):\njaxpr_in_unknowns = [Fa...
Python
Apache License 2.0
google/jax
More tests for transpose
260,631
18.08.2022 22:07:58
25,200
29482a2ef6b60d319a6a4cee00e20a067fb42d2d
Copybara import of the project: by Jake VanderPlas jnp.remainder: match numpy's behavior for integer zero denominator
[ { "change_type": "MODIFY", "old_path": "jax/_src/numpy/ufuncs.py", "new_path": "jax/_src/numpy/ufuncs.py", "diff": "@@ -529,8 +529,6 @@ def frexp(x):\ndef remainder(x1, x2):\nx1, x2 = _promote_args_numeric(\"remainder\", x1, x2)\nzero = _constant_like(x1, 0)\n- if dtypes.issubdtype(x2.dtype, np.inte...
Python
Apache License 2.0
google/jax
Copybara import of the project: -- 0cf7b33e9e166f21a05bbddb04f95dad89a5f7a9 by Jake VanderPlas <jakevdp@google.com>: jnp.remainder: match numpy's behavior for integer zero denominator PiperOrigin-RevId: 468621345
260,631
19.08.2022 04:57:07
25,200
764c268ff6d3d6d69c30f6a2dae91739f0a82730
Increase the threshold to use tuple_args to 2000 args for TPUs.
[ { "change_type": "MODIFY", "old_path": "jax/_src/api.py", "new_path": "jax/_src/api.py", "diff": "@@ -916,7 +916,10 @@ def xla_computation(fun: Callable,\nxla.sharding_to_proto, in_parts_flat)),\nresult_shardings=(None if out_parts_flat is None else map(\nxla.sharding_to_proto, out_parts_flat)))\n- ...
Python
Apache License 2.0
google/jax
Increase the threshold to use tuple_args to 2000 args for TPUs. PiperOrigin-RevId: 468675921
260,335
11.08.2022 19:39:50
25,200
b3b4ffbc218a190eb4785c31cdcbf1411659c417
make slicing compilation a little faster For the common special case of int or slice-of-int/None indexing, we can generate a lax.slice rather than a lax.gather. That makes compilation a little faster, and makes the generated jaxpr a bit more wieldy too, to process in transformations and to read when pretty-printed.
[ { "change_type": "MODIFY", "old_path": "benchmarks/api_benchmark.py", "new_path": "benchmarks/api_benchmark.py", "diff": "@@ -546,6 +546,13 @@ def bench_slicing_compilation(state):\nwhile state:\njax.jit(lambda x: (x[0], x[1], x[2])).lower(x).compile()\n+@google_benchmark.register\n+@google_benchmar...
Python
Apache License 2.0
google/jax
make slicing compilation a little faster For the common special case of int or slice-of-int/None indexing, we can generate a lax.slice rather than a lax.gather. That makes compilation a little faster, and makes the generated jaxpr a bit more wieldy too, to process in transformations and to read when pretty-printed.
260,332
19.08.2022 11:08:11
25,200
c2e521807c0615244a5a1363a58d9ee73bafbce5
Add support to test gpu jaxlib nightly in CI instead of prebuilt jax/jaxlib
[ { "change_type": "MODIFY", "old_path": ".github/workflows/slurm_job_scripts/multinode_pytest.sub", "new_path": ".github/workflows/slurm_job_scripts/multinode_pytest.sub", "diff": "#!/bin/bash\n-#SBATCH -A arbitrary\n+#SBATCH -A ci-jax-gpu\n#SBATCH -p compute\n#SBATCH -N 2 # number of nodes\n-#SBATCH...
Python
Apache License 2.0
google/jax
Add support to test gpu jaxlib nightly in CI instead of prebuilt jax/jaxlib
260,447
19.08.2022 12:25:37
25,200
d37b711dd46b19baf75d990e14493c80586fcd37
[sparse] Add batch count and batch stride to matrix descriptors.
[ { "change_type": "MODIFY", "old_path": "jaxlib/cuda/cusparse.cc", "new_path": "jaxlib/cuda/cusparse.cc", "diff": "@@ -78,17 +78,21 @@ cudaDataType DtypeToCudaDataType(const py::dtype& np_type) {\n// Returns the descriptor for a Sparse matrix.\nSparseMatDescriptor BuildSparseMatDescriptor(const py::d...
Python
Apache License 2.0
google/jax
[sparse] Add batch count and batch stride to matrix descriptors. PiperOrigin-RevId: 468760351
260,558
22.08.2022 15:46:38
-7,200
88f605123302903e3b0c081b9de688959d57cc21
target_total_secs has type `int` but used as type `None` "filename": "benchmarks/benchmark.py" "warning_type": "Incompatible variable type [9]", "warning_message": " target_total_secs is declared to have type `int` but is used as type `None`." "warning_line": 86 "fix": int to Optional[int]
[ { "change_type": "MODIFY", "old_path": "benchmarks/benchmark.py", "new_path": "benchmarks/benchmark.py", "diff": "@@ -83,7 +83,7 @@ def benchmark(f: Callable[[], Any], iters: Optional[int] = None,\ndef benchmark_suite(prepare: Callable[..., Callable], params_list: List[Dict],\n- name: str, target_to...
Python
Apache License 2.0
google/jax
target_total_secs has type `int` but used as type `None` "filename": "benchmarks/benchmark.py" "warning_type": "Incompatible variable type [9]", "warning_message": " target_total_secs is declared to have type `int` but is used as type `None`." "warning_line": 86 "fix": int to Optional[int]
260,631
23.08.2022 10:37:34
25,200
a73a6a8de082c58667ab732d12762b8c5c9f1dbd
Shut down preemption sync manager if enabled.
[ { "change_type": "MODIFY", "old_path": "jax/_src/distributed.py", "new_path": "jax/_src/distributed.py", "diff": "@@ -87,6 +87,8 @@ class State:\nif self.service:\nself.service.shutdown()\nself.service = None\n+ if self.preemption_sync_manager:\n+ self.preemption_sync_manager = None\ndef initialize_...
Python
Apache License 2.0
google/jax
Shut down preemption sync manager if enabled. PiperOrigin-RevId: 469497215
260,335
23.08.2022 19:06:57
25,200
88a212e3bbdaabde113f448667095f4554201705
in ad_checkpoint WrapHashably.__eq__, check _both_ are hashable fixes
[ { "change_type": "MODIFY", "old_path": "jax/_src/ad_checkpoint.py", "new_path": "jax/_src/ad_checkpoint.py", "diff": "from functools import partial\nimport operator as op\n-from typing import Callable, Optional, List, Tuple, Sequence, Set, Union, Any\n+from typing import (Callable, Optional, List, T...
Python
Apache License 2.0
google/jax
in ad_checkpoint WrapHashably.__eq__, check _both_ are hashable fixes #12063
260,510
24.08.2022 11:46:51
25,200
91b1e01f60c9a11421eec9ca60d2533310e7328b
Update to pxla dispatch path to handle ordered effects with a single device
[ { "change_type": "MODIFY", "old_path": "jax/interpreters/mlir.py", "new_path": "jax/interpreters/mlir.py", "diff": "@@ -813,19 +813,19 @@ def lower_jaxpr_to_fun(\ntoken_types = [token_type() for _ in effects]\ninput_types = [*token_types, *input_types]\noutput_types = [*output_token_types, *token_ty...
Python
Apache License 2.0
google/jax
Update to pxla dispatch path to handle ordered effects with a single device PiperOrigin-RevId: 469785153
260,310
25.08.2022 11:15:03
25,200
28b14b1242d0e5cc31172ebdd5b8ee87ea2d49c1
Fix jax2tf GDA bug.
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/jax2tf.py", "new_path": "jax/experimental/jax2tf/jax2tf.py", "diff": "@@ -55,9 +55,11 @@ from jax._src.lax import windowed_reductions as lax_windowed_reductions\nfrom jax._src import lib as jaxlib\nfrom jax._src.lib import xla_client\n+...
Python
Apache License 2.0
google/jax
Fix jax2tf GDA bug. PiperOrigin-RevId: 470036308
260,510
25.08.2022 10:25:05
25,200
260f1d8b843483df46cf397ae5a1afc0abc9c64f
Add option to generate perfetto trace without generating link
[ { "change_type": "MODIFY", "old_path": "jax/_src/profiler.py", "new_path": "jax/_src/profiler.py", "diff": "@@ -72,12 +72,14 @@ class _ProfileState:\nself.profile_session = None\nself.log_dir = None\nself.create_perfetto_link = False\n+ self.create_perfetto_trace = False\nself.lock = threading.Lock(...
Python
Apache License 2.0
google/jax
Add option to generate perfetto trace without generating link
260,631
25.08.2022 15:14:01
25,200
cc8d406bdbf4dc782a2d9e05ea08e56360887be8
Copybara import of the project: by Jake VanderPlas generate lax.slice instead of lax.gather for more indexing cases
[ { "change_type": "MODIFY", "old_path": "jax/_src/numpy/lax_numpy.py", "new_path": "jax/_src/numpy/lax_numpy.py", "diff": "@@ -77,8 +77,8 @@ from jax._src.numpy.util import ( # noqa: F401\n_register_stackable, _stackable, _where, _wraps)\nfrom jax._src.numpy.vectorize import vectorize\nfrom jax._src....
Python
Apache License 2.0
google/jax
Copybara import of the project: -- d5ee729a7f5d583c8e212c6a9f1b98221b13cbdc by Jake VanderPlas <jakevdp@google.com>: generate lax.slice instead of lax.gather for more indexing cases PiperOrigin-RevId: 470094833
260,332
25.08.2022 15:27:07
25,200
4b1a2eaaec5a48846b3ba50c140bf721b5d180e6
combine gpu tests
[ { "change_type": "MODIFY", "old_path": ".github/workflows/slurm_job_scripts/multinode_pytest.sub", "new_path": ".github/workflows/slurm_job_scripts/multinode_pytest.sub", "diff": "@@ -33,6 +33,7 @@ read -r -d '' setup_cmd <<EOF\npython3.8 -m pip install --pre jaxlib -f https://storage.googleapis.com...
Python
Apache License 2.0
google/jax
combine gpu tests
260,403
26.08.2022 00:24:55
25,200
3c9178745ba0300f66626603bb2e4e0eab8b914a
Add runtime type check in named_scope to ensure that name is a string.
[ { "change_type": "MODIFY", "old_path": "jax/_src/api.py", "new_path": "jax/_src/api.py", "diff": "@@ -3180,6 +3180,8 @@ def named_scope(\n... logits = w.dot(x)\n... return jax.nn.relu(logits)\n\"\"\"\n+ if not isinstance(name, str):\n+ raise ValueError(\"named_scope name argument must be a string.\"...
Python
Apache License 2.0
google/jax
Add runtime type check in named_scope to ensure that name is a string. PiperOrigin-RevId: 470177071
260,689
26.08.2022 11:37:04
-7,200
024ae47e79cd7855cd02e573d5569c9d951411cd
Switch from pocketfft to ducc All credit goes to Martin Reinecke
[ { "change_type": "MODIFY", "old_path": "build/LICENSE.txt", "new_path": "build/LICENSE.txt", "diff": "@@ -4332,8 +4332,8 @@ Copyright 2019 The TensorFlow Authors. All rights reserved.\nlimitations under the License.\n--------------------------------------------------------------------------------\n-...
Python
Apache License 2.0
google/jax
Switch from pocketfft to ducc All credit goes to Martin Reinecke <martin@mpa-garching.mpg.de>.
260,332
29.08.2022 09:00:19
25,200
a571db18db953ee2af78066fea1111150032f4c1
Enable one gpu per process in multinode GPU CI
[ { "change_type": "MODIFY", "old_path": ".github/workflows/slurm_job_scripts/multinode_pytest.sub", "new_path": ".github/workflows/slurm_job_scripts/multinode_pytest.sub", "diff": "#SBATCH -J \"ci-jax-gpu\" # job name\n#SBATCH --exclusive # exclusive node access\n#SBATCH --mem=0 # all mem avail\n-#SB...
Python
Apache License 2.0
google/jax
Enable one gpu per process in multinode GPU CI
260,510
29.08.2022 14:52:14
25,200
4936713c88a678990a4a3f1f03cac0aed2c301b1
Use cases_from_list to limit number of tests
[ { "change_type": "MODIFY", "old_path": "tests/state_test.py", "new_path": "tests/state_test.py", "diff": "@@ -373,13 +373,13 @@ class StatePrimitivesTest(jtu.JaxTestCase):\nself.assertEqual(jaxpr.eqns[2].primitive, state.get_p)\nself.assertEqual(jaxpr.eqns[3].primitive, state.get_p)\n- @parameterize...
Python
Apache License 2.0
google/jax
Use cases_from_list to limit number of tests
260,510
29.08.2022 17:23:56
25,200
10b2e210ed5049595bbc00cf2a8cf9fd873382ba
Enable custom_jvp/vjp in eager pmap
[ { "change_type": "MODIFY", "old_path": "jax/interpreters/pxla.py", "new_path": "jax/interpreters/pxla.py", "diff": "@@ -1082,6 +1082,18 @@ class MapTrace(core.Trace):\nfor v, s, dst in zip(out, outaxes, out_axes_thunk()))\nreturn map(partial(MapTracer, self), out, outaxes)\n+ def process_custom_jvp_...
Python
Apache License 2.0
google/jax
Enable custom_jvp/vjp in eager pmap PiperOrigin-RevId: 470855724
260,287
18.08.2022 13:33:34
0
cbcc77b37653a4b92910a33b47215250a037d3f5
Add a correct jaxpr substitution for the cond primitive The default rule only searches for jaxprs stashed immediately in params, but cond keeps a tuple of jaxprs, so we missed them.
[ { "change_type": "MODIFY", "old_path": "jax/_src/lax/control_flow/conditionals.py", "new_path": "jax/_src/lax/control_flow/conditionals.py", "diff": "@@ -671,6 +671,12 @@ def _cond_transpose(reduce_axes, cts, *args, branches, linear):\nassert next(out_iter, None) is None\nreturn [None] + out\n+def _...
Python
Apache License 2.0
google/jax
Add a correct jaxpr substitution for the cond primitive The default rule only searches for jaxprs stashed immediately in params, but cond keeps a tuple of jaxprs, so we missed them.
260,631
30.08.2022 14:03:54
25,200
4847c0929e168423e98eaa9bbdded6838db657f0
Init local variables It's UB to use uninitialized values as arguments.
[ { "change_type": "MODIFY", "old_path": "jaxlib/lapack_kernels.cc", "new_path": "jaxlib/lapack_kernels.cc", "diff": "@@ -661,7 +661,7 @@ void RealGees<T>::Kernel(void* out_tuple, void** data, XlaCustomCallStatus*) {\nconst T* a_in = reinterpret_cast<T*>(data[4]);\n// bool* select (T, T) = reinterpret...
Python
Apache License 2.0
google/jax
Init local variables It's UB to use uninitialized values as arguments. PiperOrigin-RevId: 471084634
260,631
30.08.2022 15:29:59
25,200
9c16c832341ebbd959cef8718fa63ae14ca381c1
Rollback of upgrade logistic (sigmoid) function into a lax primitive.
[ { "change_type": "MODIFY", "old_path": "docs/jax.lax.rst", "new_path": "docs/jax.lax.rst", "diff": "@@ -103,7 +103,6 @@ Operators\nlgamma\nlog\nlog1p\n- logistic\nmax\nmin\nmul\n" }, { "change_type": "MODIFY", "old_path": "jax/_src/lax/lax.py", "new_path": "jax/_src/lax/lax.py", ...
Python
Apache License 2.0
google/jax
Rollback of upgrade logistic (sigmoid) function into a lax primitive. PiperOrigin-RevId: 471105650
260,510
31.08.2022 12:18:40
25,200
311a9cb5d9da004e9dffc781ce2bd785a7a60560
Throw error when 64-bit dtypes used incorrectly in `jax.pure_callback`
[ { "change_type": "MODIFY", "old_path": "jax/_src/callback.py", "new_path": "jax/_src/callback.py", "diff": "\"\"\"Module for JAX callbacks.\"\"\"\nfrom __future__ import annotations\n-import functools\n-\nfrom typing import Any, Callable, Sequence\nfrom jax import core\nfrom jax import tree_util\n+f...
Python
Apache License 2.0
google/jax
Throw error when 64-bit dtypes used incorrectly in `jax.pure_callback`
260,310
31.08.2022 15:05:05
25,200
59c2fc9ea9c2c289501d1af6751e2b473eddf482
[jax2tf] add a new test for jax2tf gda test. Now it cover the test using gda as jax function input.
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/tests/jax2tf_test.py", "new_path": "jax/experimental/jax2tf/tests/jax2tf_test.py", "diff": "@@ -1336,11 +1336,12 @@ class XlaCallModuleTest(tf_test_util.JaxToTfTestCase):\nglobal_shape, global_mesh, mesh_axes,\nlambda idx: global_data[i...
Python
Apache License 2.0
google/jax
[jax2tf] add a new test for jax2tf gda test. Now it cover the test using gda as jax function input. PiperOrigin-RevId: 471365834
260,411
30.08.2022 23:04:31
25,200
906a212b8050ad5d6c91d6819176c0c00b64c427
[shape_poly] Add limited support for jnp.repeat with shape polymorphism Supports only the case when the `repeats` parameter is a dimension polynomials and the total_repeat_length is None.
[ { "change_type": "MODIFY", "old_path": "jax/_src/numpy/lax_numpy.py", "new_path": "jax/_src/numpy/lax_numpy.py", "diff": "@@ -2351,7 +2351,7 @@ def indices(dimensions, dtype=int32, sparse=False):\n_TOTAL_REPEAT_LENGTH_DOC = \"\"\"\\\n-Jax adds the optional `total_repeat_length` parameter which speci...
Python
Apache License 2.0
google/jax
[shape_poly] Add limited support for jnp.repeat with shape polymorphism Supports only the case when the `repeats` parameter is a dimension polynomials and the total_repeat_length is None.
260,305
01.09.2022 13:27:23
0
1c69f594fba78062cd43cf3ca70821204619d1bf
testSphHarmOrderZeroDegreeOne and test_custom_linear_solve_cholesky have been fixed in ROCm, no need to skip
[ { "change_type": "MODIFY", "old_path": "tests/custom_linear_solve_test.py", "new_path": "tests/custom_linear_solve_test.py", "diff": "@@ -227,7 +227,6 @@ class CustomLinearSolveTest(jtu.JaxTestCase):\norder=2,\nrtol={jnp.float32: 6e-2, jnp.float64: 2e-3})\n- @jtu.skip_on_devices(\"rocm\") # rtol and...
Python
Apache License 2.0
google/jax
testSphHarmOrderZeroDegreeOne and test_custom_linear_solve_cholesky have been fixed in ROCm, no need to skip
260,510
01.09.2022 15:27:27
25,200
e1410bd16bc6e9d31900c0a588e7b9d4454beb05
Use lowering as impl rule for `pure_callback`
[ { "change_type": "MODIFY", "old_path": "jax/_src/callback.py", "new_path": "jax/_src/callback.py", "diff": "\"\"\"Module for JAX callbacks.\"\"\"\nfrom __future__ import annotations\n+import functools\n+\nfrom typing import Any, Callable, Sequence\nfrom jax import core\n@@ -21,6 +23,7 @@ from jax im...
Python
Apache License 2.0
google/jax
Use lowering as impl rule for `pure_callback`
260,496
01.09.2022 14:13:48
-7,200
7cc7da8576a679c81fc4429bd4f81a25e0201cb9
Improves quickdraw example in jax2tf
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/examples/tf_js/README.md", "new_path": "jax/experimental/jax2tf/examples/tf_js/README.md", "diff": "-This directory contains examples of using the jax2tf converter to produce\n-models that can be used with TensorFlow.js. Note that this ...
Python
Apache License 2.0
google/jax
Improves quickdraw example in jax2tf
260,411
03.09.2022 08:17:38
-10,800
fe055d06ba9517cc2a68ba2ef6270ddd06950cf0
Allow get_aval to work on ShapeDtypeStruct This is necessary to be able to call jit(f).lower(ShapeDtypeStruct(...) when jax_dynamic_shapes is on. The code in partial_eval.infer_lambda_input_type calls get_aval.
[ { "change_type": "MODIFY", "old_path": "jax/_src/api.py", "new_path": "jax/_src/api.py", "diff": "@@ -2958,6 +2958,10 @@ class ShapeDtypeStruct:\nnamed = frozenset(self.named_shape.items())\nreturn hash((self.shape, self.dtype, named))\n+core.pytype_aval_mappings[ShapeDtypeStruct] = (\n+ lambda x: S...
Python
Apache License 2.0
google/jax
Allow get_aval to work on ShapeDtypeStruct This is necessary to be able to call jit(f).lower(ShapeDtypeStruct(...) when --jax_dynamic_shapes is on. The code in partial_eval.infer_lambda_input_type calls get_aval.
260,411
03.09.2022 08:31:07
-10,800
33d1c08a31625c56c59f4a90a52c2b1fd367c7e8
Update jax/experimental/jax2tf/jax2tf.py
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/jax2tf.py", "new_path": "jax/experimental/jax2tf/jax2tf.py", "diff": "@@ -270,7 +270,7 @@ def convert(fun_jax: Callable,\nReturns:\nA version of `fun_jax` that expects TfVals as arguments (or\n- tuple/lists/dicts) thereof, and returns T...
Python
Apache License 2.0
google/jax
Update jax/experimental/jax2tf/jax2tf.py Co-authored-by: Marc van Zee <marcvanzee@gmail.com>
260,546
05.09.2022 16:16:18
25,200
892dea5d8777a8d1cf008409b2dd90fcb0df788a
Fix NaNs in the gradient of jnp.interp when the spline being interpolated into contains the same knot coordinate twice.
[ { "change_type": "MODIFY", "old_path": "jax/_src/numpy/lax_numpy.py", "new_path": "jax/_src/numpy/lax_numpy.py", "diff": "@@ -988,7 +988,9 @@ def interp(x, xp, fp, left=None, right=None, period=None):\ndf = fp[i] - fp[i - 1]\ndx = xp[i] - xp[i - 1]\ndelta = x - xp[i - 1]\n- f = where((dx == 0), fp[i...
Python
Apache License 2.0
google/jax
Fix NaNs in the gradient of jnp.interp when the spline being interpolated into contains the same knot coordinate twice. PiperOrigin-RevId: 472333594
260,424
06.09.2022 14:59:59
-3,600
82f74d18984c8b8a137b763f4bbcbeef8134b5d5
Checkify: Remove some `err` <-not-> `pred` flipping. Now all internal APIs deal in terms of errs (error thrown if True), and only the check API takes a pred (error thrown if False).
[ { "change_type": "MODIFY", "old_path": "jax/_src/checkify.py", "new_path": "jax/_src/checkify.py", "diff": "@@ -125,11 +125,11 @@ init_error = Error(False, 0, {})\nnext_code = it.count(1).__next__ # globally unique ids, could be uuid4\n-def assert_func(error: Error, pred: Bool, msg: str,\n+def asser...
Python
Apache License 2.0
google/jax
Checkify: Remove some `err` <-not-> `pred` flipping. Now all internal APIs deal in terms of errs (error thrown if True), and only the check API takes a pred (error thrown if False).
260,510
02.09.2022 10:11:58
25,200
b2a5d2c3bbd39210a0a5ff03002449f51d115e17
Add partial_eval_custom rule for `for_loop`
[ { "change_type": "MODIFY", "old_path": "jax/_src/lax/control_flow/for_loop.py", "new_path": "jax/_src/lax/control_flow/for_loop.py", "diff": "@@ -33,7 +33,7 @@ from jax._src import dtypes\nfrom jax._src import source_info_util\nfrom jax._src import state\nfrom jax._src.util import (partition_list, m...
Python
Apache License 2.0
google/jax
Add partial_eval_custom rule for `for_loop`
260,546
06.09.2022 11:21:57
25,200
dc4591dd6cd126b2a942fc89322ef65a8a0d867f
Fix NaNs in the gradient of jnp.interp when the spline being interpolated into contains knots that are small and nearby.
[ { "change_type": "MODIFY", "old_path": "jax/_src/numpy/lax_numpy.py", "new_path": "jax/_src/numpy/lax_numpy.py", "diff": "@@ -989,8 +989,9 @@ def interp(x, xp, fp, left=None, right=None, period=None):\ndx = xp[i] - xp[i - 1]\ndelta = x - xp[i - 1]\n- dx0 = dx == 0 # Protect against NaNs in `f` when ...
Python
Apache License 2.0
google/jax
Fix NaNs in the gradient of jnp.interp when the spline being interpolated into contains knots that are small and nearby. PiperOrigin-RevId: 472511203
260,332
06.09.2022 11:45:59
25,200
5f1858f533281bf5193601d94145bb4f5fece5ba
Add pytest marker inside the test only if pytest is present in the env
[ { "change_type": "MODIFY", "old_path": ".github/workflows/slurm_job_scripts/multinode_pytest.sub", "new_path": ".github/workflows/slurm_job_scripts/multinode_pytest.sub", "diff": "@@ -40,7 +40,7 @@ EOF\nread -r -d '' cmd <<EOF\ndate \\\n&& python3.8 -m pip list | grep jax \\\n-&& python3.8 -m pytest...
Python
Apache License 2.0
google/jax
Add pytest marker inside the test only if pytest is present in the env
260,335
06.09.2022 15:05:41
25,200
3c811b1520dbbea0f4eb929536444a21a8ae0a53
fix bugs, infeed/outfeed must be considered effectful
[ { "change_type": "MODIFY", "old_path": "jax/_src/ad_checkpoint.py", "new_path": "jax/_src/ad_checkpoint.py", "diff": "@@ -439,6 +439,8 @@ def remat_jvp(primals, tangents, jaxpr, prevent_cse, differentiated, policy):\nad.primitive_jvps[remat_p] = remat_jvp\nremat_allowed_effects: Set[core.Effect] = s...
Python
Apache License 2.0
google/jax
fix bugs, infeed/outfeed must be considered effectful Co-authored-by: Yash Katariya <yashkatariya@google.com> Co-authored-by: Sharad Vikram <sharad.vikram@gmail.com>
260,411
03.09.2022 08:08:50
-10,800
f1fc7fe302148dad29c602031892970cab252714
[jax2tf] Refactor the experimental_native_lowering path in jax2tf This is part of a suite of refactorings aimed towards supporting pjit by jax2tf experimental_native_lowering. The goal here is to remove many references to internal JAX core APIs, and instead use the AOT APIs: jax.jit(func_jax).lower(*args). Only the experimental_native_lowering behavior should be affected.
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/jax2tf.py", "new_path": "jax/experimental/jax2tf/jax2tf.py", "diff": "@@ -17,7 +17,7 @@ import contextlib\nimport os\nimport re\nimport threading\n-from typing import Any, Callable, Dict, Iterable, List, Optional, Sequence, Tuple, Union...
Python
Apache License 2.0
google/jax
[jax2tf] Refactor the experimental_native_lowering path in jax2tf This is part of a suite of refactorings aimed towards supporting pjit by jax2tf experimental_native_lowering. The goal here is to remove many references to internal JAX core APIs, and instead use the AOT APIs: jax.jit(func_jax).lower(*args). Only the experimental_native_lowering behavior should be affected.
260,424
07.09.2022 06:34:52
25,200
79a74f37fad6def73788f3febec8480dda4f085f
Rollback of Breaks internal target with `NotImplementedError: Differentiation rule for 'assert' not implemented`
[ { "change_type": "MODIFY", "old_path": "jax/_src/checkify.py", "new_path": "jax/_src/checkify.py", "diff": "@@ -125,11 +125,11 @@ init_error = Error(False, 0, {})\nnext_code = it.count(1).__next__ # globally unique ids, could be uuid4\n-def assert_func(error: Error, err: Bool, msg: str,\n+def assert...
Python
Apache License 2.0
google/jax
Rollback of #12232: Breaks internal target with `NotImplementedError: Differentiation rule for 'assert' not implemented` PiperOrigin-RevId: 472709946
260,631
08.09.2022 03:59:30
25,200
14f1a345a147da244db46f2165709736eed27ad0
roll back breakage
[ { "change_type": "MODIFY", "old_path": "jax/_src/ad_checkpoint.py", "new_path": "jax/_src/ad_checkpoint.py", "diff": "@@ -439,8 +439,6 @@ def remat_jvp(primals, tangents, jaxpr, prevent_cse, differentiated, policy):\nad.primitive_jvps[remat_p] = remat_jvp\nremat_allowed_effects: Set[core.Effect] = s...
Python
Apache License 2.0
google/jax
roll back breakage PiperOrigin-RevId: 472949225
260,424
08.09.2022 05:25:19
25,200
b3393e3b60e76c30dfd9820ab5ef2aafb06c9607
Checkify: add jvp rule for assert_p.
[ { "change_type": "MODIFY", "old_path": "jax/_src/checkify.py", "new_path": "jax/_src/checkify.py", "diff": "@@ -27,6 +27,7 @@ from jax import linear_util as lu\nfrom jax.api_util import flatten_fun\nfrom jax.experimental import pjit\nfrom jax.experimental import maps\n+from jax.interpreters import a...
Python
Apache License 2.0
google/jax
Checkify: add jvp rule for assert_p. PiperOrigin-RevId: 472961963
260,510
06.09.2022 13:26:41
25,200
b6c3b9df190b536396a55eaf4fef1d577025136b
Split State effect into Read/Write/Accum effects and tie them to Ref avals
[ { "change_type": "MODIFY", "old_path": "jax/_src/lax/control_flow/for_loop.py", "new_path": "jax/_src/lax/control_flow/for_loop.py", "diff": "@@ -51,7 +51,9 @@ T = TypeVar('T')\nclass Ref(Generic[T]): pass\nArray = Any\n-StateEffect = state.StateEffect\n+ReadEffect = state.ReadEffect\n+WriteEffect =...
Python
Apache License 2.0
google/jax
Split State effect into Read/Write/Accum effects and tie them to Ref avals
260,510
06.09.2022 14:29:40
25,200
6967c7ef51c4831d06e9ce8b66cedbfcf61ba390
Add sound loop invariance detection
[ { "change_type": "MODIFY", "old_path": "jax/_src/lax/control_flow/for_loop.py", "new_path": "jax/_src/lax/control_flow/for_loop.py", "diff": "from functools import partial\nimport operator\n-from typing import Any, Callable, Generic, List, Optional, Sequence, Tuple, TypeVar\n+from typing import Any,...
Python
Apache License 2.0
google/jax
Add sound loop invariance detection
260,335
08.09.2022 13:45:06
25,200
47b2dfe92fe2590ed00958b2abc0e9c6ccb7d2cb
add _device attribute to PRNGKeyArray so that computation follows key placement unrelated: remove some redundant hasattr + try / except AttributeError
[ { "change_type": "MODIFY", "old_path": "jax/_src/numpy/ndarray.py", "new_path": "jax/_src/numpy/ndarray.py", "diff": "@@ -31,8 +31,7 @@ class ArrayMeta(abc.ABCMeta):\n# that isinstance(x, ndarray) might return true but\n# issubclass(type(x), ndarray) might return false for an array tracer.\ntry:\n- ...
Python
Apache License 2.0
google/jax
add _device attribute to PRNGKeyArray so that computation follows key placement unrelated: remove some redundant hasattr + try / except AttributeError
260,447
12.09.2022 10:08:50
25,200
3243e23aa528db390fdece3fa32517c28a50318b
[sparse] Lower batch-mode bcoo_dot_genernal to cusparseSpMM.
[ { "change_type": "MODIFY", "old_path": "jax/experimental/sparse/bcoo.py", "new_path": "jax/experimental/sparse/bcoo.py", "diff": "@@ -42,6 +42,8 @@ from jax._src.lax.lax import (\n_const, ranges_like, remaining, _dot_general_batch_dim_nums, _dot_general_shape_rule,\nDotDimensionNumbers)\nfrom jax._s...
Python
Apache License 2.0
google/jax
[sparse] Lower batch-mode bcoo_dot_genernal to cusparseSpMM. PiperOrigin-RevId: 473777597
260,510
12.09.2022 14:40:11
25,200
e5725f1df1831b745efece57f10a017fafaf6951
Split for_loop_test out of lax_control_flow_test
[ { "change_type": "MODIFY", "old_path": "tests/BUILD", "new_path": "tests/BUILD", "diff": "@@ -910,6 +910,16 @@ jax_test(\n],\n)\n+jax_test(\n+ name = \"for_loop_test\",\n+ srcs = [\"for_loop_test.py\"],\n+ shard_count = {\n+ \"cpu\": 10,\n+ \"gpu\": 10,\n+ \"tpu\": 10,\n+ },\n+)\n+\njax_test(\nname ...
Python
Apache License 2.0
google/jax
Split for_loop_test out of lax_control_flow_test PiperOrigin-RevId: 473848277
260,510
13.09.2022 09:26:32
25,200
dc4922fd0827ba0f27c7c04ab3ff50b8fa8c4fab
Bump shards on for_loop_test
[ { "change_type": "MODIFY", "old_path": "tests/BUILD", "new_path": "tests/BUILD", "diff": "@@ -914,9 +914,9 @@ jax_test(\nname = \"for_loop_test\",\nsrcs = [\"for_loop_test.py\"],\nshard_count = {\n- \"cpu\": 10,\n- \"gpu\": 10,\n- \"tpu\": 10,\n+ \"cpu\": 20,\n+ \"gpu\": 20,\n+ \"tpu\": 20,\n},\n)\n...
Python
Apache License 2.0
google/jax
Bump shards on for_loop_test PiperOrigin-RevId: 474038276
260,510
13.09.2022 12:33:40
25,200
ad326b99daa5ddcab314818095da380f5afaca02
Use cases_from_list to subsample enumerated cases in for_loop_test
[ { "change_type": "MODIFY", "old_path": "tests/BUILD", "new_path": "tests/BUILD", "diff": "@@ -914,9 +914,9 @@ jax_test(\nname = \"for_loop_test\",\nsrcs = [\"for_loop_test.py\"],\nshard_count = {\n- \"cpu\": 20,\n- \"gpu\": 20,\n- \"tpu\": 20,\n+ \"cpu\": 10,\n+ \"gpu\": 10,\n+ \"tpu\": 10,\n},\n)\n...
Python
Apache License 2.0
google/jax
Use cases_from_list to subsample enumerated cases in for_loop_test PiperOrigin-RevId: 474093596
260,510
02.09.2022 10:11:58
25,200
f26f1e8afcdebc4ded7932fa5928c10a216344ac
Add support for closing over `Ref`s in nested for loops
[ { "change_type": "MODIFY", "old_path": "jax/_src/lax/control_flow/for_loop.py", "new_path": "jax/_src/lax/control_flow/for_loop.py", "diff": "from functools import partial\nimport operator\n-from typing import Any, Callable, Generic, List, Optional, Sequence, Set, Tuple, TypeVar\n+from typing import...
Python
Apache License 2.0
google/jax
Add support for closing over `Ref`s in nested for loops
260,335
13.09.2022 16:06:33
25,200
49a6034fa38fd213860e92b468551fd51dd6e6e0
[dynamic-shapes] enable basic einsum support, following jax2tf shape polys
[ { "change_type": "MODIFY", "old_path": "jax/_src/numpy/lax_numpy.py", "new_path": "jax/_src/numpy/lax_numpy.py", "diff": "@@ -2908,10 +2908,11 @@ def einsum(*operands, out=None, optimize='optimal', precision=None,\nfor d in np.shape(op) if not core.is_constant_dim(d)\n}\nif not non_constant_dim_type...
Python
Apache License 2.0
google/jax
[dynamic-shapes] enable basic einsum support, following jax2tf shape polys
260,335
13.09.2022 17:51:16
25,200
71b0968f703a6fa873b3abdfe6f9b2f01f0507b0
skip some for_loop test cases on gpu due to flakey timeouts
[ { "change_type": "MODIFY", "old_path": "tests/for_loop_test.py", "new_path": "tests/for_loop_test.py", "diff": "@@ -221,6 +221,7 @@ class ForLoopTransformationTest(jtu.JaxTestCase):\n(\"sin_sq\", for_body_sin_sq, sin_sq_ref, [(4,), (4,)], 4),\n(\"reverse\", for_body_reverse, reverse_ref, [(4,), (4,)...
Python
Apache License 2.0
google/jax
skip some for_loop test cases on gpu due to flakey timeouts PiperOrigin-RevId: 474168747
260,510
14.09.2022 08:51:37
25,200
08c5753ee2a37ba2ea46a09564c99386c2dc639d
Implement unrolling for for_loop
[ { "change_type": "MODIFY", "old_path": "jax/_src/lax/control_flow/for_loop.py", "new_path": "jax/_src/lax/control_flow/for_loop.py", "diff": "@@ -109,7 +109,7 @@ def val_to_ref_aval(x) -> ShapedArrayRef:\ndef for_loop(nsteps: Union[int, Sequence[int]],\nbody: Callable[[Array, Ref[S]], None], init_st...
Python
Apache License 2.0
google/jax
Implement unrolling for for_loop
260,674
15.09.2022 11:45:33
25,200
311f85e8b0ab2b7aedf5f8686423594e4c8a5ee8
Adjust tolerance for lu test in preparation for XLA switch from cublas to cublasLt
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jax2tf/tests/jax2tf_limitations.py", "new_path": "jax/experimental/jax2tf/tests/jax2tf_limitations.py", "diff": "@@ -827,7 +827,12 @@ class Jax2TfLimitation(primitive_harness.Limitation):\ncustom_numeric(\ndtypes=[np.float32, np.complex64], de...
Python
Apache License 2.0
google/jax
Adjust tolerance for lu test in preparation for XLA switch from cublas to cublasLt PiperOrigin-RevId: 474618770