hash
stringlengths
40
40
msg
stringlengths
1
131k
author
stringlengths
1
33
email
stringlengths
0
57
date
int64
1,447B
1,698B
04ca99cc59bd5754d43718fe7ea018df0a5b9073
Simplify Device API PiperOrigin-RevId: 559550786
Haibo Huang
hhb@google.com
1,692,827,574,000
c69f7eb64f1b284e1efaede8d2d16d7c7aa68ddd
Fix memory leak in Rendezvous API PiperOrigin-RevId: 559552415
Haibo Huang
hhb@google.com
1,692,827,875,000
9dbbe006419d50174b3711c44e1df578a7714dff
Optimize BroadcastAdd6DSlow in TFLite reference kernel. It removes redundant dimensions, compresses compressible dimensions, and handles broadcasting. PiperOrigin-RevId: 559554841
Jared Junyoung Lim
jaredlim@google.com
1,692,828,380,000
84fa457347d92305c9a7bacb1da6a8ed0944741f
Unit test for Rendezvous API PiperOrigin-RevId: 559560973
Haibo Huang
hhb@google.com
1,692,829,725,000
2e927ced12de17eb1a6564735f35ee76fecfba94
[tflite-gpu] Add helper functions so that GPU Async can be zero-copy. PiperOrigin-RevId: 559561240
Grant Jensen
grantjensen@google.com
1,692,829,777,000
e1fa66c2e5f84179dd6ec5d1a2075f573ec42a72
lite: add option to disable serialization of stablehlo ops PiperOrigin-RevId: 559573264
Zichuan Wei
zichuanwei@google.com
1,692,832,428,000
ec0540a5a70ef22d3db915d198db22e296879f39
output incorrectly taken from inputs PiperOrigin-RevId: 559579114
A. Unique TensorFlower
gardener@tensorflow.org
1,692,833,859,000
1706f7613c88a4ff21956d5796306501b6ffcc12
Rendezvous API fortify: nullify deleted variables PiperOrigin-RevId: 559584664
Haibo Huang
hhb@google.com
1,692,835,219,000
26b5341b0d8e5f5382854bef7eecf2e4d8321249
Add support to discover multiple GPUs. PiperOrigin-RevId: 559585088
Changhui Lin
changhuilin@google.com
1,692,835,316,000
c3af98bca9e153d6a68473b8be94ec58302b6424
Fix startup config options startup options cannot be affected by --config=... flags, as they are initialized before the configs are evaluated. PiperOrigin-RevId: 559586164
Austin Anderson
angerson@google.com
1,692,835,575,000
606a03ba1226b18946640536f2b251cc7920d299
Add additional dialects to GetHloModules The func dialect was found missing, I added some others that we have commonly in ingress too. PiperOrigin-RevId: 559587193
Jacques Pienaar
jpienaar@google.com
1,692,835,816,000
21f553c44158bfb044b05278e09208f9c1417253
Add `experimental_write_callbacks` to `CheckpointOptions`. PiperOrigin-RevId: 559594012
Yu Feng
feyu@google.com
1,692,837,696,000
e46b690df18637c58771906059010939df958937
lite: flatbuffer_import: convert stablehlo attribute to tensor type instead of vector type PiperOrigin-RevId: 559606029
Zichuan Wei
zichuanwei@google.com
1,692,841,163,000
5d18993f81c9990a35dce7c34ad233ad14c29591
Release the `TensorHandle` lock when waiting for `OpIdAndOutputNum`. PiperOrigin-RevId: 559621835
Russell Power
power@google.com
1,692,846,984,000
f2f703aa361d41c65364443995cd964d8ba52a3a
compat: Update forward compatibility horizon to 2023-08-24 PiperOrigin-RevId: 559685208
A. Unique TensorFlower
gardener@tensorflow.org
1,692,867,852,000
36d8091284cc54c50ecd056ab6683b4a1c41b76b
Update GraphDef version to 1598. PiperOrigin-RevId: 559685225
A. Unique TensorFlower
gardener@tensorflow.org
1,692,867,857,000
f8916e252e546a773ddae6bd2206b8a9516206f4
[XLA:GPU] Triton GEMM: fix handling of degenerate fragments of split tensor dimensions. PiperOrigin-RevId: 559704209
Ilia Sergachev
sergachev@google.com
1,692,873,844,000
289bf39526d33f8fdf9cc0970cbcffd837268fb1
Add missing dependent dialect to rank specialization. PiperOrigin-RevId: 559711590
Stephan Herhut
herhut@google.com
1,692,876,150,000
928efe9b36c480af9fd28c9d4af38fb455fdcd68
Call TfLiteInitializeShimsForTest() in SetUpTestSuite(), rather than in SetUp(). This is needed because some tests invoke TFLite API functions/methods in the test fixture's constructor, before SetUp() has been called. PiperOrigin-RevId: 559718230
Fergus Henderson
fergus@google.com
1,692,878,513,000
46e403cb7fd27d377e687a07437346caff683368
Explain how new instructions are handled during dfs traversal. PiperOrigin-RevId: 559719036
Adrian Kuegel
akuegel@google.com
1,692,878,741,000
25b4c9fd886f5f905d0460f65e555b64f5c69883
Rename tf_to_kernel to hlo_to_kernel and remove last dependencies on tensorflow dialects. PiperOrigin-RevId: 559720803
Stephan Herhut
herhut@google.com
1,692,879,360,000
9958a34b74839484cbc2bf79f28c89aaa8dc7dbe
Add type annotations. PiperOrigin-RevId: 559723269
Shashank Viswanadha
shashankvi@google.com
1,692,880,177,000
75cb7394f3a6e5c52065344719179e6d4a067d46
[XLA:GPU] Check for hlo_module_ not being nullptr PiperOrigin-RevId: 559730555
Tamás Danyluk
tdanyluk@google.com
1,692,882,400,000
9048a15d439b859f1656710f8cf866e4537af5d8
Do not choose bitcasts as fusion roots. This just adds indexing overhead without any benefit, and may require extra buffers. Bitcasts outside of fusions are no-ops. We still allow to fuse a bitcast producer into an already existing fusion. Otherwise they would act as fusion blockers. PiperOrigin-RevId: 559732964
Adrian Kuegel
akuegel@google.com
1,692,883,146,000
7fb51777aa28f70f6d9f3bf5f6e919778efbac2f
Allow cuBLAS LT FP8 gemms to run on Ada GPUs. cuBLAS supports running FP8 gemms on Ada GPUs (compute capability 8.9) but XLA previously only ran them on Hopper GPUs (compute capability 9.0). cuDNN still only supports FP8 on Hopper, so we cannot run FP8 convolutions on Ada yet. PiperOrigin-RevId: 559733091
Reed Wanderman-Milne
reedwm@google.com
1,692,883,186,000
8bb9a24e5a1f8d41dfe15480fdd5d43761a1ffd6
[XLA:GPU] Enable single wave autotuning by default This change should be relatively safe, but if you encounter any problem, please check with --xla_gpu_single_wave_autotuning=false. This change speeds up the autotuning compilation (not the overall compilation time) ~1-2x, but may cause some increase in memory usage. ...
Tamás Danyluk
tdanyluk@google.com
1,692,883,252,000
ce2c2951a923b72bb52d78f398260885d664ff7d
[XLA:GPU] Avoid leaking mlir::Operations in TritonWrapper PiperOrigin-RevId: 559740141
Tamás Danyluk
tdanyluk@google.com
1,692,885,336,000
405dc5d4bd058d18d73fed7a2dac5f8ebd1477ae
[NFC] Explicitly set dialects to `usePropertiesForAttributes=0` in preparation for https://reviews.llvm.org/D158581 (flipping the default to `1`) to land. This allows us to switch dialects to use properties one by one. PiperOrigin-RevId: 559751065
Christian Sigg
csigg@google.com
1,692,888,413,000
37a598ad832d6b438eaeb05f9f9cadceefb9335c
Update TFRT dependency to use revision http://github.com/tensorflow/runtime/commit/9ca593e8dc03e4d11582fa20a2ff0dc5b5c70e4d. PiperOrigin-RevId: 559761231
A. Unique TensorFlower
gardener@tensorflow.org
1,692,890,799,000
cce02236eae10c87086e5a54a2006fd271260fe5
Add variable quantization option test for float16 in lite conversion PiperOrigin-RevId: 559761410
Doyoung Gwak
doyounggwak@google.com
1,692,890,846,000
6b01491a8726929d3f9060709f5f153168feeb52
[XLA:GPU] Load binary if ptx is missing in ResolveConstantGlobals In a future CL we will leave the ptx empty if the cubin is available, and this CL is needed to make that work. Also update related comments. PiperOrigin-RevId: 559767896
Tamás Danyluk
tdanyluk@google.com
1,692,892,385,000
11fc64b1ef9b37692ff9ae58799f7a7fe4ff4365
[XLA:GPU][NFC] Refactor Triton fusion analysis. Prepare the analysis class to process not only GEMM fusions: rename the analysis class and several functions, move several functions into the fusion context, update the dimension order construction interface, update comments, change some pointers to references. PiperOri...
Ilia Sergachev
sergachev@google.com
1,692,893,539,000
aa5673adf05f1508cadd961f0cb28964bf5934d2
[XLA] Relax the restriction on passing preferred element types PiperOrigin-RevId: 559777378
Amit Sabne
asabne@google.com
1,692,894,435,000
c71cae7b140bfea31f106f9fd2f583af047c8d8a
Ensure that outfeed ops have their shardings preserved as they go through auto-sharding. PiperOrigin-RevId: 559782721
A. Unique TensorFlower
gardener@tensorflow.org
1,692,895,561,000
3f1ca93b581fa4c1bc1f9365349c1a484eb4a79c
[XLA] Delete `_xla_host_transfer_original_type` and `_xla_host_transfer_is_lower_bits`. PiperOrigin-RevId: 559786239
Ce Zheng
zce@google.com
1,692,896,316,000
9a16aeac0dc7503681345633d0fa60c84704ce3d
Add streamz metric to count gpu compiled programs. PiperOrigin-RevId: 559790245
A. Unique TensorFlower
gardener@tensorflow.org
1,692,897,017,000
f26dc498d89c503985433953fed5bba2ff55619e
Use similarly signed types for comparison In the downstream TFLM project, comparing int and size_t types results in sign-compare compiler warnings. PiperOrigin-RevId: 559793295
RJ Ascani
rjascani@google.com
1,692,897,584,000
8c2b83f5593a618a48402af6a04629980a127fb5
Remove TPU specific init in NextPluggableDevice. The only remaining framework init is in tensorflow/core/tpu/tpu_api_dlsym_initializer.cc which handles both SE and TFRT cases. This change completes the TPU init simplification related to using PJRT plugin. Also fixes third_party/tensorflow/core/tpu/tpu_model_server_in...
Jieying Luo
jieying@google.com
1,692,897,876,000
174d3230094e35e80eebdb03c23e6ae0f5730d6f
[NFC] Add comment to indicate default value of XLA_PYTHON_CLIENT_MEM_FRACTION PiperOrigin-RevId: 559795181
Rahul Joshi
jurahul@google.com
1,692,897,927,000
68d398dc16ed901b142ccd7f57b34e0324c5e6a0
Support degenerate case of mul operation. PiperOrigin-RevId: 559796643
A. Unique TensorFlower
gardener@tensorflow.org
1,692,898,216,000
0b8d45fdd073e29511a9d6ba20bd6bbe9aa2bf99
PR #5152: [ROCM] Update failing xla/service/gpu tests Imported from GitHub PR https://github.com/openxla/xla/pull/5152 Copybara import of the project: -- c8e616402d84ca56904fa2338e10d13ae65b7324 by Dragan Mladjenovic <Dragan.Mladjenovic@amd.com>: [ROCM] Update failing xla/service/gpu tests Merging this change clos...
Dragan Mladjenovic
Dragan.Mladjenovic@amd.com
1,692,898,246,000
62479f9b3e6c54f4dbcf99e4e026450606d7d63a
Integrate LLVM at llvm/llvm-project@382b97554dd3 Updates LLVM usage to match [382b97554dd3](https://github.com/llvm/llvm-project/commit/382b97554dd3) PiperOrigin-RevId: 559804563
Krasimir Georgiev
krasimir@google.com
1,692,899,625,000
4b34b37da2b7d9852f6f5c096c4ad1516f87c2b5
[XLA/Conditional code motion] Remove invalid get-tuple-element operations when moving operands inside conditional branches. PiperOrigin-RevId: 559810879
A. Unique TensorFlower
gardener@tensorflow.org
1,692,900,756,000
d1a5c9af980cdfb4da1abb5abf52590103f5b80a
Always return error status if serialization to FlatBuffer failed. An error is not emitted to the DiagnosticHandler in all cases in which serialization can fail. This change bases failure on the result of the `MlirToFlatBufferTranslateFunction` method call rather than the value of the status handler's status. PiperOri...
Arian Arfaian
aarfaian@google.com
1,692,900,932,000
22f749bf0bd07f804b98eaf029ccbbbbbb22b9c9
Update TFRT dependency to use revision http://github.com/tensorflow/runtime/commit/812b5fd6c28cd43131344174faee6adee20b1619. PiperOrigin-RevId: 559816076
A. Unique TensorFlower
gardener@tensorflow.org
1,692,901,635,000
934d5040c94cdccd165503e02f62682a89a4a10c
Add micro LIBTPU_EXCLUDE_C_API_IMPL to remove API implementations from libtpu PiperOrigin-RevId: 559818210
Haibo Huang
hhb@google.com
1,692,902,017,000
28649cd109f6cec8a0e02293d8b9eeb261c3335d
Merge pull request #61332 from Intel-tensorflow:kanvi/New_Layernorm_pattern PiperOrigin-RevId: 559818560
TensorFlower Gardener
gardener@tensorflow.org
1,692,902,673,000
b726a7ce9954459c811b2a97d79423010b58101c
Update FindTpuIdx for the DCNMessage. PiperOrigin-RevId: 559820271
Clive Verghese
cliveverghese@google.com
1,692,902,398,000
c0a59d7d11f967e1adac45a71add4f281c3423f4
[xla:gpu] Fix GetLatencyBetween. We didn't return the intended latency for Send/SendDone and Recv/RecvDone, due to a problem in using IsAsyncPair and a problem in checking opcode for the instruction pair. PiperOrigin-RevId: 559825578
Bixia Zheng
bixia@google.com
1,692,903,445,000
616c71eeddf09c706ba1e289009770b16b8aa81d
Make remaining targets under tensorflow/python/ have strict dependencies. PiperOrigin-RevId: 559852588
Juan Martinez Castellanos
juanantoniomc@google.com
1,692,909,057,000
9ab898c000cffa664829b31f1e1c78a57b467887
Add headers BUILD targets for tensorflow C API constructs Context: Will remove `tf_tensor` impl from libtpu, which requires some cleanup of headers and separation of interface and implementation PiperOrigin-RevId: 559853472
David Silverstone
silverstone@google.com
1,692,909,251,000
a94b7e777a2c2621aa80e7e15ae9e45c6bf2b5cd
put rocm config back in bazelrc
weihanmines
wei.han3@amd.com
1,691,781,285,000
39afd5255c8d4d0a89d46a8af20e98ffb571f09c
[XLA:Python] Remove --enable_tpu flag from the Python client. This is no longer needed now TPU support is shipped as a PJRT plugin. PiperOrigin-RevId: 559855269
Peter Hawkins
phawkins@google.com
1,692,909,602,000
c1244dd1cf9ace7f8b473f0e78504cf4580089a7
DTensor: Clean up C++ headers, includes and build_targets Context: Will remove `tf_tensor` impl from libtpu, which requires some cleanup of headers and separation of interface and implementation PiperOrigin-RevId: 559860084
David Silverstone
silverstone@google.com
1,692,910,556,000
6182de75f9b2b765c986fcf44e2bd8f4d31b7433
Add the op stats pass in the TFLite converter for printing stats of ops that are not converted to TFL ops. PiperOrigin-RevId: 559861123
Yishuang Pang
ypang@google.com
1,692,910,768,000
331276a552d40b9c15acaf4d7203d9eca20ae63d
Remove never called StreamExecutor::PlatformDeviceCount, ::SupportsDnn, ::SupportsBlas, and ::SupportsFft methods. PiperOrigin-RevId: 559869460
A. Unique TensorFlower
gardener@tensorflow.org
1,692,912,306,000
e5b257fee18494348d43b3b4318473ba29f9d9ac
Remove unused MultiPlatformManager methods. PiperOrigin-RevId: 559885166
A. Unique TensorFlower
gardener@tensorflow.org
1,692,915,447,000
f69bcdec0ca15c2ee85c92c5fcf5d848576f34c1
Eliminate unused Platform::ForceExecutorShutdown, ::GetPeerAccessMap, and EnablePeerAccess. PiperOrigin-RevId: 559885414
A. Unique TensorFlower
gardener@tensorflow.org
1,692,915,502,000
935ce2ea1c3c848d213f04c9ff68ea4c51e02f89
PR #5103: Force the cublasCreate call before we start the graph. Imported from GitHub PR https://github.com/openxla/xla/pull/5103 Calling cublasCreate() need locks and can sync the context. If we have kernels that are already running and are waiting for something, like a NCCL kernel, we end up with deadlock. [Section...
Frédéric Bastien
frederic.bastien@gmail.com
1,692,915,939,000
487a0015dc4e9a3b4e2c4b66ed290bcdf207cc01
Add support for transferring a tensor with PjRtTensorBuffer. PiperOrigin-RevId: 559887956
Changhui Lin
changhuilin@google.com
1,692,916,043,000
de806069cfc73867d1e5cb6a9ad5e42fd696a66f
Integrate StableHLO at openxla/stablehlo@d9a17eb PiperOrigin-RevId: 559892197
Kevin Gleason
gleasonk@google.com
1,692,916,973,000
ca6374646ea9e98c59994dc24dd88b57680c3059
[TF][MLIR] Fix `RegionBranchOpInterface` method in `tf.WhileRegion` op This commit fixes the following two minor bugs in the `getSuccessorRegions()` interface method of the `tf.WhileRegion` op:- 1. The `cond` region successor of both `body` and the op is incorrectly marked to have the arguments of `body` (instead of ...
Srishti Srivastava
srisrivastava@google.com
1,692,917,640,000
79c43c3d54adaa759240e80f2758183c53547def
Update variable name to adhere to the style guide. PiperOrigin-RevId: 559897422
Juhyun Lee
impjdi@google.com
1,692,918,107,000
5322fd40cd58cfa8c551e602fede7a3be19fff95
[PJRT] Fix checking for output sharding Output sharding for empty tuple needs to have one "replicated" element. PiperOrigin-RevId: 559899447
Marcello Maggioni
maggioni@google.com
1,692,918,516,000
4a5a381c95918a3f2484a154016aaf518a58c720
lite:stablehlo: add serialization for stablehlo ops: abs, and, cos, exp, floor, log, min, negate, or, power, remainder, rsqrt, select, sub & tanh PiperOrigin-RevId: 559900601
Zichuan Wei
zichuanwei@google.com
1,692,918,797,000
669c19aee668ec5c6eda630bc1c1a0bdca8b3158
[XLA:GPU] Parametrizes Triton Softmax tests from ir_emitter_triton. Currently, F32 and F16 are tested. PiperOrigin-RevId: 559905690
A. Unique TensorFlower
gardener@tensorflow.org
1,692,919,966,000
f8763b7ea65f1757288fdbfa56553a4cc3cf70bc
Makes `memory_space` a pure virtual method in PjRtBuffer. PiperOrigin-RevId: 559906991
A. Unique TensorFlower
gardener@tensorflow.org
1,692,920,255,000
6ee904dca19a5fd584b2cf8341769e54c03b5e8f
#tf-data Add log statements for symbolic checkpointing. PiperOrigin-RevId: 559909678
Wilsin Gosti
wilsin@google.com
1,692,920,895,000
d9e5091f8e517d65ddcbec6bad00ab241bcaa153
Add `_copy_trackable_to_cpu()`, `_serialize_to_tensor()`, and `_restore_from_tensor()` to `DelegatingTrackableMixin`. PiperOrigin-RevId: 559918267
A. Unique TensorFlower
gardener@tensorflow.org
1,692,922,933,000
07eff3770b8414aa5814b5e4d9d2e7a6ac1d6b3d
Avoid clamping in Lowering UQ mhlo.add for i32 storage type in ConvertMHLOQuantToInt pass For i32 storage type, clamping against int32_max/min is no-op. PiperOrigin-RevId: 559922810
A. Unique TensorFlower
gardener@tensorflow.org
1,692,924,247,000
844e7e5f3c227f4bb32be67da5f36e919a91d11c
#tf-data-service Include snapshot path in the log message. PiperOrigin-RevId: 559929314
Yang Chen
yangchen@google.com
1,692,925,917,000
3d50778c6d32129afd13abfdcbb7e8e118feb1d4
Fix delegation of VarHandle nodes for unsupported datatype VarHandle nodes don't have a type associated with it. We rely on a ReadVariable/AssignVariable to figure out the data type. When visiting VarHandle, we currently incorrectly assume we can always handle it, but it really depends on the data type, which we only ...
Zhi An Ng
zhin@google.com
1,692,926,861,000
159f2b1d3cdb6c783e4d6d389ded63cc86baf62f
Allow i8->i32 convert op for zero point constants in `ComposeUniformQuantizedDotGeneralOpWithTwoQuantizedActivations`. Currently the `mlir::odml::ComposeUniformQuantizedDotGeneralOpWithTwoQuantizedActivations` assumes the zero point values are constant folded into i32 values. However, there are use cases where the zer...
Dan Suh
dansuh@google.com
1,692,927,062,000
1e6e95965950f076d8c1880567055cfaa09796d5
Add missing elem-wize functions to xla operation semantics doc
Alexander Pivovarov
apivovarov@gmail.com
1,692,933,627,000
c6ecfeac886e6c39193d422d577cb23b55f3530c
Integrate LLVM at llvm/llvm-project@1ff0bdb86dbf Updates LLVM usage to match [1ff0bdb86dbf](https://github.com/llvm/llvm-project/commit/1ff0bdb86dbf) PiperOrigin-RevId: 559954302
A. Unique TensorFlower
gardener@tensorflow.org
1,692,934,045,000
a7f3934a67900720af3d3b15389551483bee50b8
Better support for reversed sharding strategies (generated due to the presence of the kReverse HLO op). This implies handling some cases where we coudl not previosuly compute resharding costs. PiperOrigin-RevId: 559960750
A. Unique TensorFlower
gardener@tensorflow.org
1,692,936,384,000
7d9fb81b8a605f736ddbb99a5397d8a69fb4f4e2
Update TFRT dependency to use revision http://github.com/tensorflow/runtime/commit/32c232c66f5224228d1544158c458a9e8368a5cd. PiperOrigin-RevId: 559965494
A. Unique TensorFlower
gardener@tensorflow.org
1,692,938,003,000
4eb74b85d75cf823b8f6faf350e6c4e13a397d4d
Add a proper jax config for memories so that we can iteratively develop and enable it. PiperOrigin-RevId: 559977015
Yash Katariya
yashkatariya@google.com
1,692,940,993,000
a4e65d7b453387b124c7a34aa39de1a9a1c8af44
Internal Code Change PiperOrigin-RevId: 559996011
A. Unique TensorFlower
gardener@tensorflow.org
1,692,946,256,000
6bd368dd8c509aef655ba6c5615bfc8e0cf75abc
Internal Code Change PiperOrigin-RevId: 559999040
A. Unique TensorFlower
gardener@tensorflow.org
1,692,947,137,000
1290b847d80684c1f6a3f193d9ffdf8431659e35
Merge pull request #61690 from apivovarov:fix_xla_op_sem PiperOrigin-RevId: 560020394
TensorFlower Gardener
gardener@tensorflow.org
1,692,953,675,000
b79a11c222085ab958ab5e9ef41450d437bf0ff3
Update GraphDef version to 1599. PiperOrigin-RevId: 560023468
A. Unique TensorFlower
gardener@tensorflow.org
1,692,954,224,000
20108002d564107ec8613b057d053331f7834ae4
compat: Update forward compatibility horizon to 2023-08-25 PiperOrigin-RevId: 560023473
A. Unique TensorFlower
gardener@tensorflow.org
1,692,954,225,000
9af1d271c240e22d20196cbc3a0c993450b35d4e
Update TFLite `SingleOpModel` to test ops that use the new `BuiltinOptions2` schema union. PiperOrigin-RevId: 560030520
Quentin Khan
qkhan@google.com
1,692,956,441,000
bfe5732c0423c893a23eff90b35aa90027c41c38
[XLA:GPU] Remove dependency of ProfileExecutable on Executable::module() This will allow not storing the HLO module in the autotuning executables, thus saving memory. PiperOrigin-RevId: 560036254
Tamás Danyluk
tdanyluk@google.com
1,692,958,242,000
62296aa33e61aad357d43489ce6df0d0846b9ded
[XLA:GPU] Unset some debug options for autotuning compilations This would probably just waste memory. PiperOrigin-RevId: 560047723
Tamás Danyluk
tdanyluk@google.com
1,692,961,688,000
cd0f81459f474b69c9200f2da66aaca704f5348b
[XLA:GPU] Autotuning: Also try convolution algorithms returned by heuristic_mode_a. On A100, heuristic_mode_b currently does not return enough fast algorithms. So as a workaround, also try the algorithms returned from heuristic_mode_a. This comes at a cost though, as we will try more algorithms during autotuning. Pip...
Adrian Kuegel
akuegel@google.com
1,692,961,732,000
5eaf14034fefaef4258f55bd87be609b32e11093
PR #5074: XLA runtime support for fused attention Imported from GitHub PR https://github.com/openxla/xla/pull/5074 This PR add XLA Runtime support for cuDNN runtime based Fused attention feature in XLA. Copybara import of the project: -- 4fe6ba760c00e3f82e631ff8f7d6404dbdd5cf9e by Ayan Moitra <amoitra@nvidia.com>: ...
Ayan Moitra
amoitra@nvidia.com
1,692,964,149,000
d8031689066ca52c4a8f8f10f47c524fbd979c7a
PR #5185: Add a debug option to toggle reduction epilogue fusion Imported from GitHub PR https://github.com/openxla/xla/pull/5185 This is to follow up on the original [pr](https://github.com/openxla/xla/commit/f91402af93e86752c5e7c5d0a42ec54f89c007fe) to enable reduction epilogue fusion in xla gpu fusion modules. We ...
TJ Xu
tjx@nvidia.com
1,692,964,888,000
5c79f0cf5943ef5fdf1dd809e2538eeceaa7512c
Enable DUS tests on GPU. They are passing now, most likely since when we have the DynamicUpdateSlice in-place emitter. PiperOrigin-RevId: 560066499
Adrian Kuegel
akuegel@google.com
1,692,967,840,000
27ac120221813b19352299ead14ac5a0e5a5e3dc
Enable convolution test on GPU which is working now. PiperOrigin-RevId: 560067697
Adrian Kuegel
akuegel@google.com
1,692,968,259,000
61b272d94361b8b80896b77e5bad35ed45c3b3b4
#tf-data Reuse the `FunctionLibraryDefinition` rather than creating a new one for every identity map function check in `map` transformations to speed up `noop_elimination` rewrite. PiperOrigin-RevId: 560076996
Wilsin Gosti
wilsin@google.com
1,692,971,142,000
cc3c71696b077c85f989ba7ca4d1474901cc617a
Enable ManyParametersIntoWhileLoop test on CPU and GPU. This runs reasonably fast now, it is only the second-slowest test after ThreeThousandParametersAndOutputElements test which runs more than twice as long. PiperOrigin-RevId: 560080763
Adrian Kuegel
akuegel@google.com
1,692,972,240,000
b10b155d44bcb3bed5f10065326c8bb7f164f734
Increase number of tuple parameters in LargeTuple test. Previously, a bigger number resulted in timeouts or OOM on different backends. This seems to work now. PiperOrigin-RevId: 560080982
Adrian Kuegel
akuegel@google.com
1,692,972,301,000
35dd4818cb17c2c1a6981e52fb4842b84652a314
[XLA:GPU] Triton GEMM: support more broadcasts. This requires these additional changes: - Broadcasts now rely on the dimension analysis class. Because Softmax and GEMM emitters share the broadcast generation code the dimension analysis of Softmax fusions is added. - Softmax fusions contain reductions, so a minimal s...
Ilia Sergachev
sergachev@google.com
1,692,972,777,000
4ceb8dc4bc27bf2d0dd5517318c9002ced72bfd7
Enable ReduceWindow test on GPU. Apparently this was failing at some point due to a PTXAS bug that got fixed by now. PiperOrigin-RevId: 560082727
Adrian Kuegel
akuegel@google.com
1,692,972,828,000
65e35d558215f44e5a542ec5c586df8a307f1140
Integrate LLVM at llvm/llvm-project@2acf00bd0ac2 Updates LLVM usage to match [2acf00bd0ac2](https://github.com/llvm/llvm-project/commit/2acf00bd0ac2) PiperOrigin-RevId: 560101344
Krasimir Georgiev
krasimir@google.com
1,692,977,960,000
f3631aff4b00ed0ffa5d197548d4bb831052d993
Update TFRT dependency to use revision http://github.com/tensorflow/runtime/commit/7807f8c4fcb384394aed0dc05dce7604bc60e4b6. PiperOrigin-RevId: 560110922
A. Unique TensorFlower
gardener@tensorflow.org
1,692,980,325,000
dd3182796834b1888b2380e419c4cf8f5b2e791d
Add control dependencies for input output buffer alias. If needed, we add control dependency to enforce output-related instructions after the users of input buffers. PiperOrigin-RevId: 560122049
A. Unique TensorFlower
gardener@tensorflow.org
1,692,982,755,000
83373223fce5da628907042e16d221755684ec21
clip_by_global_norm: Support tensors in t_list with different dtypes. Casts the scale factor to each tensor's dtype before multiplying. Motivation: When using a mixture of float32 and bfloat16 variables, avoids trivial dtype mismatch errors. PiperOrigin-RevId: 560122639
RJ Skerry-Ryan
rjryan@google.com
1,692,982,863,000