Develop upstream sync 240123 #2375
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
PiperOrigin-RevId: 599021699
PiperOrigin-RevId: 599023315
…port.cc to tensorflow/compiler/mlir/lite/utils/const_tensor_utils.h for shared usage PiperOrigin-RevId: 599023637
PiperOrigin-RevId: 599035702
…art #2 PiperOrigin-RevId: 599037622
…art #3 PiperOrigin-RevId: 599039077
…) the number of iterations. For such loops, use that value instead of the static estimate we have for more accurate cost modeling. PiperOrigin-RevId: 599040661
PiperOrigin-RevId: 599048699
…erleave PiperOrigin-RevId: 599049008
PiperOrigin-RevId: 599052754
I had to duplicate error handling macros in a few files, this is a temporary hack and will be removed with direct uses of nccl and cuda APIs. PiperOrigin-RevId: 599057949
… is enabled. Currently the proto is dumped every time the dumping is enabled for the module. PiperOrigin-RevId: 599058121
PiperOrigin-RevId: 599064645
PiperOrigin-RevId: 599086843
PiperOrigin-RevId: 599086860
PiperOrigin-RevId: 599087162
…r tests Imported from GitHub PR openxla/xla#8033 Merging this change closes tensorflow#8033 PiperOrigin-RevId: 599110626
We had several callers of FindNonTrivialHero which called it with a fusion instruction. This was usually an indicator that the call was not actually needed at all and we should just use the instruction itself as the hero. Also adjust ChooseFusionKind to check the producer, too, whether it is a kInput fusion. PiperOrigin-RevId: 599125417
… Type Imported from GitHub PR openxla/xla#8402 This PR adds BF16 support in oneDNN Matmul op by allowing the Dot op to maintain the BF16 type until handled by OneDnnMatMulRewriter pass. Copybara import of the project: -- 4f7ddbcd5ecf7a4b3cfd140abd9a73d193e9ca39 by Mahmoud Abuzaina <mahmoud.abuzaina@intel.com>: Enable MatMul op in BF16 Merging this change closes tensorflow#8402 PiperOrigin-RevId: 599132673
With LLVM 17+ LLD fails on undefined symbols in the linker version script by default. This breaks XLA builds with errors like this: ``` ld.lld: error: version script assignment of 'global' to symbol 'initxla_extension' failed: symbol not defined ld.lld: error: version script assignment of 'global' to symbol 'init_xla_extension' failed: symbol not defined ``` The problem is that the linker version script lists symbols that only existed in Python-2 builds. Since Python-2 is not even supported anymore by pybind11 we can just remove those entries from the linker script. (This is the change which removed Python-2 support: pybind/pybind11@6493f49) PiperOrigin-RevId: 599133946
…CompileToHsaco Imported from GitHub PR openxla/xla#8506 This fixes sporadic crashes in multithreaded_compilation_test_gpu. Copybara import of the project: -- a6fc9ada24d551f35e5f01bafb2cadbcf848f41b by Dragan Mladjenovic <Dragan.Mladjenovic@amd.com>: [ROCm] Move device libs path initialization into CompileToHsaco This fixes sporadic crashes in multithreaded_compilation_test_gpu. Merging this change closes tensorflow#8506 PiperOrigin-RevId: 599134851
…ure. Also improve test coverage slightly. PiperOrigin-RevId: 599144190
PiperOrigin-RevId: 599144440
…leOp`. After calibration, statistics will be attached to the resulting `ModuleOp`. This component takes `ModuleOp` as its input and output. This implies that during the calibration process, which relies on python-level TF Session APIs (and TF runtime), it is exported to SavedModel before the calibration and imported back to ModuleOp after the calibration internally. Tests are omitted because it requires components to be exposed to the python layer because the component depends on `PyFunctionLibrary`, which should only be injected from the python layer. Using test doubles (mocks) for `PyFunctionLibrary` has been considered but discarded, because doing so would require implementing `SaveExportedModel` properly in c++, which is an overkill for the objective of simply testing `CalibrationComponent`. PiperOrigin-RevId: 599146078
This is in preparation of adding support for libnvptxcompiler. PiperOrigin-RevId: 599151463
PiperOrigin-RevId: 599155194
Updates LLVM usage to match [f3d534c4251b](llvm/llvm-project@f3d534c4251b) PiperOrigin-RevId: 599164620
Added generic caution note
The `no_cuda_asan` tag is not considered in our current build config, but `noasan` is. PiperOrigin-RevId: 599169543
An upstream Triton issue triggers UBSAN on those tests. Let's disable sanitizers on them until that's fixed. PiperOrigin-RevId: 599170685
PiperOrigin-RevId: 600581010
…d move to Thunk CollectiveExecuteParams is a companion of Thunk::ExecuteParams and has to be defined close to it. Also convert class to struct for consistency with ExecuteParams. PiperOrigin-RevId: 600581907
PiperOrigin-RevId: 600583266
PiperOrigin-RevId: 600584468
PiperOrigin-RevId: 600584560
…efore initialization and execution PiperOrigin-RevId: 600586067
PiperOrigin-RevId: 600592917
… thunks PiperOrigin-RevId: 600601012
Use `PrefetchedSplitProvider` to prefetch the splits and write them in temporary files in parallel. When the dispatcher receives GetSnapshotSplit requests, it will just move the temporary files to the split directories. It could reduce the lock time and speed up the GetSnapshotSplit requests. PiperOrigin-RevId: 600602766
…elds. PiperOrigin-RevId: 600602767
…nsorHandleData` PiperOrigin-RevId: 600609090
…g input PiperOrigin-RevId: 600623197
The change was made public in TensorFlow 2.14 and 2.15 release notes: https://github.com/tensorflow/tensorflow/releases PiperOrigin-RevId: 600623585
This is one of many CLs to transition to the new PJRT ID APIs. - Add device lookup APIs with strong typed ID, and delegate the old ones to the new ones. - Delegate local_hardware_id() to the new one with typed ID. PiperOrigin-RevId: 600636575
PiperOrigin-RevId: 600668807
…IndexingMapSimplifier. PiperOrigin-RevId: 600687133
PiperOrigin-RevId: 600703058
PiperOrigin-RevId: 600703605
PiperOrigin-RevId: 600705896
PiperOrigin-RevId: 600720209
Imported from GitHub PR openxla/xla#8696 This PR fixes a couple of minor issues to support the XLA build against the cuDNN v9. cc. @reedwm Copybara import of the project: -- fb0ae743eafea727423dd02736214fc6f31364ee by Kaixi Hou <kaixih@nvidia.com>: Fix support to cudnn v9 Merging this change closes tensorflow#8696 PiperOrigin-RevId: 600734575
|
retest cpu-pycpp please |
|
Retest Ubuntu-GPU-single please. |
i-chaochen
approved these changes
Jan 24, 2024
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
No description provided.