hash stringlengths 40 40 | msg stringlengths 1 131k | author stringlengths 1 33 | email stringlengths 0 57 | date int64 1,447B 1,698B |
|---|---|---|---|---|
8c9ecb08fae356a2f13dd826cd8d5b73cc0a3995 | Handle bool arguments when one is a prefix of the other
We had both --run and --run_xla_backend_only.
If we set --run_xla_backend_only then we got this error:
"Couldn't interpret value _xla_backend_only=true for flag run."
PiperOrigin-RevId: 575121904 | Tamás Danyluk | tdanyluk@google.com | 1,697,785,674,000 |
db1469ac7f53b4ff77f740b399c599000eeee930 | Integrate LLVM at llvm/llvm-project@01263c6c6fb4
Updates LLVM usage to match
[01263c6c6fb4](https://github.com/llvm/llvm-project/commit/01263c6c6fb4)
PiperOrigin-RevId: 575143026 | A. Unique TensorFlower | gardener@tensorflow.org | 1,697,791,093,000 |
2fbdb9f743f8c8bd1b78ece5dcef083c9c3445cb | Merge pull request #45822 from Intel-tensorflow:gaurides/placeholder2constant
PiperOrigin-RevId: 575151524 | TensorFlower Gardener | gardener@tensorflow.org | 1,697,792,939,000 |
0eab3a9dac40d20eb05497c5f95b926705262c4b | Update GraphDef version to 1655.
PiperOrigin-RevId: 575151621 | A. Unique TensorFlower | gardener@tensorflow.org | 1,697,792,538,000 |
03b245bd9dc3eeabf41ace4d6e4964702a1ef2ab | compat: Update forward compatibility horizon to 2023-10-20
PiperOrigin-RevId: 575151656 | A. Unique TensorFlower | gardener@tensorflow.org | 1,697,792,544,000 |
6a635ef1dae1cdcdaabb77bdeb40d7bb2e364a5e | Delete unused patch files of TFRT dependency.
PiperOrigin-RevId: 575154338 | Christian Sigg | csigg@google.com | 1,697,793,141,000 |
02d13cbbe0dd8747bf6111095259a40a017a3bea | Internal Code Change
PiperOrigin-RevId: 575163214 | A. Unique TensorFlower | gardener@tensorflow.org | 1,697,795,982,000 |
2bb52e8e83da7fb6946a11df4adaa8b48b469152 | Fix non-portable code that was assuming a std::atomic_flag can be initialized with a boolean.
This fixes <https://github.com/tensorflow/tensorflow/issues/61890>.
PiperOrigin-RevId: 575174050 | Fergus Henderson | fergus@google.com | 1,697,799,468,000 |
864420736135828e6326dfc5e016b46720996088 | Expose num_partitions from SpmdPartitioner.
PiperOrigin-RevId: 575175375 | A. Unique TensorFlower | gardener@tensorflow.org | 1,697,799,863,000 |
565dfbfe56f9540a16a2997699b2cd8d2d8fcc39 | [XLA] Replace Reduce(Broadcast(x), dims, Sum()) with Broadcast(x * prod(dims)).
PiperOrigin-RevId: 575180606 | A. Unique TensorFlower | gardener@tensorflow.org | 1,697,801,584,000 |
7171b0293e10be569b5a5d017705b98086feebe5 | PR #6398: Call cuStreamSynchronize if cudaMallocAsync allocator fails on allocation
Imported from GitHub PR https://github.com/openxla/xla/pull/6398
Call cuStreamSynchronize if cudaMallocAsync allocator fails, this can reduce the chance of OOM
Copybara import of the project:
--
2a03090982017395251572fd2f0e5adca2a902f9 by Shawn Wang <shawnw@nvidia.com>:
The sync allow the driver more option to find memory. So sometimes it can find memory available after a sync.
Merging this change closes #6398
PiperOrigin-RevId: 575197415 | shawnwang18 | 35983922+shawnwang18@users.noreply.github.com | 1,697,807,352,000 |
0db57007af5e33f7979bfe7a13f37b31ff62f008 | Update TFRT dependency to use revision
http://github.com/tensorflow/runtime/commit/d2cb06f5b7913b00519f8df16a45a6772424406f.
PiperOrigin-RevId: 575207251 | Christian Sigg | csigg@google.com | 1,697,810,490,000 |
b956861dc6cd77e879b95c7b4dcb8caacf7331d7 | Update wheel_verification in official to support arm64
Update wheel_verification in official to support arm64 | Michael Hudgins | michaelhudgins@google.com | 1,697,811,625,000 |
52fdad01ad0e9742e6a84813399fe78b1f5e92f9 | Merge pull request #62119 from tensorflow:shmishra99-patch-1
PiperOrigin-RevId: 575216703 | TensorFlower Gardener | gardener@tensorflow.org | 1,697,813,642,000 |
ace76d5aa52a802797d456c53a5bb2dcfa784391 | Merge pull request #62143 from gautam1858:patch-4
PiperOrigin-RevId: 575217034 | TensorFlower Gardener | gardener@tensorflow.org | 1,697,813,976,000 |
989c77e2a2340ee55dfa5a83906cc4c2e0800171 | Header cleanups in pywrap_quantize_model.cc
Adds some missing headers (like for `absl::string_view` or `pybind11::detail::type_caster`). Also annotates `IWYU pragma: keep` for headers that are actually required.
PiperOrigin-RevId: 575219102 | Dan Suh | dansuh@google.com | 1,697,813,936,000 |
c56a3b42276bd1a236f3318dcc33984cdd0623d5 | export TFCI_PYTHON_VERSION in setup | Michael Hudgins | michaelhudgins@google.com | 1,697,817,683,000 |
458d1bfde28c059c2f364960c396e4b377249c79 | Remove previous change | Michael Hudgins | michaelhudgins@google.com | 1,697,818,028,000 |
a10cd98fd565af502a69cea7394da0b749dab09f | Update wheel_verification.bats | Michael Hudgins | michaelhudgins@google.com | 1,697,818,052,000 |
71bb4fea6cefb63e8c873605872c0f90bf4bb7f9 | Update rename_and_verify_wheels.sh | Michael Hudgins | michaelhudgins@google.com | 1,697,818,120,000 |
9b985b21eab2a839bb8a385caa9b5d697d51f6b0 | Add some commented-out load statements for a cc_proto_library build macro.
These are needed for the Google internal build.
PiperOrigin-RevId: 575238688 | Fergus Henderson | fergus@google.com | 1,697,818,947,000 |
2973185bd390b9eca7ff372676490b0e67cdd450 | Pass TFCI_PYTHON_VERSION to the docker run command
This is needed by some scripts ran within docker | Michael Hudgins | michaelhudgins@google.com | 1,697,820,370,000 |
898005b873f540de2df17baa1c472c7421f86b0e | Update wheel_verification.bats | Michael Hudgins | michaelhudgins@google.com | 1,697,820,397,000 |
b0feade52cfc9946d35a02445b15935950347a15 | Update rename_and_verify_wheels.sh | Michael Hudgins | michaelhudgins@google.com | 1,697,820,410,000 |
f6242163e772492dbc92efb3ea317f718d9a7d36 | [stream_executor] Push/pop CUDA context.
PiperOrigin-RevId: 575245583 | Chris Jones | cjfj@google.com | 1,697,820,591,000 |
4b12dc10b9e8271dd4e7e7f7a2c0c170b6841a5b | Redirect more references from the framework target to the new single-source-file targets.
PiperOrigin-RevId: 575249090 | Juan Martinez Castellanos | juanantoniomc@google.com | 1,697,821,371,000 |
6010a70637e4f8659dda0377c835280472241f55 | Increase matmul_op_test test size, since it periodically times out.
PiperOrigin-RevId: 575259580 | Antonio Sanchez | cantonios@google.com | 1,697,823,565,000 |
7aff0d8a47702bfe569acc9bb696601b97e747ee | Add DynamicEmbedding to tf_keras
PiperOrigin-RevId: 575259991 | Divya S | divyasreepat@google.com | 1,697,823,650,000 |
ad838d320482726d32286f7c7b899d33e968be41 | [XLA] Add config to not re-use the insertion buffer for forward pipelining.
It's a lambda accepting the instruction to move as input and return a bool (HloPredicate)
PiperOrigin-RevId: 575261187 | Marcello Maggioni | maggioni@google.com | 1,697,823,907,000 |
9fded73c802caa867c604368782af9ddc7488e02 | Redirect more references away from the `lib/io:lib` target and onto the new single-source-file targets.
PiperOrigin-RevId: 575265177 | Juan Martinez Castellanos | juanantoniomc@google.com | 1,697,824,695,000 |
54a440c753446d328d804faa5658c417c8c1bae8 | Display actual number of shards in `device_put` error message if it's not == 1
PiperOrigin-RevId: 575268963 | Alistair Muldal | alimuldal@google.com | 1,697,825,464,000 |
e89772b25909a55e99d640fce55a3e38a342e178 | Fix rescaling for Dot Like ops
This lowers the threshold for skipping rescaling in lowering Dot Like ops. Thus makes the result more accurate.
This should not affect prod use cases since the quantizer should produce dot like op that has exactly equal lhs_scale*rhs_scale and result_scale.
PiperOrigin-RevId: 575278763 | A. Unique TensorFlower | gardener@tensorflow.org | 1,697,827,575,000 |
3b6d8243c1ce03f446d521f53e3ac72a044903ba | Add `device_type` method to `DeviceBase` to avoid down-casting then accessing the method.
PiperOrigin-RevId: 575284482 | Haoyu Zhang | haoyuzhang@google.com | 1,697,828,767,000 |
bc19312a08bf25fbd08d7803c488a4407cdd70e9 | Merge pull request #62186 from tensorflow:wheel_verification_arm64
PiperOrigin-RevId: 575287825 | TensorFlower Gardener | gardener@tensorflow.org | 1,697,829,809,000 |
bfb2603a286d73dc928a0b724efed12ed0480dcd | Improve the regex for finding bazel commands in build logs.
This will allow it to find commands which don't have anything extra
between `bazel` and the name of the command i.e. `test|build`.
PiperOrigin-RevId: 575288632 | A. Unique TensorFlower | gardener@tensorflow.org | 1,697,829,701,000 |
862ae464e440dbd10313f4eb2f9d666c69463f1e | Split tpu:tpu_test_wrapper_test target into single-source targets.
PiperOrigin-RevId: 575294561 | Juan Martinez Castellanos | juanantoniomc@google.com | 1,697,831,216,000 |
25c61d2ca860ad7f0a842f2f9052dfe67dca9391 | #tf-data-service Fix dataset file already exists error.
The dataset files and the dispatcher state can get inconsistent if the
dispatcher fails after creating a dataset file but before updating its state.
In this case, the dispatcher state is the source of truth, and the dataset
file should be synced with the dispatcher state.
PiperOrigin-RevId: 575300343 | Yang Chen | yangchen@google.com | 1,697,832,698,000 |
0230b4d3203fecc94b8be933d2febe05231b11b0 | PR #6460: [ROCm] bf16 supports for dnn
Imported from GitHub PR https://github.com/openxla/xla/pull/6460
This PR has added existing bf16 supports from our code base to upstream.
@akuegel @ddunl
Thanks in advance!
Copybara import of the project:
--
62562dffc7f13107104cdf37c8bee1825ad04ba7 by Pavel Emeliyanenko <pavel.emeliyanenko@amd.com>:
bfloat16 support for rocm
--
7f07d3e2d02bacb89999e4d69feb2c7655969e1a by Chao Chen <cchen104@amd.com>:
more comments on UseNhwcLayoutForRocm()
Merging this change closes #6460
PiperOrigin-RevId: 575301454 | Chao | cchen104@amd.com | 1,697,832,978,000 |
22605654a200a0b261309c3a615f2f0dc04d9716 | [XlaCallModule] Add back support for version 4.
In cl/574450204 we increased the minimum supported version to 5, because after March 23rd, 2023 only version 5 and higher were serialized.
Instead of rolling back all the changes, we only adjust the minimum supported version. This restores support, except for models that have shape polymorphism.
PiperOrigin-RevId: 575315983 | George Necula | necula@google.com | 1,697,836,520,000 |
33c0f436e79d671e4fbd0aa63b449f1fb1a9be12 | #tf-data-service Update the snapshot ops counter to use the plural form.
This is more consistent with other counters.
PiperOrigin-RevId: 575317203 | Yang Chen | yangchen@google.com | 1,697,836,792,000 |
42ab814d6c49265286e0b52b50f1ddc323cc95ad | [xla] improve control-flow handling in dynamic padder
The purpose of this change is to improve handling of dynamic shapes in the
presence of control flow. Before this change, the DynamicPadder would always
choose to insert a PadToStatic op after Parameter and Infeed ops when their
output is used by a control flow op (While and Conditional). This is
undesirable in cases where the ops inside the control flow can directly consume
the dynamic shape withut requiring PadToStatic, since (among other reasons) we
do not want to call SliceToDynamic before calling these ops.
This required significant changes to DynamicPadder and DynamicDimensionInference:
- Instead of blindly inserting PadToStatic when control flow is present, we now
use HloDataflowAnalysis to find the uses of a dynamic Parameter/Infeed. If none of
the uses require PadToStatic, we do not insert it.
- An awkward case exists where a dynamic input to a While loop body is not used
in a way that requires PadToStatic, but the output of the while loop *does*
require PadToStatic. (For example, the input could be dead while the output
was produced by an Infeed followed by some elementwise operation.) In these
cases we should insert PadToStatic, to ensure the dynamic size is always
available even when the while loop condition is always false.
- To handle the above, we must run DynamicDimensionInference on while loop
bodies before running it on the computation that contains the While op. This
ensures that the dynamic sizes of the while body's outputs are inferred and we
are sure that the PadToStatic is required.
- Since we are now potentially running DynamicDimensionInference on computations
more than once, we must be sure that second and later runs do not produce
conflicting inferences. This is achieved by tracking whether the dynamic
sizes have been already inferred for an op in the output shape of the op;
i.e. when the sizes are inferred we clear the dynamic dimensions from the op's
output shape. We avoid running inferences on ops that output static shapes,
and postpone inference if the dynamic sizes of operands have not been
inferred.
The changes described above required that tf2xla's dynamic shape handling be
tightened up. The DynamicPadder changes described above require that dynamic
dimensions are always semantically correct, and tf2xla/XlaBuilder did not always
produce correct dynamic dimensions.
- Do not generate broadcasts where non-broadcast dimensions are dynamic
(i.e. where the broadcast introduces new dynamic dimensions that were not
present in the operand). These broadcasts don't make sense because there is
no operand to provide the size of the new dynamic dimension. For similar
reasons, iota should never produce dynamic dimensions.
- Change tf2xla's handling of binary ops to explicitly handle cases where an
input dimension size is <=1. This is a somewhat weird case but is necessary
to support some TF graphs. XLA only broadcasts degenerate dimensions of size
1, so binary TF ops now explicitly broadcast these dimensions when necessary.
- Correctly propagate dynamic dimensions in TPU image resize kernels.
PiperOrigin-RevId: 575334042 | A. Unique TensorFlower | gardener@tensorflow.org | 1,697,841,326,000 |
96677d5f862b35ed1c092719b998cd71a8af7cfc | [tflite] Add the stablehlo gather kernel
PiperOrigin-RevId: 575336512 | Majid Dadashi | majiddadashi@google.com | 1,697,842,034,000 |
7db699dcae3f5156f5c18ac9dfdbcd9b94073582 | [xla:gpu] Remove silent recovery from failed CUDA graph updates
Trying to silently recover from CUDA graph update errors does not work well in practice as it can silently replace an error with a performance degradation, or even worse, another error.
Instead we should fix the root cause of these errors and honestly report the status up the call chain.
PiperOrigin-RevId: 575344185 | Eugene Zhulenev | ezhulenev@google.com | 1,697,844,207,000 |
4ed3ec437354682d41959850c1bc7fc6baa13228 | Integrate LLVM at llvm/llvm-project@49af6502c6dc
Updates LLVM usage to match
[49af6502c6dc](https://github.com/llvm/llvm-project/commit/49af6502c6dc)
PiperOrigin-RevId: 575358328 | A. Unique TensorFlower | gardener@tensorflow.org | 1,697,849,164,000 |
904e48fc27cb43ff4961781dc1b0b1f5da3e44f2 | Move FFT wrapper into the DUCC dependency.
This allows us to better isolate the special
build flags that DUCC requires in order to enable
exceptions and RTTI.
PiperOrigin-RevId: 575358706 | Antonio Sanchez | cantonios@google.com | 1,697,849,328,000 |
8a0f6552fd6eb170c512dcbc4a5154d8d60d8faf | #tf-data-service Fix dataset file already exists error.
A previous fix [1] did not address all the cases. If the user
has a requested a dataset ID, this could also happen. So apply
the fix to both cases. Since the dispatcher state is the source
of truth, if the state does not have a dataset [2], any existing
dataset file should be overwritten.
[1] https://github.com/tensorflow/tensorflow/commit/25c61d2ca860ad7f0a842f2f9052dfe67dca9391
[2] https://github.com/tensorflow/tensorflow/blob/33c0f436e79d671e4fbd0aa63b449f1fb1a9be12/tensorflow/core/data/service/dispatcher_impl.cc#L586
PiperOrigin-RevId: 575358928 | Yang Chen | yangchen@google.com | 1,697,849,438,000 |
91152fd1bcebd3023c32b55b1562e2a8d5fdf6d5 | Update TFRT dependency to use revision
http://github.com/tensorflow/runtime/commit/1942aae3f632d3cf3f6bd65fddef77a79f486604.
PiperOrigin-RevId: 575363226 | A. Unique TensorFlower | gardener@tensorflow.org | 1,697,851,456,000 |
5184130f817d35136d12ea4534143eb92c57f69d | Internal change only.
PiperOrigin-RevId: 575364791 | A. Unique TensorFlower | gardener@tensorflow.org | 1,697,852,224,000 |
57e6377cf9879e33f3612f1ffd3619b6513e5296 | Update Eigen to commit:aa6964bf3a34fd607837dd8123bc42465185c4f8
PiperOrigin-RevId: 575378306 | Antonio Sanchez | cantonios@google.com | 1,697,858,409,000 |
5f2067a8b6465a25e5ac57d2d0c58baec84b7487 | Update TFRT dependency to use revision
http://github.com/tensorflow/runtime/commit/fa9afb1ad2a60a44f690a44cedf9863abe841230.
PiperOrigin-RevId: 575381723 | A. Unique TensorFlower | gardener@tensorflow.org | 1,697,860,476,000 |
ef21067179fb138c15bbf9ab87a6fe1cda4dcbed | Update TFRT dependency to use revision
http://github.com/tensorflow/runtime/commit/f6b5570b2e04978d6362a0f307982c56fb0e01cd.
PiperOrigin-RevId: 575416238 | A. Unique TensorFlower | gardener@tensorflow.org | 1,697,875,659,000 |
da960835f0d5625518e857ba5f087e223c108952 | Internal Code Change
PiperOrigin-RevId: 575417130 | A. Unique TensorFlower | gardener@tensorflow.org | 1,697,876,074,000 |
056eb8fbfb650eb32df70bc252040a4e674a10f0 | Update GraphDef version to 1656.
PiperOrigin-RevId: 575423294 | A. Unique TensorFlower | gardener@tensorflow.org | 1,697,878,939,000 |
d93d327201ad3c8c997ce1338236ace43b233b57 | compat: Update forward compatibility horizon to 2023-10-21
PiperOrigin-RevId: 575423357 | A. Unique TensorFlower | gardener@tensorflow.org | 1,697,878,960,000 |
e41e4ce374c4218ef49cd3e5b595370e392bb736 | Improve documentation of Hlo input formats for Hlo runner.
PiperOrigin-RevId: 575489893 | A. Unique TensorFlower | gardener@tensorflow.org | 1,697,916,338,000 |
bb107e15316774657be59d0e87a15a3a2f9917b9 | Internal change only.
PiperOrigin-RevId: 575528207 | A. Unique TensorFlower | gardener@tensorflow.org | 1,697,942,001,000 |
b27b5d3ea312a73b64193d02e852717dbf3c33c0 | [PJRT:C] Add PJRT_Executable_Fingerprint to support AOT compilation.
PiperOrigin-RevId: 575543476 | A. Unique TensorFlower | gardener@tensorflow.org | 1,697,949,602,000 |
05ebf196ace1103092a787c281a5186c1d44734a | Update GraphDef version to 1657.
PiperOrigin-RevId: 575577263 | A. Unique TensorFlower | gardener@tensorflow.org | 1,697,965,324,000 |
24843f557a52a70e94c2e55270537adad15202fc | compat: Update forward compatibility horizon to 2023-10-22
PiperOrigin-RevId: 575577264 | A. Unique TensorFlower | gardener@tensorflow.org | 1,697,965,324,000 |
fa7533e3e169abd941295774e97da5f0475a786c | [XLA] Fix all-reduce reassociate pattern for dynamic-slice for when both inputs of an add are the same.
PiperOrigin-RevId: 575641202 | Marcello Maggioni | maggioni@google.com | 1,698,002,678,000 |
d9cb7572cc0eb6b16838dca02f34326e95eebb23 | Merge pull request #62139 from freedomtan:fix_nnapi_delegate_cos_op
PiperOrigin-RevId: 575709674 | TensorFlower Gardener | gardener@tensorflow.org | 1,698,039,625,000 |
fa8ea72f55b0cdc2607d08be2230479e127e50f5 | Merge pull request #61877 from bvschaik:make-android-namespaces-unique
PiperOrigin-RevId: 575709696 | TensorFlower Gardener | gardener@tensorflow.org | 1,698,039,927,000 |
eb552cc9b1cdd10d130c897971c0eb48d6a6f394 | Merge pull request #62144 from tensorflow:dependabot/pip/urllib3-2.0.7
PiperOrigin-RevId: 575716733 | TensorFlower Gardener | gardener@tensorflow.org | 1,698,042,109,000 |
3e5c68c0675bd6d07368d6f64f2e84c2f0cf8722 | Add `FunctionLibraryDefinition::AddFunctionDef()` overload that allows moving the FunctionDef into the library.
PiperOrigin-RevId: 575734498 | Derek Murray | mrry@google.com | 1,698,047,723,000 |
7a0c371d29b5d73d8322a6d0ae165c3b8d14127e | [XLA:GPU] Put in a crude heuristic that fusing reduces breaks coalescing
This will eventually be replaced with a proper coalescing analysis.
PiperOrigin-RevId: 575747123 | Benjamin Kramer | kramerb@google.com | 1,698,051,556,000 |
455d082956e5ffca521fcd519bf909fcf543b8a9 | Update GraphDef version to 1658.
PiperOrigin-RevId: 575747643 | A. Unique TensorFlower | gardener@tensorflow.org | 1,698,051,723,000 |
2d782eb0365a8329d6a59fa0d12dab4ab807086b | compat: Update forward compatibility horizon to 2023-10-23
PiperOrigin-RevId: 575747672 | A. Unique TensorFlower | gardener@tensorflow.org | 1,698,051,728,000 |
dc28c324462b2d93ee24354f39c761dc900acedd | Print optional attr-dict on `stablehlo.reduce`.
In the compact version of printing, any custom attributes a user of StableHLO sets will not appear. This makes sure they appear if they exist.
PiperOrigin-RevId: 575778539 | Bart Chrzaszcz | bartchr@google.com | 1,698,061,065,000 |
a39d7df972aaedc3a2bf4a986e6e3d48f5f3eab5 | PR #6475: Fix flaky test bitcast_dtypes_expander_test
Imported from GitHub PR https://github.com/openxla/xla/pull/6475
`bitcast_dtypes_expander_test` failed when executed in Docker container
```
//xla/service:bitcast_dtypes_expander_test FAILED in 3 out of 3 in 0.7s
```
Reason - swapped variable names in `S64toS32` test
```
%shift-right-logical.11 -> %shift-right-logical.12
%constant.12 -> %constant.11
```
To resolve the issue we can use pattern based var names instead.
```
%[[VAL_10:.*]]
%[[VAL_11:.*]]
```
Copybara import of the project:
--
b99ed813b39c6b53f1aa0981dbaae060d8cb5aa8 by Alexander Pivovarov <pivovaa@amazon.com>:
Fix flaky test bitcast_dtypes_expander_test
Merging this change closes #6475
PiperOrigin-RevId: 575793343 | Alexander Pivovarov | pivovaa@amazon.com | 1,698,065,787,000 |
6f30fc1fd7ff45940aa2ce17042facde66394948 | Integrate LLVM at llvm/llvm-project@e558be51bab0
Updates LLVM usage to match
[e558be51bab0](https://github.com/llvm/llvm-project/commit/e558be51bab0)
PiperOrigin-RevId: 575794929 | Benjamin Kramer | kramerb@google.com | 1,698,066,244,000 |
a02bae3e033fc325c322b12c596411b181108cbc | Add libcublas-12-2 to prevent the libnvinfer8 dependency from pulling cuda 12.3 dependencies.
Otherwise, when pulled, the 12.3 will change the alias of /usr/local/cuda to 12.3, causing failures.
PiperOrigin-RevId: 575799576 | Michael Hudgins | michaelhudgins@google.com | 1,698,067,639,000 |
c251034d99a7e20126d7742136c93fa045dc078f | [XlaCallModule] Add more logging.
Log the op attributes on `--vmodule=xla_call_module_op=3`.
PiperOrigin-RevId: 575812286 | George Necula | necula@google.com | 1,698,071,375,000 |
3441b35bb02d0cfdb77a890dc272326aa6b735ea | Introduce tf.GeneratorDatasetRegion op.
PiperOrigin-RevId: 575813013 | Matthias Kramm | kramm@google.com | 1,698,071,568,000 |
1b57ad592d0afc7e51e56a4b32f60bff20e653c7 | Update TFRT dependency to use revision
http://github.com/tensorflow/runtime/commit/4e2f93c2c98c3ee32fd62ef7dbd5cc75c80011b8.
PiperOrigin-RevId: 575813261 | A. Unique TensorFlower | gardener@tensorflow.org | 1,698,071,640,000 |
ba095590a538bd9daed0a7186531f77de7ef0443 | [XLA:GPU] Avoid scheduling already scheduled modules.
PiperOrigin-RevId: 575813414 | Ilia Sergachev | sergachev@google.com | 1,698,071,685,000 |
68e88d2a0f72ad623e4155a3d9683ff2bc5f6e3c | [XLA:GPU] Handle NaNs correctly in minimum and maximum
Fixed that IrEmitterTriton "swallowed" NaNs in minimum and maximum.
Fixed that ElementalIrEmitter "swallowed" NaNs in the case of minimum(x, NaN) on GPU. That was likely caused by an llvm error.
PiperOrigin-RevId: 575817645 | Tamás Danyluk | tdanyluk@google.com | 1,698,072,773,000 |
c3a5443f5b3afd94fa2d1aa1b7da894756681f3d | Add forwarding shim for the C++ XNNPack plugin.
PiperOrigin-RevId: 575825788 | Fergus Henderson | fergus@google.com | 1,698,074,736,000 |
729644ec7665b3cc5b1d86bb2eb464583a25189a | Add boilerplate code to add verify clustering passes.
PiperOrigin-RevId: 575845487 | Arturo Schmidt | arturoschmidt@google.com | 1,698,079,370,000 |
319d697f87b4c5c6a9fc7cb3740c896f82a2a55a | Update TFRT dependency to use revision
http://github.com/tensorflow/runtime/commit/18a68a86884cc3b033586e725cb65557420fe3ea.
PiperOrigin-RevId: 575847883 | A. Unique TensorFlower | gardener@tensorflow.org | 1,698,079,905,000 |
4f36d36a1c14565b6221c8c90bb3e46ee26db7d5 | Touch up the requirements' updater README.
PiperOrigin-RevId: 575850088 | A. Unique TensorFlower | gardener@tensorflow.org | 1,698,080,335,000 |
ca428728015daa000ce890e3584e51d502140076 | [xla:gpu] Allow users to toggle command types that are enabled in command buffers
Before this CL the libraries supported by GPU graphs are determined a integer level. This CL allows the user to have fine-grained control on whether a library should be enabled in graphs.
Example usage: XLA_FLAGS=--xla_gpu_command_buffer_command_types=FUSION,CUBLAS
PiperOrigin-RevId: 575856977 | Anlun Xu | anlunx@google.com | 1,698,081,617,000 |
1ce130020c89d6a8c8910dad06579ec28521ab92 | Refactor Requantize in ConvertMHLOQuantToInt pass
Consolidate implementations of requantize and use the same one as TF quantizer.
PiperOrigin-RevId: 575864178 | A. Unique TensorFlower | gardener@tensorflow.org | 1,698,083,038,000 |
d734e0d9ee4462e8e51e1341e2db760d7aa9da29 | Removed the unused `device` argument.
PiperOrigin-RevId: 575866461 | Dateng Lin | datenglin@google.com | 1,698,083,460,000 |
affb9df19ef6d3a7755cfc133a97098acc8107a1 | [XLA:Runtime] Moved the FFT thunk to a new folder and removed unused dependencies, as part of a thunk clean up, and updated the necessary directories pointing to this thunk. #5758
PiperOrigin-RevId: 575867727 | A. Unique TensorFlower | gardener@tensorflow.org | 1,698,083,690,000 |
360b6e3e127156303a10d419470edab9615eea83 | Make costs of each batch accessible to RPC handler.
PiperOrigin-RevId: 575879128 | Shan Han | hanshan@google.com | 1,698,085,827,000 |
a9e19d0de09d81e1a46d32f86ec04ab5203d7bf0 | Add a 3.12 Docker container.
PiperOrigin-RevId: 575881215 | A. Unique TensorFlower | gardener@tensorflow.org | 1,698,086,261,000 |
f2a52bd533aafa688a9e9582da083187d80caa00 | Logs a warning if no default sharding strategy can be found for a given instruction, and reports the total # of such nodes.
PiperOrigin-RevId: 575887322 | A. Unique TensorFlower | gardener@tensorflow.org | 1,698,087,508,000 |
d5d0deec1c29292fe3d1a23a8bd0382d20cfe0c5 | #tf-data destroys `element` before the `mutex_lock l(*mu)` is destroyed.
This will guarantee that the iterators created inside of parallel interleave iterator get destroyed before itself.
Otherwise, `CancelThreads(true)` might finish before `element`'s destructor is called because of `outstanding_threads_finished_cond_var_.wait(l);`
PiperOrigin-RevId: 575888988 | Jim Lin | jimlintw@google.com | 1,698,087,852,000 |
4fd71da7305f7d461af2ad19a07e0271e0b9d542 | [stream_executor] Use `StreamIsCapturing` function.
PiperOrigin-RevId: 575890152 | Chris Jones | cjfj@google.com | 1,698,088,085,000 |
9e31959b058e19a7f63f5dabcec1418a80c6cc74 | Remove dependency from "GetMinibatchSplitsWithPhysicalReplica" to a flag.
The flag could have a different value in different processes, which would invalidate the inferred shape. The dependency also inflates the size of the op library, by adding a dependency on a kernel library (and transitively on MLIR), and pulls in unrelated ops which complicates wrapper generation.
PiperOrigin-RevId: 575900618 | Derek Murray | mrry@google.com | 1,698,090,375,000 |
1c22230666d36c520ee6c41fbb8be8f5e90097e3 | Fixes typo in function args comment
PiperOrigin-RevId: 575904949 | A. Unique TensorFlower | gardener@tensorflow.org | 1,698,091,347,000 |
10ac33ac80c5355f72fa0916cf87a82031d96599 | Move error logging into lower cluster to runtime ops
PiperOrigin-RevId: 575910019 | Mason Chang | masonchang@google.com | 1,698,092,396,000 |
0ed832a9ea2e3213539169322769430c85703141 | Refactor xla.bzl to not repeat deps list in xla_cc_{binary,test}
In preparation to create `xla_protos_all` to simplify further
PiperOrigin-RevId: 575911998 | David Dunleavy | ddunleavy@google.com | 1,698,092,800,000 |
57b76a1da45ba56d309e00f263fc7fdc214d7904 | Remove dependencies on other op libraries from //third_party/tensorflow/core/tpu/ops:sparse_core_ops.
These dependencies are unnecessary, and they can introduce duplicate op wrappers when we generate Python bindings.
PiperOrigin-RevId: 575921758 | Derek Murray | mrry@google.com | 1,698,094,836,000 |
2328ddf3b3439e19b68cb3650c2c45573e4706c6 | Add visibility for //learning/metadata/artifactoid/cc
PiperOrigin-RevId: 575932165 | A. Unique TensorFlower | gardener@tensorflow.org | 1,698,097,003,000 |
be859e03626080af41b54b01bfdb98a766be4c37 | Sprawling .pyi updates related to pybind11 PRs #4876.
PiperOrigin-RevId: 575941719 | Ralf W. Grosse-Kunstleve | rwgk@google.com | 1,698,099,098,000 |
cb1f99598a2e9505ddabdffcf050fde1961ceabe | Add nightly libtpu download to newer scripts
Trying to get these in sync with the current jobs, which are failing,
probably because of the wrong .so being used.
PiperOrigin-RevId: 575978433 | Austin Anderson | angerson@google.com | 1,698,108,138,000 |
860fdd1dad2e1f95b7f47d59cbc5849988a1d19e | Use random input for numerical tests in convert_tf_quant_to_mhlo_int_test
This CL adds helper functions to generate random input and compare test results.
I found that TF kernels and the lowering pass have different round schemes for quantize/rescaling etc: floor(x+0.5) vs round_nearest_even. So there maybe +/-1 errors. The lowering pass uses the latter, which is consistent with TF quantizer. (Requantize in TF kernel doesn't use the former, thus the good agreement).
Therefore I removed most Q/DQ pairs in test cases so that they don't interfere with the evaluation of other ops.
PiperOrigin-RevId: 575992943 | A. Unique TensorFlower | gardener@tensorflow.org | 1,698,113,384,000 |
ddab1a76974f56b68af2bcdfce96c68e7a784efc | Add ResourceGather to the list of quantizable ops
PiperOrigin-RevId: 576004858 | Thai Nguyen | thaink@google.com | 1,698,117,846,000 |
3a0645b863a154db0711f61bdedd2d20bb423a05 | Internal Code Change
PiperOrigin-RevId: 576031178 | A. Unique TensorFlower | gardener@tensorflow.org | 1,698,127,909,000 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.