hash stringlengths 40 40 | msg stringlengths 1 131k | author stringlengths 1 33 | email stringlengths 0 57 | date int64 1,447B 1,698B |
|---|---|---|---|---|
e8312884e13a32b25437d261c882797258ec1868 | Check in generated pyi files for some py_extension targets.
PiperOrigin-RevId: 559157635 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,726,182,000 |
d40d897bc6f9d94ba20ebb8807c87250bf3826d5 | Check in generated pyi files for some py_extension targets.
PiperOrigin-RevId: 559157880 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,726,238,000 |
21044235ceefb149722d70872950239932fd63fe | Check in generated pyi files for some py_extension targets.
PiperOrigin-RevId: 559158542 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,726,364,000 |
ef8be1c83e8efb8128bc6e271cd80e219f75f1c9 | Check in generated pyi files for some py_extension targets.
PiperOrigin-RevId: 559159625 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,726,552,000 |
4e8734d92a9d1dbad94e3ead18fb38708e3fb87c | Remove deprecated functions with no callers.
PiperOrigin-RevId: 559160225 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,726,660,000 |
113042db83881abb986e07bd68e13e7f36860d39 | Update documentation for RemoveCustomCallWithShapeAssertion conversion pattern. It is not safe to remove shape assertion in some cases. JAX may trace the program differently based on the shape polymorphism specification, for example, if the program contains a conditional on "x.shape[0] % 2 == 0" that conditional would ... | Yishuang Pang | ypang@google.com | 1,692,726,942,000 |
5230772d95e22a084be0e1eecd1b57c7117f6a88 | Create tf_keras directory (current).
PiperOrigin-RevId: 559162259 | Francois Chollet | fchollet@google.com | 1,692,727,023,000 |
12857504743627ac01722fefc5aebd7fb7827ff0 | Check in generated pyi files for some py_extension targets.
PiperOrigin-RevId: 559167100 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,727,952,000 |
3ddcdf765b2cec39580cbb183f093e1d3ce142e4 | Add QuantizeWeightsPass
This pass covers const weights used by TF/TF XLA ops with f32/bf16
precision. After quantization, a constant will be replaced by a
quantized const + ops for dequantize combination. In the quantize op,
without an identity op, const will be folded. In the dequantize op, a
cast op is required to b... | Songyi Han | syhan@google.com | 1,692,728,306,000 |
e9d815b97d849057fd32099d1810999ab8abae21 | Add an new util function for PJRT execution.
PiperOrigin-RevId: 559175820 | Changhui Lin | changhuilin@google.com | 1,692,729,413,000 |
dc03e4dfee0a079a55a5f2724d6404b01b48cf7c | Properly propagate attributes when clustering with top-level compilation markers
This fixes what attributes of top-level functions are copied when creating clusters based on top-level compilation markers.
PiperOrigin-RevId: 559178124 | Jian Cai | jiancai@google.com | 1,692,729,858,000 |
cd2362f57e2ed4ef5b569d61804c177e534bbe55 | Coordinate memory usage between model and legacy prefetch tuner.
Previously, the legacy prefetch autotune (used by default) would grow prefetch buffers without regard for the autotune memory budget. This could cause out of memory errors for memory-intensive input pipelines.
This CL adds coordination to ensure that au... | Jim Lin | jimlintw@google.com | 1,692,730,070,000 |
aa1204b2220383bb0ea6fc636c1e10e1617a55aa | Do not pass args to RendezvousDoneCallback
The args is not used for now. And it will constructs another pair of DeviceContext and cause memory leak.
PiperOrigin-RevId: 559186823 | Haibo Huang | hhb@google.com | 1,692,731,569,000 |
1e34315b0a167491ae800dce72aab208fda6c95e | Check in generated pyi files for some py_extension targets.
PiperOrigin-RevId: 559187989 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,731,803,000 |
4c9e54e132dd2e948c8038bc013abaf3f9776176 | Fix a typo in comments of import_model.cc
This is needed for jumping to matching parenthesis works correctly in vim.
PiperOrigin-RevId: 559188007 | Jian Cai | jiancai@google.com | 1,692,731,806,000 |
9c0d3aaa0bb81e9b915b1778c3330a126a4594e0 | Merge pull request #61663 from Intel-tensorflow:mraunak/rsync_resolution
PiperOrigin-RevId: 559191099 | TensorFlower Gardener | gardener@tensorflow.org | 1,692,732,828,000 |
a8eb1ad0bc1b3454652d2af82af96d9757dcf53d | Fix tflite ios build.
PiperOrigin-RevId: 559193347 | Antonio Sanchez | cantonios@google.com | 1,692,733,050,000 |
d9c7a2290c5d6c9fc173379704b38c19b15cb3f6 | Fill in Verify No Outside Compilation Markers pass to fail if we have xla_outside_compilation attribute or tf_device.launch op.
PiperOrigin-RevId: 559195841 | Mason Chang | masonchang@google.com | 1,692,733,612,000 |
c52ee137b4f9a1b28818f4831001935c5008768d | Import openai/triton from GitHub.
PiperOrigin-RevId: 559198106 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,734,124,000 |
707094616a3f253513415252e3cec8d52517c6a3 | Remove metadata from `CustomNestTraceType.__hash__` to allow for unhashable metadata.
With this change, tf.function cache can still store `CustomNestTraceType` with different metadata in the same internal hash slot (based on `CustomNestTraceType.__hash__`) and then use the equality test `CustomNestTraceType.__eq__` (w... | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,735,185,000 |
b1c864f484ecfee01915e567eee12678efb4c09f | Make all targets under tensorflow/tools/** have strict dependencies.
PiperOrigin-RevId: 559210421 | Juan Martinez Castellanos | juanantoniomc@google.com | 1,692,736,702,000 |
ad1117d5776769e5ad4925c5dcee2d3186a2ef1f | In tsl/platform/default tf_proto_library, ignore `create_kotlin_proto` parameter.
PiperOrigin-RevId: 559212313 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,737,099,000 |
27f08c2b715916ab8642641b558d8fa560957f79 | Update Eigen to commit:2873916f1ca24e3282bf6e0150545d34a16b5224
CHANGELOG
=========
2873916f1 - Rename plugin headers to .inc.
6e4d5d483 - Add IWYU private pragmas to internal headers.
328b5f908 - Add temporary macro to allow unaligned scalar UB.
a798d0765 - Fix tensor stridedlinearbuffercopy
8d9f46703 - fix boost mp ... | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,739,018,000 |
7701a45e3893bbf8c451f6437d039172db89d548 | Update XNNPACK and pthreadpool dependencies
PiperOrigin-RevId: 559222529 | Marat Dukhan | maratek@google.com | 1,692,739,150,000 |
4881499130eba8958983c8a95e538c871b472c36 | Use tf2xla implementation of RandomUniform for MLIR bridge
PiperOrigin-RevId: 559223581 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,739,372,000 |
29c3734e80ab8b291fea36b12bf9348e49741150 | Fix profiler_analysis.proto build
PiperOrigin-RevId: 559226127 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,739,942,000 |
e57f0b7b445ba273cd8b68cc81def3ad65e74d0f | Replace absl/log and absl/log:check with tsl/platform:logging
This is more portable.
PiperOrigin-RevId: 559227058 | David Silverstone | silverstone@google.com | 1,692,740,149,000 |
ef34fb4d7ef8e100da794be9fd36e543a9b0ef1a | Add `is_stateless` parameter to collectives. When false (the default), behavior is the same as before. When true, no dependencies are added and the `instance_key` should be used to distinguish collective TF ops into collective groups.
PiperOrigin-RevId: 559239478 | Michael Delorimier | mdel@google.com | 1,692,742,865,000 |
612df8d11b288180abcd5440b6065c2e369234cb | Fix the missing tensor of residual which is connected with input of subgraph when input type is uint8
PiperOrigin-RevId: 559241231 | Doyoung Gwak | doyounggwak@google.com | 1,692,743,299,000 |
025562c526cd78fdd228cdf3746692d21966d615 | [XNNPack] Support transient indirection buffers
Add new flag TFLITE_XNNPACK_DELEGATE_FLAG_TRANSIENT_INDIRECTION_BUFFER and pass it to runtime_v4. When specified, conv operators in xnnpack will not create persistent memory for indirection buffer. This can reduce memory usage especially when input image dimensions are b... | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,743,725,000 |
11819dd47852f6b085a51a7476d91ed53dfee18b | [xla:gpu] Preparation for enabling cuBLAS gemm in GPU graphs
1. Set cuda level 2 to enable cuBLAS
2. Add a fallback mechanism when graph update fails
PiperOrigin-RevId: 559244181 | Anlun Xu | anlunx@google.com | 1,692,743,979,000 |
0cbae3e8919078d75e25eeacbefbe7e6d53ea0d6 | Go: Update generated wrapper functions for TensorFlow ops.
PiperOrigin-RevId: 559246665 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,744,600,000 |
1db41d41ffe69144b232ea896c30d695594356e0 | Make region<->function conversion optionally pass arguments through while condition.
For now, this is guarded by a flag (allow-passthrough-args), to work around DTensor breakages that happen if this functionality is turned on globally.
PiperOrigin-RevId: 559249407 | Matthias Kramm | kramm@google.com | 1,692,745,277,000 |
c349be1a3a43d6077017fdf6c57dd1a7fddcba4a | Remove parallel_for/pfor.py's dependency on parsing_ops.py.
PiperOrigin-RevId: 559249520 | Fiona Lang | flang@google.com | 1,692,745,305,000 |
33bc3149cea787a25dcab33b908a926ab134dee1 | Fix dlpack breakage introduced in https://github.com/openxla/xla/commit/0b3cbfe4bc7cd68dc20924bd878cdfb2faa1a169.
I accidentally released the GIL while trying to increase a Python
refcount, oops.
PiperOrigin-RevId: 559253549 | Skye Wanderman-Milne | skyewm@google.com | 1,692,746,244,000 |
b3afb311f4474442e411ba43bc0169a6d292811e | Fix missed target in profiler_analysis.proto
PiperOrigin-RevId: 559257120 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,747,109,000 |
4c0f39ebc433a7541b8237ed5448ecbd5b602126 | Clean up compilation cache GRPC targets in tensorflow/core/tpu/kernels
PiperOrigin-RevId: 559257699 | David Silverstone | silverstone@google.com | 1,692,747,261,000 |
925501f3de8ecea3b8b2cc651dfc040dde115067 | Switch from deprecated perftools::gputools namespace to stream_executor namespace.
PiperOrigin-RevId: 559259740 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,747,755,000 |
ecc6a61c119e9a106c125cd889b17428e3d9b33a | Switch from deprecated platform_kind() function to platform()->id().
PiperOrigin-RevId: 559262098 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,748,367,000 |
18761d3fb844ea089edbb463345d4f2a17e91768 | move bridge filtering specific matchers to shared header
add tests
PiperOrigin-RevId: 559267591 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,749,846,000 |
b31a1ecbb99e37bc7bab95bbef6903e6cde9a04e | Add new pass "ColocateTPUCopyWithDynamicShape".
PiperOrigin-RevId: 559269454 | Matthias Kramm | kramm@google.com | 1,692,750,384,000 |
d5d26fd613f66606a9143f7888088b28a3c6a520 | Cast the quantized tensor and the zero point tensor to i32 to prevent underflow.
Also adds an integration test to detect this type of numeric errors.
PiperOrigin-RevId: 559270557 | Dan Suh | dansuh@google.com | 1,692,750,721,000 |
7ade094bfbaef22a79e1f7d16eca921419ae9033 | Remove FindAndLoadTpuLibrary in StreamExecutor based PJRT client.
StreamExecutor based PJRT client in JAX has been deprecated on cloud TPU. TFRT based PJRT client uses PJRT_Plugin_Initialize.
PiperOrigin-RevId: 559270586 | Jieying Luo | jieying@google.com | 1,692,750,732,000 |
2c8b29b07369e595bb57cb14c026c57bab3c2598 | Fixes more memory leaks
PiperOrigin-RevId: 559271797 | Haibo Huang | hhb@google.com | 1,692,751,121,000 |
99d2293e6546db053884d919b9ed9bd2cbc33377 | Update ops-related pbtxt files.
PiperOrigin-RevId: 559272170 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,751,250,000 |
2725337f6fd3f37116fe9bfabd92d7eb66d364ac | Add traceme for both async start and end in TF executor
PiperOrigin-RevId: 559281192 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,754,320,000 |
3a6da77b15689a6409a6da865bbbf265f742366f | Rollback enabling FP8 in XLA runtime.
Rolls back part of https://github.com/openxla/xla/commit/4c2bf94fbbe6a78b46ccdb207b40a5495108f7a6. I kept the fixes to XLA runtime but rolled back the enabling of XLA runtime for FP8.
According to NVIDIA, this made FP8 models run training ~5x slower.
PiperOrigin-RevId: 559286863 | Reed Wanderman-Milne | reedwm@google.com | 1,692,756,632,000 |
8fd7f37369cdbf1a4f68195696cf9de386a999d8 | Add _manual_axes support to NamedSharding. This is needed because
custom_partitioning may produce manually sharded axes.
PiperOrigin-RevId: 559288864 | Parker Schuh | parkers@google.com | 1,692,757,433,000 |
325a5c8562c0948147f3b1cdecdfdb5e8bda89de | Adds a flag for skipping op operands
It's useful if we want to skip partial operands of an op instead of skipping the whole op. For example, `--ops_operands_to_skip TRANSPOSE_CONV:0,2` stands for skipping the TRANSPOSE_CONV operands indexed 0 and 2.
PiperOrigin-RevId: 559298425 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,760,956,000 |
34e8d4b45ba241097a8b8cce576ef8e28c1dd07e | lite: add serialization support for stablehlo.custom_call & stablehlo.reduce
also enable convert to stablehlo flatbuffers by default
PiperOrigin-RevId: 559304245 | Zichuan Wei | zichuanwei@google.com | 1,692,763,278,000 |
81bdf378864d253e4f03bfeee354a18eac77c859 | Internal Code Change
PiperOrigin-RevId: 559310555 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,765,767,000 |
5f8bafee8489ec4796e13b6a3f1564077a6050c1 | Add compilation timer around combined bridge.
Move compilation timer into shared header.
PiperOrigin-RevId: 559310679 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,765,811,000 |
4ca59e7524ad7dcdfe272b2a265f1c8f73cdce18 | Internal Code Change
PiperOrigin-RevId: 559330749 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,773,501,000 |
15aeaf65189288ec48f94e8f7017c7a399879ab9 | Merge pull request #61670 from elfringham:fix_pip
PiperOrigin-RevId: 559335018 | TensorFlower Gardener | gardener@tensorflow.org | 1,692,775,129,000 |
4f7a3c7d5cfc2c671fb09c8cfa06acf5e7ac7b75 | [xla] remove dynamic_parameter_binding from HloModule
PiperOrigin-RevId: 559346639 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,778,445,000 |
8c5766987e1164d5b93aeb112c514088f914758e | PR #5146: [ROCM] Fix error handling in GpuDriver::GraphExecUpdate
Imported from GitHub PR https://github.com/openxla/xla/pull/5146
hipGraphExecUpdate API call fails if hip_result is expected but no error_node is provided.
Copybara import of the project:
--
e4020b4345c34f6ebfb8412be6e0c9eb4fd423f7 by Dragan Mladjenov... | Dragan Mladjenovic | Dragan.Mladjenovic@amd.com | 1,692,780,844,000 |
2d9225270ed317437ab7d329deffc6d3e5a5fae6 | Update GraphDef version to 1597.
PiperOrigin-RevId: 559356287 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,781,363,000 |
c56eb6e6e87b3f42b2704a95083942873b4b7b95 | compat: Update forward compatibility horizon to 2023-08-23
PiperOrigin-RevId: 559356294 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,781,365,000 |
d75e1e8d390d286359da2f6eb1a3f421ebca3f3d | [XLA:GPU] Add missing flag_list entry for xla_gpu_collect_cost_model_stats.
PiperOrigin-RevId: 559358035 | Oleg Shyshkov | shyshkov@google.com | 1,692,781,836,000 |
4f9c14f9538d320e3329b9c4f0d0fa9cfb6bd6fe | Update image_ops_impl.py
Updated 1-D tensor to 0-D tensor and added space | Surya | 116063290+SuryanarayanaY@users.noreply.github.com | 1,692,782,375,000 |
94eb84b051eb1435c1930ddad8f474c188a234b9 | [XLA:GPU][NFC] Refactor Triton GEMM rewriter: use a context class to hold attributes of a fusion scope.
In addition generalize the way the parameters of the dimension split are passed between dot fusion scopes and update the method checking for supported dimension orders.
PiperOrigin-RevId: 559358482 | Ilia Sergachev | sergachev@google.com | 1,692,781,979,000 |
53fb01819c8101706fb01962b85e0eee36e30b11 | [XLA:GPU] Triton GEMM: fix handling of degenerate fragments of split tensor dimensions.
PiperOrigin-RevId: 559380378 | Ilia Sergachev | sergachev@google.com | 1,692,788,827,000 |
d29ed7a10fc8afe07d29efbd209390ff589a3690 | Avoid use after move in LayoutNormalization.
We can normalize the constant without moving the literal and creating a new
constant. This avoids the use after move in ReplaceInstruction() VLOG call.
Now that we do not create a new constant anymore, we need to make sure that we do
not revisit any of the normalized users ... | Adrian Kuegel | akuegel@google.com | 1,692,790,500,000 |
4d9174b7d0f36f5e49e207d809f9081fc9726496 | Do not rely on pattern rewrites when remapping to jit.
The pattern rewriter only adds complexity and makes the outcome less
predictable, in particular as it also moves constants around. Instead,
just mutate the IR directly.
PiperOrigin-RevId: 559388404 | Stephan Herhut | herhut@google.com | 1,692,791,595,000 |
b03802f937f381681f191b775f8672469ac636f6 | [XLA:GPU][NFC] Make transitive includes direct in SoftmaxRewriterTriton.
PiperOrigin-RevId: 559391309 | Benjamin Chetioui | bchetioui@google.com | 1,692,792,507,000 |
fd9764c2c0de59b5de7fe8b2f31dd0c2b4bc288f | Add support for multiple contiguous axes to reverse
PiperOrigin-RevId: 559414146 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,799,743,000 |
a6db3ab7db96ec0a5471d5c4d025a8f3ff2d48cd | Simplify env prep steps
Previously, users were instructed to copy one of the canon env files, then
change it to avoid permissions errors. Now, since the envs are all shell scripts,
the "sample" config is used instead to select a base configuration and then
overwrite all of the special CI options. TFCI_GIT_DIR is also ... | Austin Anderson | angerson@google.com | 1,692,800,196,000 |
92b23438fc19e78ede41c358881ceadb166cd3cb | Integrate LLVM at llvm/llvm-project@c058eb998aa6
Updates LLVM usage to match
[c058eb998aa6](https://github.com/llvm/llvm-project/commit/c058eb998aa6)
PiperOrigin-RevId: 559420432 | Krasimir Georgiev | krasimir@google.com | 1,692,801,374,000 |
09d7b642ef13e13c4f07ffd59d2226a592d028f6 | Move kernel definitions for kernel generator from TF dialect to mhlo/chlo.
PiperOrigin-RevId: 559423399 | Stephan Herhut | herhut@google.com | 1,692,802,025,000 |
d95078cf844e12012f5e8419b305666fdf55da9e | Use tflite memory planner for caching 4bit weights
PiperOrigin-RevId: 559434970 | David Rim | davidrim@google.com | 1,692,804,532,000 |
03b7745a117a47e05b41eed744da0b7761c8bfbb | Add traceme for both async start and end in TF executor
PiperOrigin-RevId: 559440323 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,805,621,000 |
cc30f664ac249b889f6ff4666c40b1e44ff20219 | Update TFRT dependency to use revision
http://github.com/tensorflow/runtime/commit/a5d1d5d4d02110cabbde25a3c752877bcba77173.
PiperOrigin-RevId: 559440373 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,805,629,000 |
209593b7719aa733b1e02bc1826873cb418fc64c | Prevent null pointer being passed to memcpy
Ensure that the vector just created has valid data as this is
passed to memcpy as the destination. Passing a null pointer here
will result in a segfault. | Andrew Goodbody | andrew.goodbody@linaro.org | 1,692,806,473,000 |
782eeacba5087457ef529b4798dab612d51d92c4 | Avoid DCHECK failure in LayoutNormalization.
MakeBitcastConvertToHlo() has special handling if no convert is needed and will
just return the operand. This means no new HloInstruction is created in such a
case, and we would call SetVisited() with an instruction that we already have
marked as visited. Add a guard for th... | Adrian Kuegel | akuegel@google.com | 1,692,806,927,000 |
508145a2c1f5d72c398f83353649355cdcf3f242 | [XLA] Change approach in reverting scatter optimization in SPMD prepare.
The previous code had a typo (using rhs updates for lhs scatter).
While at it change the implementation to code that is closer to the optimization
we want to revert
from
// scatter0 = scatter(operand0, indices0, updates0)
// scatter1 = ... | Marcello Maggioni | maggioni@google.com | 1,692,806,952,000 |
299e3517be47bb4cff0be8733fdd8542784b561f | Removes standard_or_nightly(...) for TensorFlow dependencies in setup.py.
Here is how it works:
- For TF nightly release we use the nightly packages of TB, Keras, etc as provided in setup.py on the `master` branch.
- For TF release the version of Keras, TB, etc is updated from nightly to the released package only on t... | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,807,047,000 |
8ff06a116079901db5857a0e5862107a5705a583 | Simplify rendezvous function binding
With this, the `TfCThunkRendezvous` can be passed by value. Nothing to keep alive on TF side.
PiperOrigin-RevId: 559449408 | Haibo Huang | hhb@google.com | 1,692,807,526,000 |
cde89bf41432cb2da384171d6e2a2024f89a7d91 | Save one TensorInterface construction
PiperOrigin-RevId: 559456473 | Haibo Huang | hhb@google.com | 1,692,809,098,000 |
086ba24cc5d7a47ee11734851ca50ef51590bb8a | Merge pull request #61672 from tensorflow:SuryanarayanaY-patch-1
PiperOrigin-RevId: 559465907 | TensorFlower Gardener | gardener@tensorflow.org | 1,692,811,116,000 |
a37fa92c1825c3cd8dfec28cec710d0157587c39 | Integrate LLVM at llvm/llvm-project@f806be5eaae1
Updates LLVM usage to match
[f806be5eaae1](https://github.com/llvm/llvm-project/commit/f806be5eaae1)
PiperOrigin-RevId: 559468470 | Krasimir Georgiev | krasimir@google.com | 1,692,811,310,000 |
aa0cd569f93766c3fb198086c2d5dbe2e10baa8c | Add return type for distribute_lib.get_strategy().
PiperOrigin-RevId: 559469091 | Fiona Lang | flang@google.com | 1,692,811,405,000 |
990453422eb852601ab084b4c0ac4570dc4b4b90 | Make PJRT compile and execute utils visible from TFRT GPU.
PiperOrigin-RevId: 559480728 | Changhui Lin | changhuilin@google.com | 1,692,813,452,000 |
b71360e34045edc8a9fa497811913174c8bbb820 | Update TFRT dependency to use revision
http://github.com/tensorflow/runtime/commit/8a789c16688c8aca1f9a788d0870f0f863c6a458.
PiperOrigin-RevId: 559481530 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,813,600,000 |
41f7ec133c1d2c5303414b84ba96f75f0c2d3b99 | Add support for dcn collective stats in xspace_to_tools_data conversion
PiperOrigin-RevId: 559481741 | Surbhi Jain | sjsurbhi@google.com | 1,692,813,631,000 |
a481868c5bee878f786eb7cf41e5a8a7b95012ec | Clean up compilation cache GRPC targets in tensorflow/core/tpu/kernels
PiperOrigin-RevId: 559487630 | David Silverstone | silverstone@google.com | 1,692,814,665,000 |
1c962b5114ffb2cda34023dd1a77e3481a92c9f6 | Ensure that CompileOptions serializes deterministically.
CompileOptions has two serialization mechanisms: Py pickle and
SerializeAsString. Neither mechanism serializes deterministically.
Deterministic serialization (also called idempotent serialization
or in-order serialization) ensures that a given structure
serializ... | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,815,623,000 |
4d95c93e723267b1ceba25d86c633fb4b71ff24a | Add `experimental_write_callbacks` to `CheckpointOptions`.
PiperOrigin-RevId: 559493341 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,815,759,000 |
88a17fff5b379632072b75104bd1589e467bd754 | #tf-data Fix build errors caused by generating Kotlin protos.
PiperOrigin-RevId: 559495296 | Yang Chen | yangchen@google.com | 1,692,816,165,000 |
f9dff0433870952fbe273dbe3ffd4dac7f3cad78 | [tflite] Enable threading on XNNPACK Eager API for the binary ops: div, sqrt, max/min, mul, squared_diff.
PiperOrigin-RevId: 559503579 | Grant Jensen | grantjensen@google.com | 1,692,817,866,000 |
8887b96602a170af4363aed59625dacc169e7431 | Implementation for rotate + tests.
PiperOrigin-RevId: 559508489 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,818,950,000 |
c051d3f18988bd6bfec775d90cc44b5e598a727e | Remove unused deprecated enum and associated functions.
PiperOrigin-RevId: 559508955 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,819,053,000 |
4ca0940b8d704d3f7d33e3941a33fc858ec55633 | Update XNNPack dependency
Bring in fix for Bazel build on x86-64 Mac
PiperOrigin-RevId: 559510426 | Marat Dukhan | maratek@google.com | 1,692,819,384,000 |
5c0b46e52318bfb8e9d0b202da68e8c032abd774 | [TPU] Remove unused `HostTransferManager` interface
It has been replaced with C APIs to accomplish the original goal in a more secure manner.
PiperOrigin-RevId: 559525936 | David Silverstone | silverstone@google.com | 1,692,822,480,000 |
d1941b5030db12896c7579d8df756d668de9ef2b | No public description
PiperOrigin-RevId: 559531175 | Ryan M. Lefever | lefever@google.com | 1,692,823,486,000 |
796895fe5de0c11ecd1b5f936a4309c1943fc4d9 | disable streaming trace viewer if cloud vertex AI, because data are stored in blobstore and we can not add cache for serving streaming trace viewer.
PiperOrigin-RevId: 559536264 | Jie Sun | jiesun@google.com | 1,692,824,534,000 |
8950f4cdb519ba3915d7ab12f9804c552a71b11c | Implement RegionBranchOpInterface for WhileRegionOp.
PiperOrigin-RevId: 559536333 | Matthias Kramm | kramm@google.com | 1,692,824,544,000 |
8f4804efe6197e3d6f26ec18b7dcaaf5abf5e14e | Add string representation for default value markers
PiperOrigin-RevId: 559537281 | Faizan Muhammad | fmuham@google.com | 1,692,824,743,000 |
ffe5d2c2723adf969b3da78fb76ce1ceb9302f92 | [XLA:GPU] Add flag to compile all autotuning configs in one "wave"
The flag --xla_gpu_single_wave_autotuning is false by default.
If set to true, compilation will use more memory, but it can be faster because it utilizes CPU cores better.
Test: Covered by existing TritonAutotunerTest.
PiperOrigin-RevId: 559543465 | Tamás Danyluk | tdanyluk@google.com | 1,692,826,054,000 |
5ebfa80f866ca0ef06929e7ef6a6de8629fef2a7 | Remove unused GetMemAllocs interface and associated infrastructure that was never used.
PiperOrigin-RevId: 559546715 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,826,747,000 |
b074db40f5d4ee52c635337226ce374808d8b2d9 | Add API_AVAILABLE annotation to address -Wunguarded-availability warning.
PiperOrigin-RevId: 559549305 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,827,304,000 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.