hash stringlengths 40 40 | msg stringlengths 1 131k | author stringlengths 1 33 | email stringlengths 0 57 | date int64 1,447B 1,698B |
|---|---|---|---|---|
3bd0c5231c292b98474aa480655337a040f55598 | [XLA:GPU] Don't embed HloModule, HloProto & BufferAssignment in autotuning executables
This saves host RAM during compilation.
PiperOrigin-RevId: 560124483 | Tamás Danyluk | tdanyluk@google.com | 1,692,983,165,000 |
11030308c5d25df5b36f8a583f1b4607e4ea2b7f | Add a check to check if all sharding strategies are dropped due to infinity costs
PiperOrigin-RevId: 560125787 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,983,405,000 |
8eaa97bbccfbdac049ff5aec4c159279ed0b0b05 | Remove untested, unused Stream methods.
PiperOrigin-RevId: 560127191 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,983,693,000 |
bc0b42bb74185a351e4c258db47ce7f5a0272588 | Remove unused ExecutorForDeviceWithPluginConfig method.
PiperOrigin-RevId: 560130478 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,984,342,000 |
7445696c0867bf4ebec2aa866297fac71fc49e2f | When replicating small tensors, we piggyback on the existing support for user sharding annotations in auto-sharding. As implemented previously, annotations added for replicating small tensors were being dropped when user annotations are asked to not be preserved. This CL fixes this behavior and excludes annotations add... | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,984,407,000 |
ea279498128a58cf955c155932a560d8617907a1 | [XLA:GPU] Fix Triton GEMM fusion logic.
Always fuse operations reducing the input DRAM traffic even if they have multiple users.
PiperOrigin-RevId: 560138383 | Ilia Sergachev | sergachev@google.com | 1,692,985,857,000 |
2a5cbb86c5f977649ae53f07346c51600600e155 | TF C Experimental Grappler: Clean up C++ headers, includes and build_targets
Context: Removing `tf_tensor` impl from libtpu, which requires some cleanup of headers and separation of interface and implementation
PiperOrigin-RevId: 560141249 | David Silverstone | silverstone@google.com | 1,692,986,400,000 |
36f85b350dd14d3fceb63ce07ef7c861e94ecbf2 | Add `experimental_write_callbacks` to `CheckpointOptions`.
PiperOrigin-RevId: 560143549 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,986,803,000 |
7a4e103f05cee009f9b8d2fb7f4e4a8f59b6a3b2 | Add C++ gradient for ClipByValue
PiperOrigin-RevId: 560146342 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,987,377,000 |
ece328dabf48f09e73ad62e950e1e04f096c3762 | [XLA:GPU][NFC] Use more move semantics in GpuCompiler::RunBackend
PiperOrigin-RevId: 560149321 | Tamás Danyluk | tdanyluk@google.com | 1,692,987,951,000 |
52c79e03d875a43dad55ed35b185a9c357707a34 | TF C API: Break `tf_buffer`'s false dependency on `tf_status` and `tf_tensor`
PiperOrigin-RevId: 560158754 | David Silverstone | silverstone@google.com | 1,692,989,915,000 |
f176e72a86463eddc768c33bf2f6c8a021a471aa | TF Lite: 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.
Certain files would refer to incomplete types without this change.
PiperOrigin-RevId: 560160699 | David Silverstone | silverstone@google.com | 1,692,990,345,000 |
93c1d9658ed406551fd1ab05a0564f5efaef888f | Allow input >= output on size in `input_output_alias`.
PiperOrigin-RevId: 560165008 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,991,436,000 |
32556a323e976c65754546cb6e75f32c2e9d0d74 | [MemorySpaceAssignment] Make comments in tests more informative and fix the language in other comments.
PiperOrigin-RevId: 560171075 | Subhankar Shah | subhankarshah@google.com | 1,692,992,890,000 |
b95a16ad293c39d2c3a25941fffcccc9a2e29b41 | [XLA:GPU] Parametrizes Triton Softmax tests from softmax_rewriter_triton_test. Currently, F32 and F16 are tested. BF16 will be included in a followup CL which adds BF16 support.
PiperOrigin-RevId: 560177215 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,994,412,000 |
c8269ac7021cf212e49ae7c19436c49b53750da2 | Selectively compile TF_Tensor C API into TF but not libtpu
Context: Will remove `tf_tensor` impl from libtpu, which requires some cleanup of headers and separation of interface and implementation
PiperOrigin-RevId: 560180326 | David Silverstone | silverstone@google.com | 1,692,995,149,000 |
25e4a6b44e2e4292dc751625d283c88b2caf0c6e | Cleanup TF C API Kernels library dependencies and type access
Context: Will remove `tf_tensor` impl from libtpu, which requires some cleanup of headers and separation of interface and implementation
PiperOrigin-RevId: 560181290 | David Silverstone | silverstone@google.com | 1,692,995,407,000 |
f4b266951aec468c4ee0a3ffe6d4e65742699261 | Update cpuinfo dependency
PiperOrigin-RevId: 560182105 | Marat Dukhan | maratek@google.com | 1,692,995,602,000 |
20bbd21f94c6f3a65cf4d02a9f9eb574189331e7 | next_pluggable_device: don't link `tf_tensor` C API impl into libtpu
Optionally remove it in Next Pluggable Device library targets
Context: Will remove `tf_tensor` impl from libtpu, which requires some cleanup of headers and separation of interface and implementation
PiperOrigin-RevId: 560183938 | David Silverstone | silverstone@google.com | 1,692,996,076,000 |
ae5fdb710f4a39f0ea619fe0a496eac97bbcec7c | Link `tsl_status_internal` in TF Eager C API
Context: Removing `tf_tensor` impl from libtpu, which requires some cleanup of headers and separation of interface and implementation
PiperOrigin-RevId: 560185147 | David Silverstone | silverstone@google.com | 1,692,996,373,000 |
69f993dca927c35907b7edaa3b84ffee6268f1cd | PR #5240: Support CollectivePermute in fp8 gemm pattern matching
Imported from GitHub PR https://github.com/openxla/xla/pull/5240
Support AllGather in fp8 gemm pattern matching.
Copybara import of the project:
--
fb25577449faf13ef424a827b4896a6f7e2c0f53 by wenscarl <shuw@nvidia.com>:
Add CollectivePermute to recurs... | wenscarl | shuw@nvidia.com | 1,692,996,383,000 |
5221bb57353c75987029e12de666145b0a7d17fb | Fix `MetaFilename` not found error blocking compilation
PiperOrigin-RevId: 560194887 | David Silverstone | silverstone@google.com | 1,692,998,613,000 |
a6dde0fe616c4cc6cbd9cf7a9587048147237d58 | tflite-gpu: Update OpenGL MEAN kernel
There was "4 x" size increasement was made in the allocating `tile_sum`
buffer which doesn't look necessary. Rollback the change to save GL memory.
Verified with mean_test and lstm_test.
PiperOrigin-RevId: 560195197 | Terry Heo | terryheo@google.com | 1,692,998,702,000 |
59c4a04d59b73740ce04ecdd2c874a3e2c517839 | TFRT: 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: 560202484 | David Silverstone | silverstone@google.com | 1,693,000,468,000 |
418478c23b0b9510340da328a780f356214a1911 | upgrade curl 8.1.2 -> 8.2.1 which has CVE fixes | justkw | karen.wu@intel.com | 1,693,002,040,000 |
298c2b4589fa572ac373296addbf6e8f87375a05 | Fix `byte_swap_tensor` and `decode_image_op` compilation
PiperOrigin-RevId: 560207819 | David Silverstone | silverstone@google.com | 1,693,001,808,000 |
96f95574b2ebe9a3a532414676cf95abe1356851 | Support for dynamic slice in generic pattern matching for FP8 GEMMs. | Philipp Hack | phack@nvidia.com | 1,693,003,359,000 |
5bfb5698d08b30b826fbbe850fdeac005d96efda | [XLA:GPU] Triton GEMM: support more broadcasts.
PiperOrigin-RevId: 560222362 | Ilia Sergachev | sergachev@google.com | 1,693,005,716,000 |
d5bae0b38ea3273c5ffaf6e01d7130e5a32e165b | 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: 560223271 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,005,977,000 |
3e9667925736787358cb9ea6430b643016929f1b | Support for dynamic slice in generic pattern matching for FP8 GEMMs. | Philipp Hack | phack@nvidia.com | 1,693,006,817,000 |
b161ab8e91a22b855943f2c540f24495c1265992 | Runtime fallback: 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: 560226430 | David Silverstone | silverstone@google.com | 1,693,006,929,000 |
7437d3a441dc705291d7a0735b77111abe5b73f2 | Integrate LLVM at llvm/llvm-project@b825808d8b51
Updates LLVM usage to match
[b825808d8b51](https://github.com/llvm/llvm-project/commit/b825808d8b51)
PiperOrigin-RevId: 560248162 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,014,647,000 |
aab7f1000efc01753c64340982cdf6993a3410ce | Update TFRT dependency to use revision
http://github.com/tensorflow/runtime/commit/828a3e176a79cf65e1396e5c9e7a41edd1901910.
PiperOrigin-RevId: 560252838 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,016,982,000 |
743395e773e78f98d037044033adf3d2bbe93ca5 | [XLA] Remove extra late depth first policy in latency hiding scheduler when depth memory pressure mode is enabled.
Doesn't improve much the pressure in the cases it helps with and makes pressure a bit worse in other cases.
PiperOrigin-RevId: 560253026 | Marcello Maggioni | maggioni@google.com | 1,693,017,076,000 |
53ff12d3a8b2d93753664a14d6f445839631ffdc | TF Python Eager and Client: cleanup 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: 560254876 | David Silverstone | silverstone@google.com | 1,693,018,055,000 |
39806998da0d750a097f4a60095965fbc63d2c48 | Integrate StableHLO at openxla/stablehlo@1ce6d356
PiperOrigin-RevId: 560265732 | Kevin Gleason | gleasonk@google.com | 1,693,022,627,000 |
fb0a4253dce9db5490a99661f0fe36466ef7dcd3 | [Memories][PJRT:C] Add `memory_space` impl to `PjRtCApiBuffer`.
PiperOrigin-RevId: 560270176 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,024,987,000 |
d685cb2aa5e3c2400a6eef21ad7cd40e8bb3b1a5 | Add a new option to `HloXlaRuntimePipelineOptions` so that users may disable the optimization pass `remove-copies-to-out-params`.
PiperOrigin-RevId: 560271427 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,025,347,000 |
97c65c3af11c8afbf2281007094b01a977b7d031 | [HloValueSemanticsAnalysis] Adds some debugging logs.
PiperOrigin-RevId: 560277889 | Jinliang Wei | jlwei@google.com | 1,693,027,704,000 |
b7051ac07bae455df0bce42da5e62d2c0a7f574e | compat: Update forward compatibility horizon to 2023-08-26
PiperOrigin-RevId: 560308742 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,040,535,000 |
bde12fd1c2d0f3bfbf078ffc55dfee9cab2ebdec | Update GraphDef version to 1600.
PiperOrigin-RevId: 560308752 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,040,537,000 |
6e3711c2ad1999aabdf83ee7705899057681faf9 | Update `CopyInsertion::RemoveUnnecessaryCopies`.
The input argument `HloOrdering* ordering` may be redundant. This function intends to "remove as many copies as possible". Thus, it should be choose the ordering by itself. If the input HloModule is scheduled, it uses `SequentialHloOrdering` with the schedule informatio... | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,058,271,000 |
4d48739c46d8fca32550edca915bc34cc2ae7d65 | Update XNNPACK dependency
PiperOrigin-RevId: 560363589 | Marat Dukhan | maratek@google.com | 1,693,071,900,000 |
a067ae6f14ba6f88ebbe710da7aa2d01f0c4f2a8 | Update GraphDef version to 1601.
PiperOrigin-RevId: 560466641 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,126,973,000 |
fd8d0c80373f6a011cfab909ee23ce4b9c42e0c2 | compat: Update forward compatibility horizon to 2023-08-27
PiperOrigin-RevId: 560466644 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,126,973,000 |
141e70b7286368e948f608dcf024ae8350742e44 | Allow input >= output on size in `input_output_alias`.
PiperOrigin-RevId: 560593222 | Pankaj Kanwar | pkanwar@google.com | 1,693,196,868,000 |
2855e676d2aa59c00ba30c054b187cf8898c000c | Update pthreadpool dependency
PiperOrigin-RevId: 560597887 | Zhi An Ng | zhin@google.com | 1,693,198,955,000 |
a6209142c12c37db859b9d328956084a66611409 | Merge pull request #61404 from halseycamilla:patch-1
PiperOrigin-RevId: 560610167 | TensorFlower Gardener | gardener@tensorflow.org | 1,693,202,978,000 |
78da3c605db34adb9b6e6858874a368df684114d | Add sparse_gradient to sparse segment reduce ops
- Adds a sparse_gradient option to the sparse segment reduce ops. When
this is enabled, the V2 gradient op is used and the gradient is
returned as an IndexedSlices object instead of a dense tensor. | Ben Barsdell | bbarsdell@nvidia.com | 1,674,121,685,000 |
b5bbcf5ee03923f745af6aa935eb8101e5acf919 | Skip unnecessary ops in embedding_lookup_sparse
- Skips the unique + embedding_lookup calls when possible and instead
passes the embedding parameters directly to the sparse segment
reduction op, which is significantly more efficient.
- Also adds a benchmark that demonstrates a 2x speedup from this change
(on a T... | Ben Barsdell | bbarsdell@nvidia.com | 1,674,122,257,000 |
4f67933e8e69b2e2d8ceda62886af38b3c1c4c6e | Note segment reduce op/API changes in Release.md | Ben Barsdell | bbarsdell@nvidia.com | 1,691,489,580,000 |
b7636076f89e994b7de79414aa9219409cf220fb | Add two missing build dependencies | Ben Barsdell | bbarsdell@nvidia.com | 1,691,647,952,000 |
10149bc816175db822aa80b0d6008f4688444e48 | Fix ShardedVariable bug in embedding_lookup_sparse
- Fixes `TypeError: Converting ShardedVariable to tensor in embedding
lookup ops is disallowed.` | Ben Barsdell | bbarsdell@nvidia.com | 1,692,189,304,000 |
a5e55b402ffe741ceac0b64eadd2f105f1d58a1a | Merge pull request #61706 from philipphack:u_fp8_dynamicslice_xla
PiperOrigin-RevId: 560628637 | TensorFlower Gardener | gardener@tensorflow.org | 1,693,208,763,000 |
c80b1f967eec4137954f3b5fe5487215e6361dcc | Update GraphDef version to 1602.
PiperOrigin-RevId: 560644511 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,213,444,000 |
3118746eb87b6fa17b7cea0bae2619fb79fc4346 | compat: Update forward compatibility horizon to 2023-08-28
PiperOrigin-RevId: 560644514 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,213,445,000 |
3450ba7f91c87560c266b735286951b6df5fc1d1 | Add a direct dependency to executor_cache.
It seems otherwise we may get a linker error with GCC in open source.
PiperOrigin-RevId: 560647768 | Adrian Kuegel | akuegel@google.com | 1,693,214,461,000 |
765314dd0b5e3fae242c9b7441ce1bda3daa8070 | Rollforward of PR #61237: [XLA:CPU] OneDNN matmul library call for XLA HLO Dot operation.
With jaxlib's `lax_numpy_test_cpu` test failure fix from https://github.com/tensorflow/tensorflow/commit/2abff97921a566d143be67451aad85e735e9f1ab
PiperOrigin-RevId: 560654825 | Penporn Koanantakool | penporn@google.com | 1,693,216,936,000 |
0c1b7a20cf133b8cbffc8a13d5e5c1854bd785f9 | Integrate LLVM at llvm/llvm-project@3823395c9d71
Updates LLVM usage to match
[3823395c9d71](https://github.com/llvm/llvm-project/commit/3823395c9d71)
PiperOrigin-RevId: 560688296 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,228,388,000 |
daedd14eddc7f4c986586d8ed596f7b7a3d8a771 | [XLA] Remove code duplication in HloCostAnalysis::HandleFusion.
PiperOrigin-RevId: 560699652 | Oleg Shyshkov | shyshkov@google.com | 1,693,231,334,000 |
7efe317835dfbcf3e1545036d0ec64c301b8ee52 | Update TFRT dependency to use revision
http://github.com/tensorflow/runtime/commit/a9932d3d9fb803d42861e899e2906688dddb5e19.
PiperOrigin-RevId: 560704709 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,232,811,000 |
c0673629822f0d2cd518e472b396a48542a4f830 | [XLA] Use readable variable names in HloCostAnalysis::(Dynamic(Update))Slice (NFC).
PiperOrigin-RevId: 560708758 | Oleg Shyshkov | shyshkov@google.com | 1,693,233,987,000 |
8675d53f3a7a83724e0d3afffd88f934924e89b4 | Fix packing operation to handle case where the remaining columns / depth (32) is
greater than half the depth (16).
PiperOrigin-RevId: 560719167 | David Rim | davidrim@google.com | 1,693,236,760,000 |
852af318dba1c17606d29ded7df9d0c66f4f2557 | The synchronization among the group of participants for acquiring
NcclCliqueState is only needed when initializing the NCCL communicator, and
should be avoided after such initialization.
We currently only avoid this synchronization for Send and Recv as doing this
for other operations may trigger deadlock.
Enable the ... | Bixia Zheng | bixia@google.com | 1,693,238,041,000 |
adb39b04e9cb116df4659a7e2de9eea27e62f25c | Add bfloat16 registration for Multinomial op.
Motivation: bfloat16 is supported via XLA bridge / MLIR, but not as a vanilla TensorFlow op. This registration will avoid user confusion when running this op in a context that does not invoke XLA.
PiperOrigin-RevId: 560729495 | RJ Skerry-Ryan | rjryan@google.com | 1,693,239,319,000 |
21fa53b5e1f25d27126d97cc1fed21d8ac713ae7 | Merge pull request #61637 from jswag180:addSubVar
PiperOrigin-RevId: 560731085 | TensorFlower Gardener | gardener@tensorflow.org | 1,693,240,021,000 |
99737c45fdd3257cda81227d3adfb3c41a8749b0 | Add direct dependencies to implementing targets.
Otherwise we may run into linker errors.
PiperOrigin-RevId: 560732287 | Adrian Kuegel | akuegel@google.com | 1,693,239,932,000 |
bca3d1c20e736b3791460a79cd08408fc77e1b02 | Export remaining TraceType methods
PiperOrigin-RevId: 560743274 | Faizan Muhammad | fmuham@google.com | 1,693,242,319,000 |
b70a39dc1787d490a045210bf3a86d49eed5603c | Check in generated pyi files for some py_extension targets.
PiperOrigin-RevId: 560744622 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,242,561,000 |
e19a9d44e01194a83695a9775d0ce9bd356153d7 | Add experimental envs for nightly TPU wheels for python 3.9 and 3.11
PiperOrigin-RevId: 560745437 | Michael Hudgins | michaelhudgins@google.com | 1,693,242,694,000 |
f23d16d41206d4ef43eacd70c9f5114e52cf47a7 | Check in generated pyi files for some py_extension targets.
PiperOrigin-RevId: 560746001 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,242,795,000 |
c3726ec2d3975541f57ef02233a18a9f9a70a018 | Delete LegalizeTFNoFallbackPass as it's no longer used.
PiperOrigin-RevId: 560746845 | Mason Chang | masonchang@google.com | 1,693,242,959,000 |
472855df347da14ec760f397a6a57e89c473fb14 | Integrate LLVM at llvm/llvm-project@3753ea83118f
Updates LLVM usage to match
[3753ea83118f](https://github.com/llvm/llvm-project/commit/3753ea83118f)
PiperOrigin-RevId: 560747694 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,243,120,000 |
c88e17f0125bae1bfb065590a07687f739879d49 | Bump up shard_count for stochastic_cast_op_test from 10 to 20
PiperOrigin-RevId: 560753324 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,244,196,000 |
d14bacc55032ecc49054f39d60a2e8ac20175e23 | Upgrade patchelf to v0.18.0 to fix certain TF shared libaries having corrupted PT_NOTE program headers.
It seems that using Patchelf <0.12 during the [wheel building process](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/tools/pip_package/build_pip_package.sh#L255-L262) corrupts the PT_NOTE program h... | Nitin Srinivasan | srnitin@google.com | 1,693,244,377,000 |
a176c0dcda9c1a86524116d38a5e4008720bbb9c | Legalize:
`mhlo.reduce(mhlo.mul, x, 1) -> tf.prod(x)`
`mhlo.reduce(mhlo.mul, x, y) -> tf.mul(tf.prod(x),y)`
PiperOrigin-RevId: 560758707 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,245,240,000 |
561717401a57355524b885b022939415048273bf | Update TFRT dependency to use revision
http://github.com/tensorflow/runtime/commit/4e490cfffb01d39ab89e9df7ce53f1e7e63bc286.
PiperOrigin-RevId: 560760644 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,245,600,000 |
58c395f95c6a8a0be66c01e7ead3b17aa059cba9 | Clean up `tpu_initialize_util`
PiperOrigin-RevId: 560763130 | David Silverstone | silverstone@google.com | 1,693,246,020,000 |
465e14e77545f8e2d1a2f4b111ca559c80344837 | Fix tf_keras build issue.
PiperOrigin-RevId: 560771619 | Scott Zhu | scottzhu@google.com | 1,693,247,672,000 |
e02e6713bf53385d95a30d203e5a23208811c8ec | Fix TF-TPU build flags | Chandra Devarakonda | chandrasekhard@google.com | 1,693,248,417,000 |
bb8d2a8b089879f35ce7d2c1a8e26bc671718c09 | [mlgo] Check aot cross-compilation support in `tfcompile`
`tfcompile` (through `saved_model_cli`) is used as a cross-compiler by e.g. Fuchsia, and that support is particularly critical as the TF pip wheel can't be installed on all the hosts that projects that would link generated native objects could.
PiperOrigin-Rev... | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,248,798,000 |
c3afb56f4ad4f30e3fdce6db817f545c5abd3ee6 | Fixing TPU flags for py39 and py311 as well | Chandra Devarakonda | chandrasekhard@google.com | 1,693,249,099,000 |
6bff9ffb8518d28a8b431b8389188f6a60b301b2 | Add a config knob to specify the number of virtual GPU devices.
This allows users to create virtual GPU devices without specifying the memory partitions. The available memory is split equally among all virtual devices.
The field will be ignored if memory partitions are specified.
PiperOrigin-RevId: 560786017 | Changhui Lin | changhuilin@google.com | 1,693,250,787,000 |
553309d95dfceb6a661925fe130c5825a5f5851a | Cleanup build targets and move them from aliased to real targets
PiperOrigin-RevId: 560787418 | Mason Chang | masonchang@google.com | 1,693,251,141,000 |
afdb0f6bf30d81734de10569df97bc77aa2feecd | Fix the path joining for `output_dir` when a non-empty `output_dir` is passed, and only pass `module_prefix` to the API generator when`_MODULE_PREFIX` is non-empty.
PiperOrigin-RevId: 560788567 | Brian Wieder | bwieder@google.com | 1,693,251,416,000 |
d9a4ee693c968d52efc380c1ced4af328863edda | [tflite] Add stablehlo.scatter kernel to tflite schema.
PiperOrigin-RevId: 560790846 | Majid Dadashi | majiddadashi@google.com | 1,693,251,919,000 |
31c50105cf90e4f8d2cf9bd93dfe5fdf9a23a2e8 | Merge pull request #61727 from tensorflow:fix_tpu_build_flags
PiperOrigin-RevId: 560791885 | TensorFlower Gardener | gardener@tensorflow.org | 1,693,252,426,000 |
c3e3988609a1ed56b414d6a1e73c921f273eeafa | replaces ::descriptor->name() with GetTypeName() for both lite and full proto.
PiperOrigin-RevId: 560796874 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,253,232,000 |
38c775cf44683f355142c306c250012fa7d4a8c6 | Selectively compile TF_Tensor C API into TF but not libtpu
Context: Will remove `tf_tensor` impl from libtpu, which requires some cleanup of headers and separation of interface and implementation
PiperOrigin-RevId: 560798161 | Haibo Huang | hhb@google.com | 1,693,253,500,000 |
3a1175e072c5666628ca0201c3bc1810324c4ae8 | Add new install method notice for tensorflow[and-cuda]
It's now possible to install Nvidia CUDA packages via Pip dependencies, thanks to a PR from @nluehr and @trevor-m (see https://github.com/tensorflow/tensorflow/pull/59825).
PiperOrigin-RevId: 560798503 | Austin Anderson | angerson@google.com | 1,693,253,568,000 |
ddc848786bdacee1a579e9b86089bb0017b0f7a2 | [PJRT] Add `stream` argument to PjRtClient::CreateViewOfDeviceBuffer
This is in preparation for creating PjRtBuffers using the new dlpack Python interface (https://dmlc.github.io/dlpack/latest/python_spec.html)
PiperOrigin-RevId: 560798913 | Skye Wanderman-Milne | skyewm@google.com | 1,693,253,658,000 |
48baac8071ae9a8c9ed8ec8515510a49970f8089 | Use raw strings for regex strings in find_cuda_config.
Fixes build failure under Python 3.12.0rc1, which reports a SyntaxWarning for these expressions due to invalid escape sequences (e.g., \w).
PiperOrigin-RevId: 560809619 | Peter Hawkins | phawkins@google.com | 1,693,255,995,000 |
0a886799b20d0923fa7cbdba437b8095b368efef | Add type annotations.
PiperOrigin-RevId: 560812193 | Shashank Viswanadha | shashankvi@google.com | 1,693,256,517,000 |
8b394dfd83fe0a1ecf3f1dde7820fe94675f4aa8 | Add a test target in `third_party/tensorflow/compiler/xla/service/copy_insertion_test.cc`.
The added test ensures that `RemoveUnnecessaryCopies` adds control dependencies to create redundancy in copies, allowing us to remove as many copies as possible.
PiperOrigin-RevId: 560820518 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,258,297,000 |
da05ac69ffc9e55894aefba1d6c968d791d462d0 | Improve cyclic cluster error message.
PiperOrigin-RevId: 560827293 | Matthias Kramm | kramm@google.com | 1,693,259,821,000 |
ec1ea7a18cb76dd877d364c267021262d9eb370e | Merge pull request #61331 from Intel-tensorflow:akhil/bf16_mean_fix
PiperOrigin-RevId: 560827705 | TensorFlower Gardener | gardener@tensorflow.org | 1,693,260,341,000 |
cfb56352240ec0d79ae9827e3a5dc08b1681f3fd | Adds EdgeTPU Model Info
PiperOrigin-RevId: 560827850 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,259,954,000 |
8ab0413add7703f0406695967c7105e871797d88 | Remove unused `DependencyHloOrdering` variable in `CopyInsertion::RemoveUnnecessaryCopies`.
PiperOrigin-RevId: 560828077 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,260,009,000 |
0d654484b8b4030ccf144e9c993baf066689bcbf | Use different input header to simplify export
The files have an include path, reduces work needed on conversion.
PiperOrigin-RevId: 560828283 | Jacques Pienaar | jpienaar@google.com | 1,693,260,048,000 |
7c753fce91151d72f0c764734b8ba15c9e25ba0a | Don't drop out of C++ fast path if mesh pointers are not equal.
This is done by returning the same object when constructing mesh if devices.shape, axis_names and flat device list matches.
PiperOrigin-RevId: 560828993 | Yash Katariya | yashkatariya@google.com | 1,693,260,198,000 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.