hash stringlengths 40 40 | msg stringlengths 1 131k | author stringlengths 1 33 | email stringlengths 0 57 | date int64 1,447B 1,698B |
|---|---|---|---|---|
3ff7a3ba0547d21ccddd3c5c81ebf4424be6ce5e | Merge pull request #61634 from 0o001:master
PiperOrigin-RevId: 565094585 | TensorFlower Gardener | gardener@tensorflow.org | 1,694,627,245,000 |
66dbd1d599dc61f3bbe2ac9441895d0ea3931cb2 | Open source global_iter_id.cc for SparseCore.
PiperOrigin-RevId: 565097655 | Hye Soo Yang | hyey@google.com | 1,694,627,266,000 |
56edb7fc3b9b5e6f1eee637ac17b7fdce395503e | Merge pull request #58400 from SaoirseARM:toupstream/int6x8_32_accum
PiperOrigin-RevId: 565106502 | TensorFlower Gardener | gardener@tensorflow.org | 1,694,629,665,000 |
c91917d7dcac800676f24b7c6dbb8fb5c4a6ca97 | Use tf2xla implementation of SliceOp instead of MLIR.
PiperOrigin-RevId: 565111478 | A. Unique TensorFlower | gardener@tensorflow.org | 1,694,629,845,000 |
74a206527e98a8da65a8c341a8ac182b2c8ee377 | [tflite-gpu] Push select_v2 dim check up from inference to parser.
PiperOrigin-RevId: 565113938 | Grant Jensen | grantjensen@google.com | 1,694,630,316,000 |
65e5b764a4d0ed22d485959d104163012d102844 | [XLA:GPU] Add missing tolerance for BF16 tests in Triton Softmax tests.
PiperOrigin-RevId: 565119660 | A. Unique TensorFlower | gardener@tensorflow.org | 1,694,631,521,000 |
5c004b4b94d58009a1863a9155fa0c49586c34a3 | Adds a new cost component for makespan.
PiperOrigin-RevId: 565121864 | A. Unique TensorFlower | gardener@tensorflow.org | 1,694,631,939,000 |
96591d041934ee40cd3c087d00501c2727eb7589 | #tf-data Scale up `"inject_io_prefetch"` experiment to 50% job level.
PiperOrigin-RevId: 565132363 | Matt Callanan | mpcallanan@google.com | 1,694,634,387,000 |
646de0c1209b6d5ec9e329642fd61237b4cb36cb | Defines an interface for makespan evaluation.
PiperOrigin-RevId: 565134913 | A. Unique TensorFlower | gardener@tensorflow.org | 1,694,634,958,000 |
6ed6cf087d12a83b7f129b72ee5e8fc8ea771e09 | Remove PluginConfig class. It was always set to PluginConfig::kDefault.
PiperOrigin-RevId: 565140913 | A. Unique TensorFlower | gardener@tensorflow.org | 1,694,636,223,000 |
67c0625e50468e980d8c97fe5bf74870ec0e7fe5 | [XLA:GPU] Fix theoretical bug where shmem_usage > shmem_budget
PiperOrigin-RevId: 565152217 | Son Tuan Vu | vuson@google.com | 1,694,638,565,000 |
b99dffb460739f08eb0c7da0630fcdee26e57274 | Open source kernel for `GetMinibatchesInCsrWithPhysicalReplicaOp` for SparseCore.
Open source sparse_core_ops_utils*
PiperOrigin-RevId: 565156105 | Hye Soo Yang | hyey@google.com | 1,694,639,253,000 |
b090b29c0ae33e525476475959166d9da6a5c0be | Update CocoaPods specs for TFLite 2.13.0
PiperOrigin-RevId: 565157773 | Yishuang Pang | ypang@google.com | 1,694,639,619,000 |
9331ee5476b7d5ed98630dbcdef161fa6417af8f | Adding sparse_core_ops_stats_handler for recording metrics
PiperOrigin-RevId: 565158613 | Hye Soo Yang | hyey@google.com | 1,694,639,805,000 |
59e2bcf69294b07d86b2660b2c9b3f1da3299bc1 | [XLA] Add WithReplicaGroups in pattern matcher and modify tests to conform to the new pattern matching format
-Add WithReplicaGroups implementation for HloInstructionPattern to match with the collective instruction's replica groups.
PiperOrigin-RevId: 565160403 | A. Unique TensorFlower | gardener@tensorflow.org | 1,694,640,144,000 |
4f75b24b0f51e407ded4486b22d6a963497c4e7e | UnimplementedError prints the type name.
Such that users can act on the classes, adding the override methods there.
PiperOrigin-RevId: 565168942 | Yu Feng | feyu@google.com | 1,694,641,937,000 |
92691ae0ac0b62b91301db470a4c1644563631c0 | Open source op for `GetMinibatchesInCsrWithPhysicalReplica` for SparseCore.
PiperOrigin-RevId: 565168980 | Hye Soo Yang | hyey@google.com | 1,694,641,943,000 |
f80b7460db914f6f0315161c42dd2d83e6c152d0 | Use PjRtFuture returned by ExecutePortable to extend the lifetime of PjRtBuffers.
The owned PjRtBuffers in `owned_executable_args` need to live until execution is complete. Currently this is achieved by blocking until all the executable outputs are ready. However, this seemed to cause performance overheads, see b/2996... | Swachhand Lokhande | swachhand@google.com | 1,694,641,980,000 |
adcfd3f69cd74ed7bae0c19d5dee72f6ea5197cc | Merge pull request #61809 from terryheo:use-ndk-r26
PiperOrigin-RevId: 565170354 | TensorFlower Gardener | gardener@tensorflow.org | 1,694,643,558,000 |
e54cae408998e0493eda56cf35ae181742b69c95 | [XLA:GPU] Limit unroll factor for column reductions
Vectorized column reductions might exceed shmem budget. Limit the unroll factors to avoid this.
PiperOrigin-RevId: 565170403 | Son Tuan Vu | vuson@google.com | 1,694,642,267,000 |
310715d91f14bf9c9ada4f8fb12c8c814feab628 | Add a test using the FlatBuffer C API (rather than the FlatBuffer C++ API)
to construct the TFLiteSettings FlatBuffer.
PiperOrigin-RevId: 565184361 | Fergus Henderson | fergus@google.com | 1,694,645,356,000 |
caac4ac308e65c18602159097ad9a11a224b8614 | Add 5c9f72faadaca7250b341b99da358e855a8d902e from abseil-cpp.
PiperOrigin-RevId: 565187417 | Clive Verghese | cliveverghese@google.com | 1,694,646,078,000 |
96d793172d7e7892f03032d0eb260b01e2e64ff5 | Open source mesh_util_test.py
PiperOrigin-RevId: 565189870 | Yu Feng | feyu@google.com | 1,694,646,616,000 |
7cf3460cd1de6cefb50780d50394d6fdd0e863b5 | Add missing backquotes in a couple of places in the release notes.
PiperOrigin-RevId: 565191065 | Fergus Henderson | fergus@google.com | 1,694,646,921,000 |
75fb8c8e8cf3a8c5eb8744c3ce79d9571c324215 | #tf-data Provide autotune with fresh values for cpu/ram budget
The loop that runs Autotune will fetch current values for available CPU and RAM on each iteration. This helps in situations where the hardware resources available to tf.data may be vertically scaled up or down based on usage during the process' lifetime.
... | A. Unique TensorFlower | gardener@tensorflow.org | 1,694,648,619,000 |
4404175d0dd55775260496f0615b48eb83f0c0d9 | Add configurable retries for SetServerDef, for fault tolerance amidst preemptions.
There's a short period during ParameterServerStrategy initialization / cluster connection in which worker preemptions will lead to UnavailableErrors from CreateContext calls. This adds configurable retries to SetServerDef so that a sing... | James Mullenbach | jmullenbach@google.com | 1,694,653,510,000 |
0d333c8ae021981e3c9be121834709267d16f2c3 | Change 2/6 for making MSA repacking slice aware.
Fix a bug in which we over-allocate space for slices, when they are colocated with larger buffers.
The interaction causing this behavior is as follows:
A) GlobalDecreasingSizeBestFitHeap::FindChunkCandidates() adds additional space to the last chunk in a sliced allocat... | Ryan M. Lefever | lefever@google.com | 1,694,654,249,000 |
ebe4f498df55278020b5c7a78d6a98b13b990294 | Fix comment bug in tsl's clean_dep
PiperOrigin-RevId: 565223624 | Jake Harmon | jakeharmon@google.com | 1,694,656,259,000 |
d8a17da3f037ce0d5c97a8ee256c3c314fdee910 | Change 3/6 for making MSA repacking slice aware.
Changed SlicedAllocationFinder to
- accept a method to determine if allocations a permitted to begin at a given offset
- expose a method to test if a sliced allocation can fit at a specific offset
PiperOrigin-RevId: 565231151 | Ryan M. Lefever | lefever@google.com | 1,694,659,269,000 |
20196d5398856168158d9d214d7a36d28570fe91 | Add highway c++ library as a dep to TensorFlow
PiperOrigin-RevId: 565237467 | Hye Soo Yang | hyey@google.com | 1,694,661,303,000 |
ab7c193a3d87ad5d8679e600aa511468d7bc2506 | lite: enable group conv -> conv2d conversion
PiperOrigin-RevId: 565268295 | Zichuan Wei | zichuanwei@google.com | 1,694,672,454,000 |
f3b30f5a0b191b09b818ae0b7afebc782f1a868d | compat: Update forward compatibility horizon to 2023-09-14
PiperOrigin-RevId: 565301845 | A. Unique TensorFlower | gardener@tensorflow.org | 1,694,682,126,000 |
f17c9f5ef5830d86705f3f3884665ad98bbf9cc2 | Update GraphDef version to 1619.
PiperOrigin-RevId: 565301847 | A. Unique TensorFlower | gardener@tensorflow.org | 1,694,682,127,000 |
ac9aa167ff6526ebf265c685324d229bfadf5aef | Select correct FC 8x16 path
PiperOrigin-RevId: 565302587 | Alan Kelly | alankelly@google.com | 1,694,682,328,000 |
2736ef7a654f737095aa9b1ace8d810766165560 | [XLA:GPU] Handle kTranspose in optimized HLO in Triton emitters.
PiperOrigin-RevId: 565307989 | Ilia Sergachev | sergachev@google.com | 1,694,684,089,000 |
e585ea7203a26cb0b10e702b4c9f193fc97325e8 | Added placeholder for internal RPC option.
PiperOrigin-RevId: 565319212 | A. Unique TensorFlower | gardener@tensorflow.org | 1,694,687,461,000 |
3da3565572613e3e00cab8ad5bf16738ef1cc684 | [Linaro:ARM_CI] Update to clang-17
Update compiler in use to be clang-17 to
maintain sync with other builds. | Andrew Goodbody | andrew.goodbody@linaro.org | 1,694,447,361,000 |
cd3d3c25b481e71fe6b41da896a0337bd0ca29d6 | Add StableHLO Quantizer as an option in TF Quantizer.
PiperOrigin-RevId: 565339452 | Jiyoun (Jen) Ha | jiyounha@google.com | 1,694,694,476,000 |
b24592bcda677a0725080af48b0370869766d958 | Const-correctness fixes for GetFusionRoots.
The output is mutable for no good reason, which causes issues when we want
to express "this fusion instruction's roots or the instruction if it's not
a fusion".
PiperOrigin-RevId: 565341699 | Johannes Reifferscheid | jreiffers@google.com | 1,694,695,287,000 |
98b0549d683f987f361a7e11713091f88741c959 | Integrate LLVM at llvm/llvm-project@bf8fd086d09c
Updates LLVM usage to match
[bf8fd086d09c](https://github.com/llvm/llvm-project/commit/bf8fd086d09c)
PiperOrigin-RevId: 565351116 | Benjamin Kramer | kramerb@google.com | 1,694,698,493,000 |
9d833cd42c29df337da1dd166b70119e03516469 | Internal Code Change
PiperOrigin-RevId: 565353457 | A. Unique TensorFlower | gardener@tensorflow.org | 1,694,699,249,000 |
ad428bcfb9008917dd276a834d3f5933ecdc3e1f | Fixes to avoid deadlocks with collectives in the pipelining while loop
PiperOrigin-RevId: 565355251 | Pat Notz | patn@google.com | 1,694,699,813,000 |
c6b4dd5ef2a48afeefe6abf8228ecdd31fa46786 | [xla] Rename LatencyHidingSchedulerPreparation pass to P2PSchedulePreparation
pass.
This is because the pass is needed to linearize point-to-point Send and Recv
chains an HLO scheduler.
Modify the GPU HLO scheduler to call P2PSchedulePreparation pass regardless
whether the latency hiding scheduler is on.
PiperOrigin... | Bixia Zheng | bixia@google.com | 1,694,704,912,000 |
a69eea2814d9ff3ec9a286c0816f8367d1816a51 | Internal changes only.
PiperOrigin-RevId: 565383371 | A. Unique TensorFlower | gardener@tensorflow.org | 1,694,707,065,000 |
704c1918a05574215d0499a74905f9a43fab01d5 | Open source op for `GetMinibatchesInCsrWithPhysicalReplica` for SparseCore.
PiperOrigin-RevId: 565384260 | A. Unique TensorFlower | gardener@tensorflow.org | 1,694,707,271,000 |
7718a3c01fec88c22c87be5385545dd2928f3c41 | [tf2xla] remove useless code.
PiperOrigin-RevId: 565395488 | A. Unique TensorFlower | gardener@tensorflow.org | 1,694,709,766,000 |
f6ece021bb02d666a81345426e4e19c1ce8b816c | update with the correct checksum | justkw | karen.wu@intel.com | 1,694,710,778,000 |
d25fabc24de08c453aa52ecbdf5ca0ce160cd170 | Uses slightly more descriptive loop iterator names (i -> node_idx, or i -> edge_idx) when processing members of the solver request object.
PiperOrigin-RevId: 565398513 | A. Unique TensorFlower | gardener@tensorflow.org | 1,694,710,407,000 |
c137d4b055e4abb44af53ef200847a09aa199bdb | [xla] Extend collective-pipeliner to support the pipelining of Send and Recv
chains.
Extend collective-pipeliner to support the backward pipelining of Recv-done,
which has the effect of pipelining a Send and Recv chain.
Extend the HLO verifier to support multiple Send and Recv pairs using the same
channel IDs, which ... | Bixia Zheng | bixia@google.com | 1,694,710,839,000 |
24a9884fac5646a459594d1193c7c796e3f3c4a5 | Define StableHLO RngBitGenerator as builtin op to schema and add a reference kernel for it.
PiperOrigin-RevId: 565413735 | Yishuang Pang | ypang@google.com | 1,694,713,422,000 |
9fa010483866844ff9b2237e454e00848f58b351 | [tflite] Temporarily add back mhlo.scatter to TF patterns
This broke some tests outside of tflite and darwinn. Temporarily adding back these patterns until the test issues are resolved.
PiperOrigin-RevId: 565415946 | A. Unique TensorFlower | gardener@tensorflow.org | 1,694,713,857,000 |
1ec5ed423d64fd1d15fee2fec86d5798f4bf6d05 | Support for GPU Memory Streamz Metrics
PiperOrigin-RevId: 565417401 | A. Unique TensorFlower | gardener@tensorflow.org | 1,694,714,138,000 |
48b73ff00f3b1523bdc6c17087bdecb07bb95daa | [TF:PJRT] Use AcquireExternalReference to create a Tensor with PjRtTensorBuffer.
Before this change, only PjRtStreamExecutorBuffer can be used to create a Tensor with PjRtTensorBuffer. After this change, a Tensor with PjRtTensorBuffer can be created with any PJRT client that supports AcquireExternalReference.
PiperOr... | Jieying Luo | jieying@google.com | 1,694,714,622,000 |
f50884936e6e5db528c2f5ab2b8566326fef72c3 | [mlir][sparse][xla] rename files for future extension.
PiperOrigin-RevId: 565428745 | A. Unique TensorFlower | gardener@tensorflow.org | 1,694,716,302,000 |
2ecb3fb78c49f5863ceb91f1d81e280076e73645 | [PJRT C API] Support passing allocator_config as an option in PJRT GPU plugin.
PiperOrigin-RevId: 565429986 | Samuel Agyakwa | sagyakwa@google.com | 1,694,716,563,000 |
32fd4805aba5fe028b2c36cdab9d14f32a3bb273 | #tf-data-service Update mutex lock to shared lock.
The dispatcher gets blocked when there are too many workers.
PiperOrigin-RevId: 565430053 | Yang Chen | yangchen@google.com | 1,694,716,580,000 |
66230bb61fc33eefce9d6427121e1560b93596f6 | Support for legalizing convolutions with dynamic batch dimension
PiperOrigin-RevId: 565431519 | A. Unique TensorFlower | gardener@tensorflow.org | 1,694,716,907,000 |
1ffe70e2ab61fc22e792b640db53a541b153a8ac | Invoke vendoring logic for Windows
PiperOrigin-RevId: 565459002 | Jake Harmon | jakeharmon@google.com | 1,694,722,939,000 |
bff19b28da2f12be395b72874c2c8d36547c30f7 | #tf-data-service Simplify `"no_compression"` experiment implementation.
PiperOrigin-RevId: 565461943 | Matt Callanan | mpcallanan@google.com | 1,694,723,539,000 |
72df60c5451ccec269d071cbff30038abb4e7d0b | Change 4/6 for making MSA repacking slice aware.
Split FindChunkCandidates() into 3 methods:
- GetMaxColocationSize
- CreateSlicedAllocationFinder
- PostProcessFindChunkCandidatesResult
We do this so that slice-aware repacking can make use of the individual methods. In particular, when more than 1 sliced allocation i... | Ryan M. Lefever | lefever@google.com | 1,694,725,526,000 |
3b721239cf375b378610dc8cea2b676faf24530b | [PJRT C API] Implement PjRtCApiBuffer::is_dynamic_dimension().
Store is_dynamic_dimension_ at PjRtCApiBuffer by reconstructing from the number of dimensions and output of calling PJRT_Buffer_DynamicDimensionIndices.
Also change PJRT_Buffer_DynamicDimensionIndices so that it only sets dyn_dim_indices the first time it... | Jieying Luo | jieying@google.com | 1,694,726,279,000 |
37c476e382db559ab5a30423b8b03a4512b82245 | Add support for float flags to compiler_options.
PiperOrigin-RevId: 565475731 | Parker Schuh | parkers@google.com | 1,694,726,344,000 |
39fcac026233e4f803c3e9c9b3054dcd2d89c5fc | #tf-data-service Use absl::Status in snapshot_manager.
PiperOrigin-RevId: 565478401 | Yang Chen | yangchen@google.com | 1,694,726,959,000 |
a5c2f0a3e8c8ffca543f7ec13d90bc7049fa69cc | Refactors the solution extraction logic into its own function.
PiperOrigin-RevId: 565489501 | A. Unique TensorFlower | gardener@tensorflow.org | 1,694,729,452,000 |
d52a753c6a51466af18db3dc58da6251608e4a8f | Merge pull request #61817 from DrMarcII:apigen
PiperOrigin-RevId: 565497974 | TensorFlower Gardener | gardener@tensorflow.org | 1,694,732,013,000 |
0761e7831af8be457b31f879d2768dd2911ea387 | Remove GenericFunction class in favor of PolymorphicFunction
PiperOrigin-RevId: 565512443 | Faizan Muhammad | fmuham@google.com | 1,694,734,774,000 |
66c2a25e1c754754b5eadc515aa39cfd1d512076 | Open source op for `GetMinibatchesInCsrWithPhysicalReplica` for SparseCore. Rollback of rollback.
PiperOrigin-RevId: 565519907 | Hye Soo Yang | hyey@google.com | 1,694,736,700,000 |
e53dd8e0b1eaf84c346fd01a73de0fbfd429dd3c | Add per-compilation resource context to LoadingResult.
Similar to LoadedClientGraph in GraphExecutor, each compiled graph should have its own resource context, so that the resource inside that references compiled graph wouldn't conflict with each other.
PiperOrigin-RevId: 565522943 | Kuangyuan Chen | chky@google.com | 1,694,737,594,000 |
d8d267596bbf592a5043eb27ef92f0a21df94dd1 | #tf-data-service Make SnapshotManager thread-safe.
Currently, one mutex guards the snapshot manager for all snapshots.
This is suspected to cause contention when the dispatcher is
responsible for multiple snapshots and many workers.
PiperOrigin-RevId: 565526669 | Yang Chen | yangchen@google.com | 1,694,738,814,000 |
01982d9ebe9b22184d75823906e43cb7d873d9a1 | Merge pull request #61799 from Intel-tensorflow:karen/upgrade-sqlite
PiperOrigin-RevId: 565545139 | TensorFlower Gardener | gardener@tensorflow.org | 1,694,745,874,000 |
b89baa4e1f45dc90fa348785452ceac0a916fc65 | Merge pull request #61855 from cjflan:cf/release-test-enhancments
PiperOrigin-RevId: 565555875 | TensorFlower Gardener | gardener@tensorflow.org | 1,694,749,878,000 |
5bef7ce69551df9ee3af189d9606c19a70d12d57 | Merge pull request #61833 from elfringham:clang17
PiperOrigin-RevId: 565556493 | TensorFlower Gardener | gardener@tensorflow.org | 1,694,750,313,000 |
8468eb8ab4973ce659796ba6c04bfc6ab3455225 | Internal Code Change
PiperOrigin-RevId: 565564801 | A. Unique TensorFlower | gardener@tensorflow.org | 1,694,752,804,000 |
2438d19bb086114ca27b627e56c5b5e25418621e | Open source op & kernel for ConvertToCooTensor
PiperOrigin-RevId: 565581273 | Hye Soo Yang | hyey@google.com | 1,694,758,003,000 |
f7c4ee46f532521a9e581f31436eac70da34d266 | Internal Code Change
PiperOrigin-RevId: 565605737 | A. Unique TensorFlower | gardener@tensorflow.org | 1,694,765,922,000 |
c6e386f07aa6d5f084e4db8b69e17e91ce7d34c1 | Clang Migration Windows | Raunak | mayank.kumar.raunak@intel.com | 1,694,768,540,000 |
3a62569b5c3a092fab42b4da40819f8ee59777fd | Update GraphDef version to 1620.
PiperOrigin-RevId: 565614806 | A. Unique TensorFlower | gardener@tensorflow.org | 1,694,768,730,000 |
9a21fcf571cab1a34435c72ac3508910fd8533ef | compat: Update forward compatibility horizon to 2023-09-15
PiperOrigin-RevId: 565614823 | A. Unique TensorFlower | gardener@tensorflow.org | 1,694,768,736,000 |
fffde0e8430c91c6f23611ee93d557e746a4b941 | [XLA:GPU][NFC] Remove outdated documentation files
PiperOrigin-RevId: 565617828 | Tamás Danyluk | tdanyluk@google.com | 1,694,769,729,000 |
674a70a9064b8007da65c471fc88c9b0b9a819cb | [XLA:GPU] Set the debug options on the extracted HLOs before doing the SplitK transform
This fixes the --xla_gpu_triton_gemm_disable_reduced_precision_reduction flag for autotuning.
PiperOrigin-RevId: 565634784 | Tamás Danyluk | tdanyluk@google.com | 1,694,775,373,000 |
ac70a0829c9e0dc7faaef850c11689a0a957a4fa | [XLA:GPU] Handle more split-k sizes by padding
Now we also allow applying the split-k transformation when the contracting dimension is not divisible by the number of splits.
We allow this by padding the contracting dimension to be divisible. The padding is not materialized. The "padding" values are masked out by sett... | Tamás Danyluk | tdanyluk@google.com | 1,694,777,632,000 |
de48970c033f95e69fbb03c489eb78f818ce0d23 | Add a new pass in the converter for MHLO -> TFLite conversions.
PiperOrigin-RevId: 565643475 | Quentin Khan | qkhan@google.com | 1,694,778,392,000 |
3452c11a32f54a0b1b7ee6d6fd6b903acb76f6f1 | Integrate LLVM at llvm/llvm-project@ae84b160d2fe
Updates LLVM usage to match
[ae84b160d2fe](https://github.com/llvm/llvm-project/commit/ae84b160d2fe)
PiperOrigin-RevId: 565658144 | Benjamin Kramer | kramerb@google.com | 1,694,783,284,000 |
ae92196dd38afd7d9be89b7efa80dcfac3b53d0d | Remove an unused #include.
PiperOrigin-RevId: 565671584 | Fergus Henderson | fergus@google.com | 1,694,787,731,000 |
c12af26ab68e4255620d824783470e18c6a45618 | [XLA:Python] Remove the interpreter client.
PiperOrigin-RevId: 565672207 | Peter Hawkins | phawkins@google.com | 1,694,787,927,000 |
8fcc1b465a354c2277dc84b7503443013c693541 | [SE] [NFC] Remove unused method
PiperOrigin-RevId: 565673897 | George Karpenkov | cheshire@google.com | 1,694,788,467,000 |
e96e46a6ef82231bce1f3e4c437b4d9fcee8b1de | Boundary functions for producer/consumer fusion.
PiperOrigin-RevId: 565678576 | Johannes Reifferscheid | jreiffers@google.com | 1,694,789,885,000 |
5a3af90ee76b3b2b5b8e5af47d7237cbc3cd8be9 | Always log when delegate cannot be applied due to dynamic tensors
PiperOrigin-RevId: 565683384 | Alan Kelly | alankelly@google.com | 1,694,791,196,000 |
775bb390c5bfe437cc4e710d6594c745d0002887 | Global rename of tf2xla/api/v1 to tf2xla/api/v2 to align with tf1/tf2
PiperOrigin-RevId: 565689095 | Mason Chang | masonchang@google.com | 1,694,792,689,000 |
324920bc7f800169cd879d27fdcbb776062c2610 | Integrate LLVM at llvm/llvm-project@3d51010a3350
Updates LLVM usage to match
[3d51010a3350](https://github.com/llvm/llvm-project/commit/3d51010a3350)
PiperOrigin-RevId: 565693453 | Benjamin Kramer | kramerb@google.com | 1,694,793,864,000 |
6ed239d72a8453c773a5d1702ce9fd7a1ce04e1f | [mlir][sparse][xla] rename some sparse primitives' name.
PiperOrigin-RevId: 565700544 | A. Unique TensorFlower | gardener@tensorflow.org | 1,694,795,621,000 |
1f7baa3b563872170bb64a96230e1ec3e0b940e1 | Add a generic pattern to lower MHLO quant ops
This pattern handles a selection of ops that require uq->int type change only. Use this pattern to consolidate and replace some previous patterns. It also add checks to ensure all UQ types in operands and results are the same.
PiperOrigin-RevId: 565707010 | A. Unique TensorFlower | gardener@tensorflow.org | 1,694,797,169,000 |
613c27c4d630288b3025790e1e94b3f57b810541 | Load rules_license separately
PiperOrigin-RevId: 565710394 | Jake Harmon | jakeharmon@google.com | 1,694,797,884,000 |
a9ae004fd0ca1eeae1d1e4a08b72491d6dd1cd5d | Internal change only.
PiperOrigin-RevId: 565712539 | Fiona Lang | flang@google.com | 1,694,798,375,000 |
e41bd3f8ddf9a85356f10977b432afb4b4f6d6a6 | Explicitly check if the input module to auto-sharding uses manual sharding, and crash if so as we currently do not support handling such modules.
PiperOrigin-RevId: 565712685 | A. Unique TensorFlower | gardener@tensorflow.org | 1,694,798,403,000 |
a0943fb1d9fda11ffc4ee5c2c05adaa1518544f6 | [XLA] Add a smoke test for the sort comparator.
The comparator needs to satisfy the strict weak ordering requirement, otherwise std::sort() may crash.
We can't really robustly verify this without considering all triples, but at least we can smoke-test on the first element that the comparator is not reflexive.
PiperO... | A. Unique TensorFlower | gardener@tensorflow.org | 1,694,799,242,000 |
a074d946391e9d9fd9c39d4f08b621f9af3dcab9 | Internal Code Change
PiperOrigin-RevId: 565719155 | A. Unique TensorFlower | gardener@tensorflow.org | 1,694,799,841,000 |
03dd8124043037ff8384b7b739ea892352fd02a5 | Undo the MHLO::BroadcastInDimOp folding pattern on splat tensor. This helps reduce model size. This pass slightly changes the patterns for average pooling:
```
1. div(reduce_window(add), const_divisor) -> div(reduce_window(add), broadcast_in_dim(const_divisor))
2. div(reduce_window(add), reduce_window(const_1, init_va... | Yishuang Pang | ypang@google.com | 1,694,800,264,000 |
2ab61291628c4cbce673bce3e360ae3b8819a0a1 | Optimize the performance of `ConvertToCooTensorOp`.
This CL combines several optimizations:
1. If the combiner is "sum", we avoid all computation and allocation related to gain-rescaling.
2. If the weights are a scalar, we broadcast the same weight to all tokens. This will avoid the need to execute a `Shape`->`Fill` ... | Derek Murray | mrry@google.com | 1,694,800,325,000 |
57c7009bd72ed8d7591a896eda33df19b3e5e007 | [XLA] Fallback to V1 sharding when it's not possible to preserve V2 in PartialTile.
PiperOrigin-RevId: 565723524 | Ce Zheng | zce@google.com | 1,694,800,693,000 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.