hash stringlengths 40 40 | msg stringlengths 1 131k | author stringlengths 1 33 | email stringlengths 0 57 | date int64 1,447B 1,698B |
|---|---|---|---|---|
7224b2e421fffc4577cc61c7cf9ccc6b750f58db | Add back six pt 2 | halseycamilla | halseycamilla@google.com | 1,692,299,534,000 |
ebc7b16cea45745fdab233f320a99ff30cb89201 | Change mesh propagation from a while loop into one pass.
DTensor never updates the mesh propagation result, so we never do a second pass on PropagateMesh.
PiperOrigin-RevId: 557896300 | Tianrun Li | tianrun@google.com | 1,692,299,229,000 |
60b9ba1d6636dfbdaab87a6353281e115519bed5 | Update opensource_only.files to solve merge conflict | halseycamilla | 111537937+halseycamilla@users.noreply.github.com | 1,692,299,911,000 |
0e4a9d3fe4244918e2284b50ee13549229605812 | allocator_registry: Put a lock around instance_ to avoid races.
In process_state.cc, the setting of instance_ can race with reading it in
allocator_registry.h. Wrapping the assignment and access in a lock
prevents the race.
While here, switch thread-safety annotations to ABSL's and normalize includes
slightly in allo... | pizzud | pizzud@google.com | 1,692,299,711,000 |
b3d1d1db07cfd025b500ef5abc3adb5599c1503e | [XLA/layout assignment] Add a const interface to return the layout constraints of a computation.
PiperOrigin-RevId: 557902779 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,300,563,000 |
20bbefe3821b3ef47e691c421f0f1f5e1b41a20f | Rollback of Make `steps_per_execution` parameters settable.
In the case of trying to tune with a custom `steps_per_execution` initial heuristic, it is helpful to be able to set to a certain value.
PiperOrigin-RevId: 557904674 | Gabriel Rasskin | grasskin@google.com | 1,692,300,989,000 |
4ffd359d7e463e55ec260e90b67f66cdc5af65d8 | Migrate Python external deps from tf_http_archive | halseycamilla | halseycamilla@google.com | 1,692,080,006,000 |
10c1789ac953bb9c63a5a3c0a4c6b46b8e6ba022 | removed extra six and updated lock files | halseycamilla | halseycamilla@google.com | 1,692,119,496,000 |
74789e933f8394ebaf76e44b8d72d44efcb387eb | Removal accidental addition of absl dep | halseycamilla | halseycamilla@google.com | 1,692,123,242,000 |
870b8840611a66ea3473fb779acd29b0120c17df | add @six back as dep pt 1 | halseycamilla | halseycamilla@google.com | 1,692,294,059,000 |
f14168b2d423adef44a9ef446893f0e6959735ad | Add back six pt 2 | halseycamilla | halseycamilla@google.com | 1,692,299,534,000 |
6cf2451882a3f9fe7315c87a52f9358f5e0255c5 | lite: extend Builtin options union and adding stablehlo.concatenate, stablehlo.reshape & stablehlo.clamp support
PiperOrigin-RevId: 557920724 | Zichuan Wei | zichuanwei@google.com | 1,692,304,267,000 |
fea90548bbae9b303f7cfef6c91f1ac49eaf9d81 | Add docstring annotations that will be used to generate docstrings for generated API modules.
Created based on https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/tools/api/generator/doc_srcs.py#L38-L104
* docstring_module_name become docstring annotations in docstring of corresponding module
* docs... | Marc Fisher | fisherii@google.com | 1,692,304,576,000 |
bf58feacdbf957e72217b435eae47644f7e7560a | Merge branch 'master' into master | halseycamilla | 111537937+halseycamilla@users.noreply.github.com | 1,692,304,872,000 |
67872ada8d90df8fe343a3b72d3905aa48b05e08 | Merge pull request #61595 from Intel-tensorflow:kanvi/upgrade_onednn_3_2_1
PiperOrigin-RevId: 557924962 | TensorFlower Gardener | gardener@tensorflow.org | 1,692,305,342,000 |
e8ae17b7ec07501bbd9a7431f56a27ddc164a34d | Remove deps of xla_ops that are not xla ops.
PiperOrigin-RevId: 557927067 | Brian Wieder | bwieder@google.com | 1,692,305,572,000 |
ca3ff84649576d6f456adeccde4e8a3e1926051b | Remove lookup_ops.py's dependency on ragged_tensor.py.
PiperOrigin-RevId: 557928650 | Fiona Lang | flang@google.com | 1,692,305,916,000 |
ca4f1dcf107991c3e6d888bbf9140ad2fccc9091 | Add leading "::" to a reference to "tflite::ErrorReporter" from a macro,
to ensure that it refers to the ErrorReporter in the global 'tflite' namespace.
This is needed in case the macro happens to be called in a scope
that contains a nested 'tflite' class or namespace.
PiperOrigin-RevId: 557931617 | Fergus Henderson | fergus@google.com | 1,692,306,495,000 |
34060e48cb0cafbb4200d62ddb06367f18af7513 | undo change in onednn_matmul.cc as a related PR has been reverted | guozhong.zhuang | guozhong.zhuang@intel.com | 1,692,306,938,000 |
f5621cbf7562e68db808ba2d3be47098f2f1e40f | Fixed syntax error | halseycamilla | halseycamilla@google.com | 1,692,307,610,000 |
0d126ba49d3b2acd407a175fde6b40c9f7186bdd | Remove autograph's dependency on RaggedTensor.
PiperOrigin-RevId: 557937402 | Fiona Lang | flang@google.com | 1,692,307,737,000 |
8dc8e4378897fa06e9dce5248b65bb37fec457fa | Fix android TFlite builds.
Some new headers are being included when building the portable FFT kernels
that were missing dependencies.
Also fixed relative include paths for android builds.
PiperOrigin-RevId: 557939574 | Antonio Sanchez | cantonios@google.com | 1,692,308,212,000 |
6ae57fb9b5391d9048f9e18a96230a12116949a3 | Consider control dependency in `tuple_simplifier`.
PiperOrigin-RevId: 557943991 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,309,160,000 |
70f50a9db16b65947695b2f1b5f4bf1b2402a70c | buildifier fix | halseycamilla | halseycamilla@google.com | 1,692,309,738,000 |
f5d2600dd2c5e90d2a81577e12e7fe6203775503 | syntax fixes | halseycamilla | halseycamilla@google.com | 1,692,309,749,000 |
1d69cb2f004651b58403551b37e207fcb61674ba | Add types for session_manager.py.
PiperOrigin-RevId: 557946709 | Fiona Lang | flang@google.com | 1,692,309,777,000 |
0c3701ce12546a2708c3fd7a4dc87ac09f4d6a8e | [XLA] Add a custom interval compare that can be used for heap packing in buffer assignment.
PiperOrigin-RevId: 557948704 | Berkin Ilbeyi | berkin@google.com | 1,692,310,209,000 |
02fae87e23f8f72bf62d6d83a4ee1f1069da05ba | Removes the `mhlo.custom_call @shape_assertion` custom call which represents an assertion that the first operand (`assert_what`) evaluates to `true`. TFLite runtime kernels support shape checking and shape inference to some extent, it is safe to remove this shape assertion for now.
PiperOrigin-RevId: 557950548 | Yishuang Pang | ypang@google.com | 1,692,310,614,000 |
75d520bd3a1caa09ee410dd3c3a2e9d359d62162 | [XLA] Add functionality to the collective pipeliner to pipeline instructions with a tree of users
We want to support cases where the instructions we want to pipeline are used by a non-trivial
computations, but still culminates into a single value.
For example something like:
b = broadcast(c)
a = all-reduce(x)
m = mu... | Marcello Maggioni | maggioni@google.com | 1,692,311,619,000 |
098231c99f0ba15428b549eb13929454baa2e00d | lite: add serialization support for stablehlo.bradcast_in_dim, stablehlo convolution & stablehlo.slice
PiperOrigin-RevId: 557955699 | Zichuan Wei | zichuanwei@google.com | 1,692,311,796,000 |
d7030ffa0e0a90378f3a4efc099cc50fd07c0577 | Add type annotations to evaluator.
PiperOrigin-RevId: 557956350 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,311,940,000 |
abf4286f719b6e3d635c60f836103d2c1af9731c | fix | halseycamilla | halseycamilla@google.com | 1,692,312,761,000 |
9339255465a042f640460fa62b879226abfd65f7 | Restructures our existing Auto Sharding functionality, resulting in a couple new files:
auto_sharding_wrapper.h ---> Declares the CallSolver() interface that wraps around CP-SAT
auto_sharding_impl.cc ---> Defines an implementation of our current high-level solving "recipe"
PiperOrigin-RevId: 557961328 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,313,125,000 |
93596226a7df33dd085712d4bed9f66ee1b55e27 | Remove dependency on simple_console_for_windows | Austin Anderson | angerson@google.com | 1,692,315,287,000 |
278ae3f29d9dac8f9b6122b39252de9dcaa1165d | Use the correct root_buffer in FallbackTensor.
PiperOrigin-RevId: 557970429 | Kuangyuan Chen | chky@google.com | 1,692,315,342,000 |
46d8a64491ddf2d0fac477a8187f8d8a52c24f25 | Trace additional CUPTI events.
Adds CUPTI tracing support for the following API calls.
- cudaMemAllocManaged
- cudaMemAllocHost
- cudaMemFreeHost
- cudaHostRegister / cudaHostUnregister
PiperOrigin-RevId: 557973145 | Justin Lebar | jlebar@google.com | 1,692,316,097,000 |
5e54fec1fe3f679d15ac704d663d315c4b390570 | Merge branch 'tensorflow:master' into master | halseycamilla | 111537937+halseycamilla@users.noreply.github.com | 1,692,318,297,000 |
5d7ba5e0fc6ad965d28348d9ec62393a20dd4900 | Integrate LLVM at llvm/llvm-project@c9f8aeec5f1a
Updates LLVM usage to match
[c9f8aeec5f1a](https://github.com/llvm/llvm-project/commit/c9f8aeec5f1a)
PiperOrigin-RevId: 557981779 | Alina Sbirlea | asbirlea@google.com | 1,692,318,419,000 |
f75894d45d91f41fe4b898f4b0cf87138d1bedc9 | [xla:gpu] Fix the logic of GPU graph eviction
timed_out is called even if eviction_timeout_seconds is nullopt, so we dereference a null pointer.
PiperOrigin-RevId: 557989213 | Anlun Xu | anlunx@google.com | 1,692,320,701,000 |
7333e0d340478dfd51bf86116236d1f3fa0eb28f | Mesh propagation code cleanup
PiperOrigin-RevId: 557989977 | Tianrun Li | tianrun@google.com | 1,692,320,930,000 |
f9528e5131be3ce45b2a0c6caf7f351081e0f7b3 | Update TFRT dependency to use revision
http://github.com/tensorflow/runtime/commit/6a643728a52d7ee229687af8bfd809eb1a1a665a.
PiperOrigin-RevId: 557990925 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,321,233,000 |
44de25b19f23acb950d7d0cddef481fa2b8c8022 | generate cupti_12_0.inc. update the stub to include the right version.
snall change to stub_codegen.sh to reflect the new destination places.
PiperOrigin-RevId: 557993033 | Jie Sun | jiesun@google.com | 1,692,321,982,000 |
4771921cb9d0f57a582e08effed576f55b9eb1e3 | [pjrt] Add a kernel for 4x4 int16 transpose
Similar to the 8x8 int4 transpose, this is annoying because each load only occupies half the SSE vectors.
However, we can quickly zip together pairs to make a 2x8 matrix.
A small number of SIMD-wise zips results in the elements appearing in transposed order.
Finally, we ca... | David Majnemer | majnemer@google.com | 1,692,322,145,000 |
7c2d8100c69d10b77aa1216017707802229f0fdd | [XLA] Implemented other direction for nested dot reordering
PiperOrigin-RevId: 557994688 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,322,587,000 |
6baea5752abb4dd82bd6823173de5d9db55f3699 | Merge pull request #60424 from Tessil:toupstream/int8_int16_log
PiperOrigin-RevId: 557996933 | TensorFlower Gardener | gardener@tensorflow.org | 1,692,323,790,000 |
5da6374e97e51a8b6bab53f6e11619fe3fd77913 | Use a CPU collective to build the core id.
This is functionally the same, but we no longer need to special case the first
TPU op. (Except for the MLIR unit tests that do not initialize Mesh::tpu_core_ids)
PiperOrigin-RevId: 558001968 | Yu Feng | feyu@google.com | 1,692,325,551,000 |
2f9fe55d2454390dd6e8274f22ea2bb1bf76794d | Support resource variables of other types
Previously we only supported resource variables that corresponded to floats, this adds support for all other data types that XNNPack also supports
PiperOrigin-RevId: 558002558 | Zhi An Ng | zhin@google.com | 1,692,325,790,000 |
3d79c0e4a663a4dd846c38cb5cdc0beeb3292d83 | Pass tensors through for DeviceContext API
PiperOrigin-RevId: 558006702 | Haibo Huang | hhb@google.com | 1,692,327,518,000 |
c8d3be19e82cf61eff6532bb8e5c12fe460c9e90 | [XLA] Modified HandleReduce and HandleSlice to check whether their operand is a dot, and if so, try to reorder the operation to one of the dot's operands
PiperOrigin-RevId: 558017027 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,330,896,000 |
1dbc8d1b0126dcac5edea9a65a1fddbfe5e337ce | Add int64 shape support to ConcatOffset
This has already been done for the Tensorflow op, now also do it for the
corresponding XLA op.
PiperOrigin-RevId: 558044956 | Adrian Kuegel | akuegel@google.com | 1,692,339,592,000 |
8491c3d61af237c7fc8e18dc3a1742e92858a507 | Include missing header to `tfl_stablehlo_pass.cc`.
PiperOrigin-RevId: 558053947 | Dan Suh | dansuh@google.com | 1,692,342,250,000 |
df531162e10fd8aee84830666dd1f32c3fd2cd59 | Factor out a convenience function for creating i8:f32 uniform quantized per-axis type.
Creates a factored out `CreateI8F32UniformQuantizedPerAxisType` function, which creates a `!quant.uniform<i8:f32:axis, scales:zps>` given `axis`, `scales`, and `zps`.
PiperOrigin-RevId: 558067416 | Dan Suh | dansuh@google.com | 1,692,345,692,000 |
01cfeb16129f54603dea3e0c35edfacab0ba62c0 | fix: rename _SLICE_ERORR | Mustafa Uzun | mustafauzun0@gmail.com | 1,692,347,418,000 |
6ce7fa449b532078161c69debff1157e342ce6a8 | PR #5076: [ROCm] Updated ROCm driver with GpuContext and fixed GraphAddMemcpyD2DNode build error
Imported from GitHub PR https://github.com/openxla/xla/pull/5076
1. update rocm_driver with more GpuContext for runtime support, which aligns with cuda_driver
2. temporarily fixed GraphAddMemcpyD2DNode (we will add it fut... | Chao | cchen104@amd.com | 1,692,347,219,000 |
57bac32886c6823016a78a69cd445479eef7073f | fix: rename uncomitted_filename | Mustafa Uzun | mustafauzun0@gmail.com | 1,692,349,364,000 |
57ec65b5cadf1ee8ee348138e40a81ba6da732a0 | compat: Update forward compatibility horizon to 2023-08-18
PiperOrigin-RevId: 558081554 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,349,342,000 |
0b3d138cb6b8c96988712cbb9d74feb773b23758 | Update GraphDef version to 1592.
PiperOrigin-RevId: 558081556 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,349,342,000 |
4e2633ad3ce02e33ff2ceb1950e1c2384979a69f | [XLA:GPU][NFC] Pass AutotuneConfig to AddAutotuningPasses directly.
Other autotuning passes which are not NVPTX only can also use this same AutotuneConfig object.
PiperOrigin-RevId: 558084029 | Son Tuan Vu | vuson@google.com | 1,692,350,061,000 |
0fbb1c4ffd3f777bbb539e7d155d6c65a8618e39 | Bump PThreadPool version
PiperOrigin-RevId: 558096445 | Zhi An Ng | zhin@google.com | 1,692,354,081,000 |
8b2f67d9c815a6bd2232a6ba51b4d896b799b4d6 | Merge pull request #61278 from Intel-tensorflow:zhoulong/variant_datatype_support_for_nextpluggabledevice
PiperOrigin-RevId: 558096493 | TensorFlower Gardener | gardener@tensorflow.org | 1,692,354,578,000 |
2e4603d22756d7bf30984019d09e3903b43fbc8f | Register DEVICE_DEFAULT for Assign(Add/Sub)VariableOp | jswag180 | jefferycole100@gmail.com | 1,692,359,368,000 |
fc52938bec4834a358cdc4071e4029c4cfed2029 | use kernel size of 5 for correspondingly named tests | Christian Steinmeyer | christian-steinmeyer@users.noreply.github.com | 1,692,359,413,000 |
4174d932e30fe9b5005acd805679c8ee79c22318 | fixed error: call to 'min' is ambiguous | Chao Chen | cchen104@amd.com | 1,692,362,252,000 |
8314fc6c21046a819ff026c15bca67e3dcb86cf5 | Remove kernel definition for bias_add as it is unused.
PiperOrigin-RevId: 558129711 | Stephan Herhut | herhut@google.com | 1,692,365,548,000 |
bd1063425e6ca354ea47e7403f1abd60d63bf28a | Remove dimension padding from constants in the `Conv2D->Add(*, %cst)` pattern
PiperOrigin-RevId: 558136203 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,367,520,000 |
2ee38c91ba5e8fd27d89c0a15838d720944e7d87 | Add CPU and GPU behavior notice and example to tf.nn.embedding_lookup function | Harshavardhan Bellamkonda | harshavardhan_b@srmap.edu.in | 1,692,370,658,000 |
0897ac90cd8574a5d80f065d2f867babdef4de77 | [XLA:GPU] Measure kernel runtime of compiler/xla/service/gpu:hlo_op_profiler_run with CUPTI tracer instead of host wall time.
This is more accurate and has less noise. The disadvantage is that the current implementation requires `--config=cuda` (due to cupti_tracer limitation).
Use median kernel runtime and reduce mi... | Christian Sigg | csigg@google.com | 1,692,373,589,000 |
92b263a54ec9729f223aee3dd756d5b0bb6ad9e2 | More refactoring to comply with various TF & XLA conventions.
PiperOrigin-RevId: 558161395 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,374,630,000 |
99c395ab1325890b082b6eb84086c2e54c109514 | 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.
This clears up the path for using send/recv to implement certain
collective_permute operations.
Enable the test in collective_ops_t... | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,375,846,000 |
979e60703cb0903c86eecb98f61fcef5a384f24e | Changing visibility to call a test function
PiperOrigin-RevId: 558179763 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,378,856,000 |
678544dcbe499942d6b306c4227b95edb9b5db4a | Add kokoro's dockerfile | Austin Anderson | angerson@google.com | 1,692,379,460,000 |
c49fd3760cf025b7ebe07383ad1c41b947987518 | Mock JAX GPU backend run on single GPU and mock multiple devices.
PiperOrigin-RevId: 558182122 | Tao Wang | wangtao@google.com | 1,692,379,320,000 |
62fa497c8306727b09cd4beb821ae7aa7ee3da4e | Serialize `use_major_to_minor_data_layout_for_callbacks` as part of `ExecuteOptionsProto` and fix a bug where `strict_shape_checking` wasn't restored correctly
PiperOrigin-RevId: 558182462 | Junwhan Ahn | junwhan@google.com | 1,692,379,366,000 |
c20905bdaf9a64bde8de3425352253947793040b | Make 6 targets under tensorflow/core/ have strict dependencies.
PiperOrigin-RevId: 558182990 | Juan Martinez Castellanos | juanantoniomc@google.com | 1,692,379,475,000 |
9cba2b2c9dc3e35221ad355d2bc839eb4c4f17d7 | Use a CPU collective to build the core id.
This is functionally the same, but we no longer need to special case the first
TPU op. (Except for the MLIR unit tests that do not initialize Mesh::tpu_core_ids)
PiperOrigin-RevId: 558183402 | Yu Feng | feyu@google.com | 1,692,379,570,000 |
d2f38355f71d2f7b704bf9218e9cf8ec1cee24f1 | Merge pull request #61619 from angerson:master
PiperOrigin-RevId: 558183664 | TensorFlower Gardener | gardener@tensorflow.org | 1,692,381,050,000 |
8f1181b8994777a2948bbf390f3bc83a648b140f | Add bazel macro to include the TF_Status implementation for Tensorflow but only headers when building LibTPU
PiperOrigin-RevId: 558192103 | David Silverstone | silverstone@google.com | 1,692,381,346,000 |
e304f337d1484b39ed22a012bf7e85d554330ff3 | Update `HloModuleGroupUtil::VisitTopologicalOrder` to allow it consider the channel instructions as a group.
PiperOrigin-RevId: 558195012 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,381,919,000 |
8419c70e2e14b79ad2c514835bf19830050c528e | Fix constraints for arm32 asm in nightly
PiperOrigin-RevId: 558196391 | David Rim | davidrim@google.com | 1,692,382,209,000 |
761478d569ca555e26838c959d799a003788206b | Force inline all the calls in the module created by a custom call.
PiperOrigin-RevId: 558196979 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,382,338,000 |
2575d31687d2782a9f10cc105f00bcc9eb49d328 | [XLA] Add a pass to mark every instruction except input and output as manual sharded.
PiperOrigin-RevId: 558197268 | Marcello Maggioni | maggioni@google.com | 1,692,382,410,000 |
af602b7c79f73dfb6d1f765f6a7f62af00d4e1cc | Fix C API support for coordination service.
Previously, the old API assumes that there is no null termination character in the middle of the input. But that assumption is broken if user passes some serialized data (e.g. protobuf to string) which may contain such characters.
This fix allows users to pass in the string... | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,382,727,000 |
19b639dca63f5ccfd7e93fb28c2716677fb31db2 | Remove duplicate ducc refs in opensource_only.files | halseycamilla | 111537937+halseycamilla@users.noreply.github.com | 1,692,383,749,000 |
fa87199c133ece4d7e3f5947cec2ec0a74d2cc50 | Add debug_name attr to VarHandleOp to allow visibility of user-defined names in the runtime.
PiperOrigin-RevId: 558204285 | James Mullenbach | jmullenbach@google.com | 1,692,383,923,000 |
5777634d62861a3209ef471d72371394fbde2362 | Go: Update generated wrapper functions for TensorFlow ops.
PiperOrigin-RevId: 558206446 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,384,386,000 |
08f25c85ee91aa5fc80b7c8e181d9928dea55e60 | Add Verification pass for lowering Quantization ops/types
PiperOrigin-RevId: 558206695 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,384,430,000 |
912a0c24d9e5f68c53286622814ddfcd9d4c6d80 | Eliminates redundant constraints from the Mixed ILP formulation.
PiperOrigin-RevId: 558211637 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,385,504,000 |
a1203dd302aed09b0f0fc2a4b5ab3ec55cdf4a66 | Update XlaAotOnlyVarHandleOp to match opdef of VarHandleOp (add debug_name).
PiperOrigin-RevId: 558216983 | James Mullenbach | jmullenbach@google.com | 1,692,386,773,000 |
2186d2590111cd6e6824c1a244f01cf13c3a93e4 | Update ops-related pbtxt files.
PiperOrigin-RevId: 558232388 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,390,543,000 |
971b54a137ac2a7eb8b204cb1746e2df619ba3bf | [JAX] Implement the `stream` argument to jax.Array.__dlpack__ for CUDA GPU
Also implements jax.Array.__dlpack_device__. See
https://dmlc.github.io/dlpack/latest/python_spec.html
This requires plumbing the raw CUDA stream pointer through PJRT and
StreamExecutor (since the GPU PJRT implementation is still based on
SE).... | Skye Wanderman-Milne | skyewm@google.com | 1,692,393,589,000 |
49d8609eec3513f92a69dc082d36e14e59b695ff | [MemorySpaceAssignment] Add compilation flag to always spill buffers from alternate to default memory to check for buffer corruption in the alternate memory by validating against normal compilation results.
PiperOrigin-RevId: 558245375 | Subhankar Shah | subhankarshah@google.com | 1,692,393,592,000 |
4233bcddc18d984ceeae87d6694c53e64301dc1b | Promote 'unknown side effect' log message from INFO to WARNING.
PiperOrigin-RevId: 558261913 | Matthias Kramm | kramm@google.com | 1,692,397,410,000 |
6856d24a16f400a0d3c2a839e04210418cdc7e71 | Add a transform for create_token to after_all
At the moment, StableHLO does not have the spec and the interpreter for `create_token`. `create_token` are a essentially a special case of `after_all` with no inputs, and this pass makes `create_token -> after_all` conversion.
PiperOrigin-RevId: 558265912 | Gunhyun Park | gunhyun@google.com | 1,692,398,379,000 |
797f78446f74e4a7da1f7411b87d9c4fc7d258e3 | #tf-data-service Record metric for missing default transfer server.
PiperOrigin-RevId: 558270686 | Matt Callanan | mpcallanan@google.com | 1,692,399,667,000 |
82911dffddb58f65a442b884db8b5a0cdb7b57ad | [PJRT C API] Split profiler related C APIs to its own struct, and only initialize it in tpu_tracer.cc.
The goal is to move non-PJRT C API initializations from PJRT plugin common path to TPU specific code path.
Add a tsl_status_helper because status_helper depends on ExecutorApiFn. Also move the dependency of proto_he... | Jieying Luo | jieying@google.com | 1,692,399,918,000 |
e9391d9005ee38e0803026eeee34ecbec5995750 | Add PjRtTopologyDescription::Serialize(). This returns a serialized version of the topology for use in cache keys.
PiperOrigin-RevId: 558273377 | Parker Schuh | parkers@google.com | 1,692,400,333,000 |
013865d064ea37a163451e70ccf27e48b2b6c22c | Move Pat to RewritePattern for ReorderFakeQuantPattern, and add an additional condition to prevent to miss required FakeQuant ops
PiperOrigin-RevId: 558275628 | Doyoung Gwak | doyounggwak@google.com | 1,692,400,904,000 |
6ebfa9f0b9f6bbfca209fce28d7bce56a505ce1b | Fix missed string replacement
I don't know how I missed this, since I did a big find-and-replace. Very strange.
PiperOrigin-RevId: 558277125 | Austin Anderson | angerson@google.com | 1,692,401,304,000 |
675255b415073633e20740f42076df01333e3624 | Add a transform for dot to dot_general
At the moment, StableHLO does not have the spec and the interpreter for `dot`. The representation of `dot` ops are a proper subset of `dot_general` ops, and this pass does a `dot -> dot_general` conversion.
PiperOrigin-RevId: 558278427 | Gunhyun Park | gunhyun@google.com | 1,692,401,670,000 |
ab6a0ebd307bf55e2a6140ec0948432e587260d5 | [xla:gpu2] Add support for lowering lmhlo.case operations
PiperOrigin-RevId: 558279868 | Eugene Zhulenev | ezhulenev@google.com | 1,692,402,068,000 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.