hash stringlengths 40 40 | msg stringlengths 1 131k | author stringlengths 1 33 | email stringlengths 0 57 | date int64 1,447B 1,698B |
|---|---|---|---|---|
cc390d943023314588f939f4a7bf2758ef8e837a | [stream_executor] CommandBuffer: add APIs for recording kernel launch and memcpy operations
This implementation is based on gpu_graph.{h.cc} and for CUDA backend adds CUDA graph nodes to the graph under construction.
https://github.com/openxla/xla/issues/5857
PiperOrigin-RevId: 568382083 | Eugene Zhulenev | ezhulenev@google.com | 1,695,692,225,000 |
d8b5490ded4b24da09d7fe08d8a6b116b3e9d452 | Remove bazel from jax's docker image and use the ones preinstalled in the sigbuild image
PiperOrigin-RevId: 568383806 | Yash Katariya | yashkatariya@google.com | 1,695,692,961,000 |
cf5e85663fef097f83f5004f112ae42758e9b035 | [stream_executor] NFC: Restrict visibility of stream_executor_pimpl target
`stream_executor_pimpl` is an internal implementation detail planned to be removed. Remove all external users of this target and clean up some of the build files to remove dependencies on other internal targets planned to be removed.
https://g... | Eugene Zhulenev | ezhulenev@google.com | 1,695,696,230,000 |
f947c9c19700e84c255a344ee0d03cc45cf9fd04 | TensorListElementShape kernel implementation.
PiperOrigin-RevId: 568396257 | Luke Boyer | lukeboyer@google.com | 1,695,697,796,000 |
c7b5860dbcc8ad8d8236d0a52603643d5886097a | Adding helper method to allow setting profile keys map.
PiperOrigin-RevId: 568413206 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,704,552,000 |
f0f6eb81e7ed9e835f0ae59194ea389abd4cfe2a | Fixes build error with clang compiler.
PiperOrigin-RevId: 568431979 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,709,377,000 |
42af713661bf7324c0a8e9b2ee21750ec417cd5e | PR #5881: [XLA:GPU] remove unused buffers to fix cudnn graph
Imported from GitHub PR https://github.com/openxla/xla/pull/5881
This PR is to fix the bug introduced in commit https://github.com/openxla/xla/pull/5184/commits/25aa1baa87abb259e3e0350737c4d39a715183a0 in runner clean up PR: https://github.com/openxla/xla/p... | Shanbin Ke | shanbinke@gmail.com | 1,695,709,478,000 |
8eba8a15c600fea12eb5f403eeed56ce9c315a03 | [XLA:GPU] Fall back to cuBLAS in TritonAutotuner if that's faster
PiperOrigin-RevId: 568440756 | Tamás Danyluk | tdanyluk@google.com | 1,695,711,936,000 |
8b579914b4b4ebeebdb93728fc06f256fa64bd64 | [stream_executor] NFC: Make stream_executor:executor_cache implementation private
PiperOrigin-RevId: 568446172 | Eugene Zhulenev | ezhulenev@google.com | 1,695,713,482,000 |
2f94ec0c2ed3d16d846f0244c6f8ddcb1ae56e86 | compat: Update forward compatibility horizon to 2023-09-26
PiperOrigin-RevId: 568466907 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,719,013,000 |
a6815d063b081eab127143de6316681847fecfb0 | Update GraphDef version to 1631.
PiperOrigin-RevId: 568466908 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,719,013,000 |
c2ae731c249b9ad452304a505929a7de8faf12a8 | Make SimplifyFPConversions check creation pass ids
This is adding an additional model to `SimplifyFPConversions` which will only
simplify chains of converts if they were generated by a previous optimization pass.
The pass is checking whether these convert ops were created by an optimization pass
or part of the input ... | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,719,191,000 |
a1fa36df2097fed9b3f912ca02b737dde5ba81e2 | [XLA:GPU] Fix detection of unsupported transposes in Triton GEMM fusions.
PiperOrigin-RevId: 568470689 | Ilia Sergachev | sergachev@google.com | 1,695,720,028,000 |
6c4d42bc1f1de15486ef50a54b99543a6ae9c679 | PR #5473: [NVIDIA XLA GPU] Added an pass to unroll loops by a factor of 2 to achieve double buffering
Imported from GitHub PR https://github.com/openxla/xla/pull/5473
This only deals with loops with known trip count.
For loops with even trip count, we unroll directly.
For loops with odd trip count, we take out 1 iter... | TJ Xu | tjx@nvidia.com | 1,695,722,927,000 |
87ce74262ac34a46d29113850c643d722c0655a3 | PR #5867: [ROCm] fixed rocm kernel link
Imported from GitHub PR https://github.com/openxla/xla/pull/5867
Fixed a recent build error due to the missing link at rocm gpu executor and some updates on rocm_driver
@akuegel @ddunl Thanks in advance!
Copybara import of the project:
--
63801cf1b46aee77ccc272e0eb6562436755... | Chao | cchen104@amd.com | 1,695,724,252,000 |
ae01629d90f6f8ef3a04de41bb459b1f5e475df6 | [XLA:GPU] Trigger Triton GEMM fusions also on kCopy input operations.
PiperOrigin-RevId: 568489918 | Ilia Sergachev | sergachev@google.com | 1,695,725,632,000 |
079ab56a5b0c5aefcdffa4f20f79c7bb137c6bd0 | Retire MLIR interpreter and related tooling.
This interpreter no longer makes sense inside XLA/isn't actively used. Retire here. Ongoing discussion with upstream to potentially move there post some refactoring, but can be recovered from commit history if and when needed rather than maintain until then.
PiperOrigin-Re... | Johannes Reifferscheid | jreiffers@google.com | 1,695,727,712,000 |
9a704c508833ca68caf272d4f44cbe944d10f8c7 | Rename FindFusionParameters -> FindFusionArguments.
Parameters are inside fusions, but this finds instructions outside the fusion, so this name is more accurate and less confusing.
PiperOrigin-RevId: 568505665 | Johannes Reifferscheid | jreiffers@google.com | 1,695,730,593,000 |
ff8e8010f77dbe98d5e811693f04e14e7cfe85d3 | Remove tracing instrumentation.
It's dead.
PiperOrigin-RevId: 568514448 | Johannes Reifferscheid | jreiffers@google.com | 1,695,733,355,000 |
c45c4c9f39b35d8b64ed0545da66a3042cb67a52 | Split initial Triton IR generation from TritonWrapper and add a Matmul filecheck test.
I'll add more test cases later.
PiperOrigin-RevId: 568534317 | Johannes Reifferscheid | jreiffers@google.com | 1,695,738,880,000 |
c5e2eff71408e935d82a2c18ebb9c22b5f9db390 | Update docker image hash in Dockerfile.
PiperOrigin-RevId: 568538855 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,740,109,000 |
aad923ba91272ffb67f3541a84da2fe21d8abefb | Integrate LLVM at llvm/llvm-project@69ba4aed004d
Updates LLVM usage to match
[69ba4aed004d](https://github.com/llvm/llvm-project/commit/69ba4aed004d)
PiperOrigin-RevId: 568544781 | Sam McCall | sammccall@google.com | 1,695,741,486,000 |
7cf4309dea6c3a6dc6cd7b416dc9e5c20b6a8fe7 | Add bisect & any scripts, allow Docker rebuilds, integrate bazelrcs
Documentation for these will come soon.
PiperOrigin-RevId: 568558257 | Austin Anderson | angerson@google.com | 1,695,744,747,000 |
38a9352c0ec79bc200222ea5e6ad4cdd2d7fa72f | Update TFRT dependency to use revision
http://github.com/tensorflow/runtime/commit/ad4ba135f9a5ca24862daf7c93e4a3b6e53c69c8.
PiperOrigin-RevId: 568559126 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,744,930,000 |
f16b9c216905ed9443b8c1e5a37a45efe66b76b3 | Clean up dependency and disable test from running under msan.
PiperOrigin-RevId: 568576434 | Gunhyun Park | gunhyun@google.com | 1,695,748,486,000 |
6bb02ebc7d9c88e35c0ef79135de7cc029e92863 | [HloValueSemanticsAnalysis] Handle while loops in einsum depth analysis.
PiperOrigin-RevId: 568578511 | Denali Molitor | dmolitor@google.com | 1,695,748,870,000 |
ee61b61f6cce45b30542a859e27a6db3791b9ab2 | [stream_executor] NFC: Remove external dependencies on stream_executor internal targets
PiperOrigin-RevId: 568583065 | Eugene Zhulenev | ezhulenev@google.com | 1,695,749,726,000 |
14f74dbd85fcc01654d05c79a34a883709ecb460 | [XLA:GPU][NFC] Pass CompileOptions directly to Run* from LLVMCompiler::Compile
The option should be used instead of just the device allocator.
PiperOrigin-RevId: 568590287 | Son Tuan Vu | vuson@google.com | 1,695,751,101,000 |
ba1bfa60c37c4a142ab170bcf67228af9c5492e2 | Don't fuse the tf.Split op and add a new attr to tf.TPUCompileAndExecute to indicate number of arguments per partition so that the kernel knows how to distribute the execution.
PiperOrigin-RevId: 568591194 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,751,255,000 |
6bf2c9b23c83d45da51b4340872dfc01d12297ef | Remove import of debug gradients in gradients_impl.
Due to the debug init file, the debug gradient imports cause a cycle with gradients_impl.py.
Consequently, regenerate pywrap_gradient_exclusions to not include debug gradients, under the assumption that debug gradients would not be used without first importing debug... | Fiona Lang | flang@google.com | 1,695,752,009,000 |
297ec1f82d90599d24033a0bbb677f79565ffeb9 | Release GIL while destructing IFRT objects
Holding GIL while destructing IFRT executables is particularly problematic since destructing `LoadedExecutable` requires waiting for all in-flight host callbacks to complete, which may be blocked by GIL.
PiperOrigin-RevId: 568596172 | Junwhan Ahn | junwhan@google.com | 1,695,752,228,000 |
e77f4b676cb98ccf0a252d6fd869cd1a978aa434 | Move replicate_constants_pass to POST_REWRITE_FOR_EXEC
To support TF1 feeding/fetching (in the "rewrite for execution"), where nodes (including Const nodes) can be replaced by a feed node, the replicate_constants_pass should be done after "rewrite for execution" so it only applies to nodes that remain Const nodes afte... | Edward Schwartz | schwartzedward@google.com | 1,695,753,622,000 |
742e788aaabec6f9769d7c728765ac919cac5ecb | Extract TPU Bridge pipeline into its own file
PiperOrigin-RevId: 568606553 | Mason Chang | masonchang@google.com | 1,695,754,199,000 |
5d485ce42700b5c184214ad7e3e8228094619867 | [stream_executor] NFC: Make command_buffer, event and kernel implementations private
PiperOrigin-RevId: 568606765 | Eugene Zhulenev | ezhulenev@google.com | 1,695,754,235,000 |
8adeb5ffa55dd02df9ce20c35624998ac718b593 | Add StreamExecutorUnloadedExecutable (de)serialization and `LoadSerializedExecutable` API.
PiperOrigin-RevId: 568607896 | Shixin Li | shixinli@google.com | 1,695,754,475,000 |
15b882e5c31531f72eeeddf5f411516eec3df9e9 | Add Sparse Core ops to MLIR.
PiperOrigin-RevId: 568607967 | Hye Soo Yang | hyey@google.com | 1,695,754,490,000 |
45cc6dd29bab7d8c79864097cc64fd9eb3126def | [stream_executor:cuda] Add a minimal test for launching CUDA kernels
PiperOrigin-RevId: 568624489 | Eugene Zhulenev | ezhulenev@google.com | 1,695,758,205,000 |
a11f8e7c00b6c7f4c02070cb73afe5213aeb52bc | Do not pad `fp16` grouped convolutions.
In such case padding may break the invariant `input_features = feature_group_size * kernel_in_features`.
https://github.com/tensorflow/tensorflow/issues/55633
PiperOrigin-RevId: 568628624 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,759,109,000 |
ef6dd0af646037971fe0d7db8f353acaa1f66187 | [tsl] platform/BUILD: sort exports_files() and remove duplicates
PiperOrigin-RevId: 568639296 | Emilio Cota | ecg@google.com | 1,695,761,461,000 |
f7ba15ae0bb6d69d9eb415a2e9e05c244e627f5f | #tf-data Turn down `"inject_io_prefetch"` experiment.
PiperOrigin-RevId: 568639355 | Matt Callanan | mpcallanan@google.com | 1,695,761,476,000 |
2d2b6147010b130f025df772b580225b4d8a243b | Only set `_arg_keywords` when concrete function is created.
(previously, it would set the arg_keywords whenever the function is accessed)
PiperOrigin-RevId: 568644270 | Katherine Wu | kathywu@google.com | 1,695,762,518,000 |
25290c3f153bc57fbc608b3abd73569a8e37565e | [tsl] platform: use fully qualified namespaces where necessary
This paves the way for changes to the logging headers.
PiperOrigin-RevId: 568649175 | Emilio Cota | ecg@google.com | 1,695,763,610,000 |
d1ee39ffb59cf239d950f0258a184106059113f8 | patch_file moved | Raunak | mayank.kumar.raunak@intel.com | 1,695,764,878,000 |
8d745464c2ad6d9a1f0198586e07a12ef82c468f | Make DeviceExecutablePersistor safe under concurrent reads+writes.
If you have multiple XLA compilations going on simultaneously in the system
(same or different processes), you can hit a race condition where
- one thread/process is reading an entry from the compilation cache at the
same time as
- another thread... | Justin Lebar | jlebar@google.com | 1,695,765,070,000 |
781c35fbd7f7eb3eab5cf5ef7c1d9594facd9edf | tf_runtime_clangcl.patch in subfolder | mraunak | 83710963+mraunak@users.noreply.github.com | 1,695,767,066,000 |
cf37a2b610e310dc7d775b229d88fef15bf8fdc7 | Create BUILD
//third_party/tf_runtime:tf_runtime_clangcl.patch: BUILD file not found in any of the following directories. Add a BUILD file to a directory to mark it as a package | mraunak | 83710963+mraunak@users.noreply.github.com | 1,695,767,213,000 |
1d649aad42c1293751fd1ac53c138ec41b001f58 | [tflite] Add stablehlo.gather kernel tests
PiperOrigin-RevId: 568664728 | Majid Dadashi | majiddadashi@google.com | 1,695,767,136,000 |
550729cf2d4ab18901372d4eb28e196521c40efe | Keep attributes in TableGen canonicalization rewrites
By default, pattern rewrites do not preserve op attributes. This is problematic for GPU graphs, where device placement of ops are stored in "device" attribute. Losing this information causesd significant performance lost in some cases. This fixes the issues by copy... | Jian Cai | jiancai@google.com | 1,695,767,760,000 |
af1855ea4b0e7d6167940a5738ae63dd99786975 | Resolve issue in which `import tensorflow;import keras` would work but `import keras;import tensorflow` would result in TF not loading Keras. This is due to the attempt to actually import `keras` in `__init__` rather than lazy-load it.
The fix requires a change of approach and the introduction of a Keras LazyLoader cl... | Francois Chollet | fchollet@google.com | 1,695,771,794,000 |
d7e7b9d0f80ba7e4d31099325e4271355066c446 | Open source Sparse Core XLA OP kernels
PiperOrigin-RevId: 568686276 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,772,287,000 |
a891c78e49fb8aabed2f1a0fb26f9a6d7ab6aa35 | Fix missing config flags for regular builds
After my previous change removed the bazelrc/ directory from ci/official, I forgot to explicitly set
the --config options to load the settings which were previously implicit.
PiperOrigin-RevId: 568691490 | Austin Anderson | angerson@google.com | 1,695,773,634,000 |
f4265451b771fee74ad8e61c34d396050dcb5e51 | Add test case to exercise TF quantizer for Conv2d with dilation
PiperOrigin-RevId: 568691561 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,773,662,000 |
63f405d5b4a7da6eae7fa0138513387201d04277 | Validate custom options buffer before parsing into flex buffer.
PiperOrigin-RevId: 568692079 | Luke Boyer | lukeboyer@google.com | 1,695,773,786,000 |
eb6ecfb7028bb02c6f3484b6681cf76cc740ceef | [stream_executor] Remove undeclared data types in CUDA 11
https://github.com/openxla/xla/issues/5744
PiperOrigin-RevId: 568693456 | Anlun Xu | anlunx@google.com | 1,695,774,146,000 |
0397b257a9dba80a732ec8b0df604df633c7158f | This CL fixes a couple small issues:
1. Failing without a crash in SetHloShardingPostProcessing when mesh selection is on. This case was not handled in a previous CL.
2. Set the chosen mesh shape in the AutoSharding class so it can be accessed during
testing. This was done before, but looks like it got lost in a pre... | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,774,687,000 |
8d4d53f8990c6424b4ab4428bbe3aa7611d20a40 | Enable replicate_constants_pass by default (in POST_REWRITE_FOR_EXEC)
PiperOrigin-RevId: 568697316 | Edward Schwartz | schwartzedward@google.com | 1,695,775,294,000 |
f1cd9a93373d417eb011ca07ad97a87f4965ee31 | Update ctc_loss to use tensor_scatter ops instead of depreciated inplace ops
PiperOrigin-RevId: 568701158 | Edward Schwartz | schwartzedward@google.com | 1,695,776,607,000 |
86b42089c749fd4eb017a1885a7fa3125e6584bd | Legalize TensorListElementShape to tfl custom.
PiperOrigin-RevId: 568712108 | Luke Boyer | lukeboyer@google.com | 1,695,780,633,000 |
4de89a208ce11d6fb61d8badd57810afc5ce86be | Fix failing cp on Mac
PiperOrigin-RevId: 568718589 | Jake Harmon | jakeharmon@google.com | 1,695,783,121,000 |
0be57e8069699ffcfeabd526c17d2cfe9ce823c0 | [PJRT C API] Implements `CopyToMemorySpace` for `PjRtCApiBuffer`.
PiperOrigin-RevId: 568726154 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,785,478,000 |
fa2eb221fcf899b0281da2b197f29b2c3dc780fb | Allow register_revived_type to override a previous registration.
This code is called at package loading and has been causing endless bugs, preventing TF to be imported altogether whenever another package in the environment has already been calling register_revived_type.
PiperOrigin-RevId: 568733482 | Francois Chollet | fchollet@google.com | 1,695,788,606,000 |
150cc936e2dea2ead97fa88ac87975f9e6fbfa4f | Internal Code Change
PiperOrigin-RevId: 568735481 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,789,490,000 |
7e330ddc4071ad5e25746ffe53aa211845c4f44d | Remove NCCL workaround
With the upgrade to LLVM-17 in our toolchain
this workaround isn't needed anymore.
PiperOrigin-RevId: 568746859 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,793,650,000 |
18354a89e9fe8c91e2944e0e20cb0235805f36af | Update run_models.py | sushreebarsa | 84765720+sushreebarsa@users.noreply.github.com | 1,695,795,319,000 |
cc279072e16a2a51fe41f423716aeccadaeda3e6 | Update debug_events_writer_test.py | sushreebarsa | 84765720+sushreebarsa@users.noreply.github.com | 1,695,795,698,000 |
7dbcb2cd968a7b53848f9a7b1ed4f465436ddb9f | Remove PjRt C API check in `PjRtArray::Reshard` since `PjRtCApiBuffer::CopyToMemorySpace` has been supported.
PiperOrigin-RevId: 568754787 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,795,955,000 |
56f579e0438e79fc39f903a90f8a2bc55b914c70 | Internal Code Change
PiperOrigin-RevId: 568769851 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,800,843,000 |
0b86e66216e547c8dd5c825457ecbd54b40c8ffe | PR #5914: [ROCm] fixing build errors for ROCm nightly tests
Imported from GitHub PR https://github.com/openxla/xla/pull/5914
This commit fixes current build errors for ROCm platform.
@akuegel @ddunl @ezhulenev : can someone of you check this commit please ?
Here I have addressed the issues mentioned by Eugene from ... | pemeliya | 141146080+pemeliya@users.noreply.github.com | 1,695,802,839,000 |
067c77203fa942f9bb27e63f0d2cb1954552c408 | compat: Update forward compatibility horizon to 2023-09-27
PiperOrigin-RevId: 568783473 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,805,388,000 |
757c6a5750faf344e3daf2c41f1319f17971af42 | Update GraphDef version to 1632.
PiperOrigin-RevId: 568784455 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,805,616,000 |
439e612a74defe4a0d6e76886e43f8890a32a44e | Extract Scheduling and PostScheduling pipelines out of CompileModuleToLLVMIR (NFC).
This allows to test these parts separately from lowering to LLVM IR.
Adjust the .hlo tests to start with scheduled modules.
PiperOrigin-RevId: 568795746 | Adrian Kuegel | akuegel@google.com | 1,695,809,191,000 |
30c6c2227de3a53b279cfa5d189eaedd29c4aa40 | Merge pull request #61527 from ROCmSoftwarePlatform:rocmfix
PiperOrigin-RevId: 568796647 | TensorFlower Gardener | gardener@tensorflow.org | 1,695,809,972,000 |
dbba95ec7ea4ea35d891f964ae25d572d33c2543 | [XLA:GPU] NFC: improve variable names.
PiperOrigin-RevId: 568811156 | Christian Sigg | csigg@google.com | 1,695,814,290,000 |
9e524f09bb697bdf5e6cdf3f1b52841f98bc4247 | validate clip_norm argument in clip_ops.py
Earlier changes causing few test fails.Hence changed the validation code and applied casting to convert dtype to dtype = values.dtype which may avoid test fails. | Surya | 116063290+SuryanarayanaY@users.noreply.github.com | 1,695,818,623,000 |
f3c0185616d771c9c07be0d1a39c1348bdc9a5cd | Update to test | Saoirse Stewart | saoirse.stewart@arm.com | 1,695,820,236,000 |
72aa8ab84e24f66762e502ba6f39f12587e2415b | PR #5936: [ROCm] Fixing stream_executor related build errors
Imported from GitHub PR https://github.com/openxla/xla/pull/5936
@akuegel: can you please review this change?
These are ongoing changes to fix XLA tests build brakes
Copybara import of the project:
--
bf5fa597ef24340efd6d675807c79d4eacb39b0f by Pavel Emeli... | pemeliya | 141146080+pemeliya@users.noreply.github.com | 1,695,821,534,000 |
6d1cc2e35a460583354dba56d7f67f2cdf354c9e | Fixed typos and grammatical errors | Guakocius | engelhardt.alexander@mein.gmx | 1,695,822,039,000 |
9d9997b511a28797fa441e8537ab08da846329d0 | Merge branch 'tensorflow:master' into master | Guakocius | 98476579+Guakocius@users.noreply.github.com | 1,695,822,112,000 |
30ea8d52641d8e98d5daf3c75bae259d58273e20 | [Tosa] Convert MLIR shape to TF before assigning the dimensions when creating dynamic dims Reshape | Thibaut Goetghebuer-Planchon | thibaut.goetghebuer-planchon@arm.com | 1,691,597,032,000 |
44c0a47ad5ad1b8978723e8ca5944568bd46eea8 | Integrate LLVM at llvm/llvm-project@aecb58005c69
Updates LLVM usage to match
[aecb58005c69](https://github.com/llvm/llvm-project/commit/aecb58005c69)
PiperOrigin-RevId: 568858383 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,828,586,000 |
c57315017a07a1750193b8dbe510efdc75dc9a28 | Update TFRT dependency to use revision
http://github.com/tensorflow/runtime/commit/9d58a7aefd3b0b184bfc30be383bdbe17c5233ae.
PiperOrigin-RevId: 568866604 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,830,590,000 |
a4b033e9147d4b5bc2fd1691c8c94772bb4557ac | Rename `LoadSerializedExecutable` to `LoadSerialized` to avoid overriding
the method.
PiperOrigin-RevId: 568873573 | Shixin Li | shixinli@google.com | 1,695,832,169,000 |
7f3a8b5f5b8e56b5893895d1908d2a3726c60b76 | Fix build breakage introduced in CL/564547199.
That CL renamed the GetOpenGlBufferReadView method
as GetOpenGlBuffer but did not make the corresponding change
to the unit test, leading to a build error.
PiperOrigin-RevId: 568874896 | Fergus Henderson | fergus@google.com | 1,695,832,445,000 |
3e337d7165f2236487a669622e4520d847fead11 | Add a test case for fusion of convert into bitcast.
Bug found by shyshkov@google.com and kramerb@google.com
PiperOrigin-RevId: 568877930 | Johannes Reifferscheid | jreiffers@google.com | 1,695,833,121,000 |
457b3ddb9648cc2362ded1ac18cea67b96887c79 | Support partially-known shapes in `SmartBroadcastGradientArgs`.
This version is able to infer static reduction axes in cases like `[?, ?, ?] x []`, `[1, ?] x [?, 1]`, etc.
Since the logic is now fully in Python (i.e. may not trigger expensive constant folding) and relatively cheap, the cache used in the previous vers... | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,835,126,000 |
1b31f7a14d0cf6379e4a4bd6ae6e50b2c5020cf4 | Move distribute docstring to distribute/__init__.py.
PiperOrigin-RevId: 568888737 | Fiona Lang | flang@google.com | 1,695,835,170,000 |
5db0660075a0ebb499e74b9dafa4868ea3cddd8c | Add MLIR tests for tf-verify-for-export pass
PiperOrigin-RevId: 568888762 | Mason Chang | masonchang@google.com | 1,695,835,173,000 |
bff714c2ea4bd358ed866ed2c5ab04ac4772221c | Code Clean | Raunak | mayank.kumar.raunak@intel.com | 1,695,836,174,000 |
e4053ca38f72f11504f95c13f69ff9ed4cb296e5 | No public description
PiperOrigin-RevId: 568888806 | Fiona Lang | flang@google.com | 1,695,835,182,000 |
b32fa41f93f2f45ab0482d262ce45934d9102517 | Increases saltiplier to eliminate the once-in-a-blue-moon issue with unit test flakiness.
PiperOrigin-RevId: 568896545 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,836,622,000 |
8e9347ba8a8a9d37c26d386178faae18f8dc6b51 | [xla] Modify P2PSchedulePreparation to work with the
CollectivPermuteDecomposer and handle pipelined Send-Recv chains.
The P2PSchedulePreparation pass now amends control dependence to express the
intended ordering of an unpipelined Send-Recv chain, as the
CollectivePermuteDecomposer no longer generates such control de... | Bixia Zheng | bixia@google.com | 1,695,837,181,000 |
5977b36696548d78544fdd54a2e13d90c35a1c95 | Implement mocking nccl call support for collective operations
PiperOrigin-RevId: 568900214 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,837,313,000 |
4b4d32acc916311ca7300cfdff4cadf5055bbc2b | [XLA] Add option to pattern match true only scalars in offset computation.
- This option will force reduce-scatter pattern matching to only recognize
true scalar computation for DUS offset computation.
PiperOrigin-RevId: 568901880 | Zixuan Jiang | zixuanjiang@google.com | 1,695,837,646,000 |
77558760a91961ed8e01610c1dc1fc6554aab799 | format fix | Raunak | mayank.kumar.raunak@intel.com | 1,695,839,012,000 |
900e7e3814e44ac3afcd5d96df1ee12ee5f5f968 | [stream_executor] Replace stream_executor_headers with stream_executor in Tensorflow pip package
This is NFC because stream_executor transitively depends on exactly the same set of headers.
PiperOrigin-RevId: 568907149 | Eugene Zhulenev | ezhulenev@google.com | 1,695,838,620,000 |
a481ebbf2594d8cd5b103842e06b4f62cd2ae44b | Change 6/6 for making MSA repacking slice aware.
Change MSA's best fit repacker to be slicing aware.
Note, the original repacker's GetTemporalBufferIntervalCompare() method was never called. Instead, GlobalDecreasingSizeBestFitHeap::GetTemporalBufferIntervalCompare() was called. Thus:
- I deleted the repacker's origi... | Ryan M. Lefever | lefever@google.com | 1,695,840,350,000 |
6f453dccff100bdbfa7d0324d46e397f6f120994 | PR #5852: [ROCm] HipBLASLt row major layout support.
Imported from GitHub PR https://github.com/openxla/xla/pull/5852
hipBlasLt currently only supports column major. This PR enables using hipBLASLt for matrices of row major layout by swapping the operands and manipulating the transpose attributes.
Copybara import of ... | wenchenvincent | 32376000+wenchenvincent@users.noreply.github.com | 1,695,840,668,000 |
a141fa94a727396217e744fa53a8845c2fc729d9 | [TSL:CUDA] Improve cuDNN stubs: handle symbols that get versions and error strings more gracefully if the library load fails.
PiperOrigin-RevId: 568917829 | Peter Hawkins | phawkins@google.com | 1,695,840,747,000 |
cb7c35369cc53ec82763dbbe65899d76a2598375 | [PJRT C API] Clarify that PJRT_Buffer_CopyToDevice and PJRT_Buffer_CopyToMemory are within the same client.
PiperOrigin-RevId: 568921998 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,841,611,000 |
7c45b3e9692e1c92587f21ccce3274102b757e3b | There was a bug in which the MSA best fit repacker was overriding GetTemporalBufferIntervalCompare(), but that override was never being executed because it's callsite was in a constructor. A previous change, removed the override.
This change makes GetTemporalBufferIntervalCompare() non-virtual, so that the mistake is ... | Ryan M. Lefever | lefever@google.com | 1,695,843,467,000 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.