hash stringlengths 40 40 | msg stringlengths 1 131k | author stringlengths 1 33 | email stringlengths 0 57 | date int64 1,447B 1,698B |
|---|---|---|---|---|
c4e5c7c6b4227ea5aedc62ae5ef3870f04e6bd74 | Move CreateTFRegionControlFlowToFunctional to graph export as it doesn't really do any Bridge logic.
PiperOrigin-RevId: 567111139 | Mason Chang | masonchang@google.com | 1,695,251,975,000 |
aca5dfa4e11a4d18085b776ecc2bf741ee958569 | Do not create dependencies among instances of an op with TF_RandomGeneratorSideEffect trait
This makes the MLIR side effect modelling of such ops consistent with ACD.
PiperOrigin-RevId: 567115291 | Jian Cai | jiancai@google.com | 1,695,252,959,000 |
6b14d69788b63fbc9fbc6677098a96c2aca3b594 | Make `PjRtMemorySpace::id()` unique within a client
The current API semantics requires memory space ids to be unique only within a memory kind. This is inconsistent with `PjRtDevice::id()` (unique within a client regardless of device types) and cumbersome (since `(memory_kind, id)` is needed to uniquely identify a mem... | Junwhan Ahn | junwhan@google.com | 1,695,253,449,000 |
86766a45c0d3a06342e2e34b5d5332131a6390c7 | [MemorySpaceAssignment] Move code related to MemorySpaceAssignment (memory_space_assignment* files) in a separate folder.
PiperOrigin-RevId: 567120948 | Subhankar Shah | subhankarshah@google.com | 1,695,254,257,000 |
44378dfdc0b10dea9a44c86b3d7210a83c304419 | Allow `std::nullopt` in `xla::ifrt::MemoryKind` constructor.
PiperOrigin-RevId: 567123142 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,254,833,000 |
e6da72d8c1c7969e409488b843fe66cdf85e341f | 1) Fix:
* test_dense_input_ragged_weights_fails,
* test_sparse_input_ragged_weights_fails.
sparse_ops.sparse_bincount is intended to handle inputs of dense tensors (including non-TensorFlow types that can be converted to dense tensors) and sparse tensors only. Ragged tensor inputs are expected to use other software vi... | Edward Schwartz | schwartzedward@google.com | 1,695,257,675,000 |
21249b3f3a157d38b25ae954a13b0ef3057e06e6 | [stream_executor] NFC: Restrict visibility of stream_executor_internal target
https://github.com/openxla/xla/issues/5761
PiperOrigin-RevId: 567134621 | Eugene Zhulenev | ezhulenev@google.com | 1,695,258,277,000 |
cb83963df2ed9192a596c234eb4ac7f18879477b | Integrate LLVM at llvm/llvm-project@afd7db48c55c
Updates LLVM usage to match
[afd7db48c55c](https://github.com/llvm/llvm-project/commit/afd7db48c55c)
PiperOrigin-RevId: 567141554 | Fangrui Song | maskray@google.com | 1,695,260,161,000 |
463c8cd5de83752f9785f5eced74d1adaac7ddc1 | [stream_executor] NFC: Remove platform/logging.h from StreamExecutor
https://github.com/openxla/xla/issues/5761
PiperOrigin-RevId: 567148608 | Eugene Zhulenev | ezhulenev@google.com | 1,695,262,059,000 |
67b2bb20ea256517cc514b83cae36aaa79f93fc4 | Update TFRT dependency to use revision
http://github.com/tensorflow/runtime/commit/6a3e3ece9c01f3e5742a297c73357d463a2fe151.
PiperOrigin-RevId: 567151719 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,262,962,000 |
413c90e89c1e6d29eef4e4dc0e3b3fbb647addb1 | Update the logic of `PjRtArray::Reshard` after `PjRtBuffer::CopyToMemorySpace` was introduced. Users should use `PjRtBuffer::CopyToMemorySpace` instead of `PjRtBuffer::CopyToDevice` when memories are supported, since the semantics of the latter one is to always copy to the default memory space of the device.
PiperOrig... | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,263,908,000 |
84c4dd54b92e2af00b4366f87b0bad7ad78a6c9d | [stream_executor] NFC: Document exported device allocator headers
Issue: #5761
PiperOrigin-RevId: 567157790 | Eugene Zhulenev | ezhulenev@google.com | 1,695,265,078,000 |
2d4a5855f674508abf2e07828a2cd212fa94f490 | Merge pull request #61927 from ROCmSoftwarePlatform:rocm_dnn_header_dep
PiperOrigin-RevId: 567170801 | TensorFlower Gardener | gardener@tensorflow.org | 1,695,269,177,000 |
9f8e4fa28ad556e1de4b3222c4e2f58bb331f774 | Fix parsing visible_devices option in GPU plugin.
PiperOrigin-RevId: 567173464 | Jieying Luo | jieying@google.com | 1,695,269,411,000 |
b1995d4f73516a8330a0355b8b06193f44bc8e1f | [PJRT C API] Fix PjRtCApiClient::ExecutableFingerprint.
When the C API returns an empty fingerprint, PjRtCApiClient::ExecutableFingerprint should return a std::nullopt.
PiperOrigin-RevId: 567173579 | Jieying Luo | jieying@google.com | 1,695,269,448,000 |
88c8f44c2860956460a4a1722a3d2e020533a6d9 | (2/N) Refactor stablehlo bridge namespaces to mlir::quant::stablehlo.
PiperOrigin-RevId: 567187656 | Jiyoun (Jen) Ha | jiyounha@google.com | 1,695,274,249,000 |
dd689216576d6af7fb6e49453909458e5dde519f | Internal Code Change
PiperOrigin-RevId: 567191671 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,275,709,000 |
3f7e5d1b5235047337e7c43d368b89dcac8a40e1 | Merge pull request #61933 from wamuir:strip-tsl-proto-zip
PiperOrigin-RevId: 567197339 | TensorFlower Gardener | gardener@tensorflow.org | 1,695,277,820,000 |
6bb880f14f1ee4042cdc5299009b2612dbae3e4f | Merge pull request #61824 from acxz:patch-1
PiperOrigin-RevId: 567201922 | TensorFlower Gardener | gardener@tensorflow.org | 1,695,279,355,000 |
82b0522f534a7821facf0746a0fb7b3d311503cd | fixed test cases | Raunak | mayank.kumar.raunak@intel.com | 1,695,282,189,000 |
6da396b975ccf5385564cf13d13795072d0411c4 | compat: Update forward compatibility horizon to 2023-09-21
PiperOrigin-RevId: 567232445 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,286,921,000 |
7f1050a6976d11bfb0bb37bdfc82350c0a238faa | Update GraphDef version to 1626.
PiperOrigin-RevId: 567232525 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,286,939,000 |
0c7098d3552a12037d3fede9784952de732675c2 | Improve shape refinement to not require inlining.
PiperOrigin-RevId: 567273141 | George Necula | necula@google.com | 1,695,297,742,000 |
de0cbb9bb70515f75bde5ff87b7d470096e7aa49 | [XLA:GPU] Clean up Target util.
We have some differences between Triton codegen and other fusion codegen,
namely for Remainder/Fmod and Cbrt. Unify that.
- Remove two unused math functions.
- Add mapping from kRemainder to kFmod.
- Use kCbrt device function in elemental_ir_emitter.
PiperOrigin-RevId: 567274915 | Adrian Kuegel | akuegel@google.com | 1,695,298,283,000 |
e20ace297d609e5f182f51a5daf8fe1d116bfbe5 | Fix permission denied on cp of headers
Remove the duplication of copies that can result in permission
denied | Andrew Goodbody | andrew.goodbody@linaro.org | 1,695,298,889,000 |
4511d6f62cc70359482145d0feea2ae583b18b2f | [XLA:GPU] Avoid unnecessary autotuning of tiny Triton GEMM fusions.
PiperOrigin-RevId: 567293724 | Ilia Sergachev | sergachev@google.com | 1,695,304,316,000 |
d95348224dc97baddfad977ef834cc2331f17787 | Add python 3.12 to JAX docker containers
PiperOrigin-RevId: 567302141 | Yash Katariya | yashkatariya@google.com | 1,695,306,732,000 |
c2655123109b494679367639a4ba890cc9017c78 | Merge same size types for PACK to reduce binary size
PiperOrigin-RevId: 567304689 | Alan Kelly | alankelly@google.com | 1,695,307,431,000 |
4f4958cc106a8f0db94e0d2750984a641c0197a7 | Pass along the shape param from tf.compat.v1.get_variable to its underlying variable creator. At this point, we have already checked that either shape is compatible w/ initial_value, or that it is not specified (None). Therefore it should always be safe to pass it. Having it can help subsequent nested variable creators... | Neel Kovelamudi | nkovela@google.com | 1,695,310,569,000 |
060f5118478e4c0cc3fb5482470ff9c0101c752d | Improve shape refinement to not require inlining.
PiperOrigin-RevId: 567326536 | George Necula | necula@google.com | 1,695,312,790,000 |
9e03bbf6d0b9073b0d4ee66eba95bf73c3c36d73 | [XLA:GPU] Trigger Triton GEMM fusions also on kCopy input operations.
PiperOrigin-RevId: 567329486 | Ilia Sergachev | sergachev@google.com | 1,695,313,414,000 |
17d45cfb8980fc6798926b2968f57d4213c6d5c8 | Disable two compiler warnings in the bazelrc
This is adding `-Wno-error=unused-command-line-argument` and `-Wno-gnu-offsetof-extensions` to the `rbe_linux_cpu` config.
We have these flags already in the `release_cpu_linux` config,
but some RBE builds don't use that config. So I'm adding
them to `rbe_linux_cpu` as wel... | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,313,582,000 |
91f89fd2abb71f41522e47a2910b63da50651973 | #tf.data The change of default memory cap for tf.data should have been rolled out with an experiment. Revert it back and will re-roll it out with an experiment.
PiperOrigin-RevId: 567332040 | Wilsin Gosti | wilsin@google.com | 1,695,314,017,000 |
6c34dcf9aa7970fcfbba4508e9437a6f17517e24 | Legalize MHLO Pad operation to TFLite
PiperOrigin-RevId: 567337044 | Quentin Khan | qkhan@google.com | 1,695,315,063,000 |
fad9aa2ac91f765c164746263151668b8230535e | Internal change only.
PiperOrigin-RevId: 567341035 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,315,892,000 |
cdd0fa4ceca4e512ef8449bc5372383db3c59d41 | Merge pull request #61937 from elfringham:perms_fix
PiperOrigin-RevId: 567351040 | TensorFlower Gardener | gardener@tensorflow.org | 1,695,318,279,000 |
7e711e5754d30c78201885e9ae0ad2eaebb109b3 | [stream_executor] NFC: Add dnn dependency to stream executor impl
https://github.com/openxla/xla/issues/5761
PiperOrigin-RevId: 567359965 | Eugene Zhulenev | ezhulenev@google.com | 1,695,319,489,000 |
0c637e6d2eb9223355693e543c28fc4a705ce3e1 | Update release branch string for Apple Silicon MacOS builds
PiperOrigin-RevId: 567363555 | Raviteja Gorijala | gorijala@google.com | 1,695,320,174,000 |
da4b96f223fee680d305d79e0ca35c6c7810ee9c | Free GPU memory Streamz Metric.
PiperOrigin-RevId: 567363752 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,320,210,000 |
4b936ccf91eaa9ea247d3451fba26113f027fc39 | fixed build failures | Raunak | mayank.kumar.raunak@intel.com | 1,695,321,246,000 |
4280d60e8bce6bf7e20ab6bdd0170024314314d8 | Add support for the kOptimizationBarrier HLO in auto-sharding. We treat the op as an elementwise op.
PiperOrigin-RevId: 567367690 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,320,942,000 |
7cdbe3889e8fcfec86ed3378f4fc57d312a00a95 | [XLA:GPU] Move dot dimension merger before padding for cuBLAS to enable more Triton fusions.
Inside the padding pass for cuBLAS the Triton GEMM rewriter tells whether a dot should be handled by Triton, if so padding is skipped and Triton is used later. Merging dot dimensions before this makes Triton handle more of the... | Ilia Sergachev | sergachev@google.com | 1,695,321,782,000 |
dad0396ad15f178b58a50d612aec47af7ffea15f | [stream_executor] NFC: Remove unused timer target
https://github.com/openxla/xla/issues/5761
PiperOrigin-RevId: 567376822 | Eugene Zhulenev | ezhulenev@google.com | 1,695,322,704,000 |
0b836ba91ec31f7b64b22471bc0e3a03090c5ce5 | Implement Zero-Point-Offset Calculation for Lowering Dot/DotGeneral Ops in ConvertMhloQuantToInt Pass
Currently this implementation supports DotGeneral and Dot (which is a specical case of DotGeneral).
It supports static shapes only. Dynamic shapes support can be added later if needed.
PiperOrigin-RevId: 567382083 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,323,729,000 |
5192adf96920853d699b9bb3ce405438934d3191 | [XLA:GPU] Trigger Triton GEMM fusions also on kCopy input operations.
PiperOrigin-RevId: 567384050 | Ilia Sergachev | sergachev@google.com | 1,695,324,179,000 |
712acc54fb2dab395e6cd80316206d53647ae9bf | Add helper to parse element type from variant. Useful when ops don't support the helper query for retrieving underlying type.
PiperOrigin-RevId: 567390216 | Luke Boyer | lukeboyer@google.com | 1,695,325,511,000 |
b893c461189c8a44ed850c10d5ba34b6a21704c2 | [xla:gpu] Depend on stream_executor implementation in hlo_graph_dumper
PiperOrigin-RevId: 567398096 | Eugene Zhulenev | ezhulenev@google.com | 1,695,327,223,000 |
d844b5b0311a1718060ccd1c98720e189d038136 | Create an python/ops/distributions init file and delete the distributions import in the python init file.
PiperOrigin-RevId: 567400961 | Fiona Lang | flang@google.com | 1,695,327,799,000 |
307cfab7f3192db349d028b87a382a5b0328c647 | TensorListLength kernel implementation.
PiperOrigin-RevId: 567415547 | Luke Boyer | lukeboyer@google.com | 1,695,330,745,000 |
8c3316dfd30e0ce8bfe323d94d35c0b2edf60b98 | Replicate small constants so they don't need to be sent to their successors. A small constant is replicated to each of its successors' devices. The maximum size of a constant to be replicated is 16 elements.
This pass is disabled by default and can be enabled with the flag replicate_small_constants.
PiperOrigin-RevId... | Derek Murray | mrry@google.com | 1,695,331,019,000 |
44f18cf198b4315e55e9dfd84e35ee748e629286 | [XLA] Add WithSharding in pattern matcher and modify tests to conform to the new pattern matching format
-Add WithSharding implementation for HloInstructionPattern to match with the sharding.
PiperOrigin-RevId: 567419758 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,331,654,000 |
fb01efc24a76eb53e808737ee882888489bdf3f9 | [XLA] Hand-roll the evaluator's sort
This allows us to not crash if the strict weak ordering requirement is not met. Also, try to detect comparators which violate strict weak ordering are passed in.
PiperOrigin-RevId: 567421978 | David Majnemer | majnemer@google.com | 1,695,332,071,000 |
f65dd98d9b5820a19dc170f2b10ff35c463e588e | Allow additional coordinator arguments to the TPU test.
PiperOrigin-RevId: 567425791 | Yu Feng | feyu@google.com | 1,695,332,959,000 |
8bebcb801b6be745c39f6cbfee180cce5e4e9523 | Decouple traceme context id from step id in tfrt threadpool interface.
PiperOrigin-RevId: 567425831 | Kuangyuan Chen | chky@google.com | 1,695,332,968,000 |
d4100db29bd4d172555eb80a431c5504ef5e17cf | Avoid runtime failures of models | Raunak | mayank.kumar.raunak@intel.com | 1,695,333,933,000 |
c9b1800ce365003aff34665ff6564a0e5554de35 | Temporarily disabling XLA CPU oneDNN Dot op rewriter because it causes performance regression in JAX.
PiperOrigin-RevId: 567429452 | Penporn Koanantakool | penporn@google.com | 1,695,333,729,000 |
24171a1a4aa7823e1092176ecde449b65d9bc484 | [stream_executor] NFC: Prepare header filegroups for defining StreamExecutor API
PiperOrigin-RevId: 567430767 | Eugene Zhulenev | ezhulenev@google.com | 1,695,334,016,000 |
8369f9479d79be57fd87c3b1f00ea1eed00dcd29 | Integrate LLVM at llvm/llvm-project@ebefe83c092e
Updates LLVM usage to match
[ebefe83c092e](https://github.com/llvm/llvm-project/commit/ebefe83c092e)
PiperOrigin-RevId: 567450893 | Fangrui Song | maskray@google.com | 1,695,338,705,000 |
216a9cea0dff4ce4b12089eff24488babb5a3f4a | #tf-data-service Use native proto casters to simplify pybind code.
This way we don't need to define the Python version of
`DataServiceMetadata`.
PiperOrigin-RevId: 567454478 | Yang Chen | yangchen@google.com | 1,695,339,627,000 |
2247d487e8d276ca774a52d51930ed3857c433d4 | [stream_executor] NFC: Remove stream_executor_headers dependencies outside of StreamExecutor
`stream_executor_headers` is an implementation detail of StreamExecutor, all external clients should depend on a regular `stream_executor` target (unless there is a reason to depend on header only target).
And it makes absolu... | Eugene Zhulenev | ezhulenev@google.com | 1,695,340,249,000 |
34ad3d985a520364be3a3aa51d894dfe482964a0 | Update TFRT dependency to use revision
http://github.com/tensorflow/runtime/commit/73bfccd957234f8cc02c69dc50078288b6d4db8c.
PiperOrigin-RevId: 567459578 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,340,941,000 |
83b712c02c2fccff2667e706794c722edb0bce3b | Fixes to avoid deadlocks with collectives in the pipelining while loop
PiperOrigin-RevId: 567462558 | Praveen Narayanan | pravnar@google.com | 1,695,341,772,000 |
e70006f20cb98ac17876fadf3ab9aa6d17149aeb | Restore TSL/XLA headers in tensorflow/include (again)
The last attempt at this had two problems:
1) Some files were duplicated, which caused permissions errors when copied over (fixed by #61937). I still need both copies to get the right files, so added -n (no-clobber) to prevent overwriting.
2) cp -r should be cp -R ... | Jake Harmon | jakeharmon@google.com | 1,695,348,876,000 |
7e47da962783934779c7611f4566407e26929d4c | [stream_executor] NFC: Clean up kernel_spec headers and warnings
Remove unused OpenCL kernel loaders.
https://github.com/openxla/xla/issues/5761
PiperOrigin-RevId: 567504087 | Eugene Zhulenev | ezhulenev@google.com | 1,695,355,693,000 |
1518a980e1e0af880a7de819106a522c72f76036 | Internal change only.
PiperOrigin-RevId: 567508075 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,357,416,000 |
9adcd215fc5da2c8f2c2270c47bcfc373fb65e7a | Internal change only.
PiperOrigin-RevId: 567511750 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,358,653,000 |
e24946f338a92e7ee731d7fe8d765b544054821d | Internal Code Change
PiperOrigin-RevId: 567515434 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,359,796,000 |
a851496e9c8b75448cb3052352d14674ee972237 | Internal Code Change
PiperOrigin-RevId: 567516040 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,359,992,000 |
19c8aba6902f54cab505a17f76a9214bf615ab3d | Internal change only.
PiperOrigin-RevId: 567519140 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,360,986,000 |
acd860ca55cf3773ca2a1674dcfac5024a878bc0 | Add ml_dtypes headers to tensorflow/include
PiperOrigin-RevId: 567520259 | Jake Harmon | jakeharmon@google.com | 1,695,361,284,000 |
b73d62ac5c0585f3ddd34769bec51e12afb8389c | [stream_executor] NFC: Clean up xla/stream_executor:executor_cache depencencies and fix warnings
PiperOrigin-RevId: 567520429 | Eugene Zhulenev | ezhulenev@google.com | 1,695,361,331,000 |
41154a69d36384c90d75043a8052c045e037b084 | Internal Code Change
PiperOrigin-RevId: 567527024 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,362,991,000 |
2635d85c1cac2562197f5aedfeba4274c0454e9c | [XLA] Fix big memory allocation and compile time slowdown for WhileLoopInvariantCodeMotion.
The pass allocates in some cases the input for the while-cond computation, but it really only
needs one of the tuple elements of the parameter, not the whole tuple.
Allocating the whole tuple (because the inputs to the loop ca... | Marcello Maggioni | maggioni@google.com | 1,695,368,201,000 |
aeb240d9029c006ef7c66f308bee82854202e1fb | compat: Update forward compatibility horizon to 2023-09-22
PiperOrigin-RevId: 567558968 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,373,308,000 |
a5ff20d0a39f1f6638448076c4eb01613d127d73 | Update GraphDef version to 1627.
PiperOrigin-RevId: 567558973 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,373,310,000 |
bdbcdcea22474ea8278b61ecf688365634d9c9f3 | [XLA:GPU] Fix upcasting of dot operands for triton GEMM fusions
PiperOrigin-RevId: 567569309 | Tamás Danyluk | tdanyluk@google.com | 1,695,376,742,000 |
bbfa1338df681da1c76e5f65e4567242f420fd31 | fix presubmit failure | Raunak | mayank.kumar.raunak@intel.com | 1,695,377,234,000 |
8b0ec1ea93faa08eadbead5b96cc02720cb3569a | Fixed typos in TF doc
Several typos fixed in the TF documentation. | sushreebarsa | 84765720+sushreebarsa@users.noreply.github.com | 1,695,379,204,000 |
9cdc613d71a61e5b93f2690f7d522dd341c49948 | Fixing typos | sushreebarsa | 84765720+sushreebarsa@users.noreply.github.com | 1,695,379,549,000 |
91de7cef66b4bb75f33a3e79b7e9e39805216953 | Update debug_events_writer_test.py | sushreebarsa | 84765720+sushreebarsa@users.noreply.github.com | 1,695,379,731,000 |
d29f78b5748cf293c572ba9f107de8e9a0e8fdb2 | Merge pull request #61910 from fadara01:aarch64_update_to_onednn3.2.1
PiperOrigin-RevId: 567576879 | TensorFlower Gardener | gardener@tensorflow.org | 1,695,379,817,000 |
fb36fd261aeb2c7a11ebe3fe1f533685b4c030dc | Update debugger_cli_common.py | sushreebarsa | 84765720+sushreebarsa@users.noreply.github.com | 1,695,379,833,000 |
899323ee1aadbf5d2cb4cfa598eb25224862a86d | [XLA:GPU][NFC] Move out some constants in OptimizeHloPostLayoutAssignment
This is a preparation for a later CL.
PiperOrigin-RevId: 567577255 | Tamás Danyluk | tdanyluk@google.com | 1,695,379,550,000 |
c51351f28b39b3c74946127cb96d501d6d4a8169 | Integrate LLVM at llvm/llvm-project@3b0f812b9af4
Updates LLVM usage to match
[3b0f812b9af4](https://github.com/llvm/llvm-project/commit/3b0f812b9af4)
PiperOrigin-RevId: 567577565 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,379,693,000 |
89b9d8d33903f0b448f543af7cc4e6030bd4e44d | Update TFRT dependency to use revision
http://github.com/tensorflow/runtime/commit/e2e1f9f197af799cbf558eacd26221695a7971ce.
PiperOrigin-RevId: 567584745 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,382,131,000 |
422d09c0badb4bc56fb554a8a50e7e7346a2c972 | [XLA:GPU][NFC] Always profile the cuBLAS version of GEMMs in TritonAutotuner
Even if no validation is requested.
This is a preparation for a later CL.
PiperOrigin-RevId: 567587347 | Tamás Danyluk | tdanyluk@google.com | 1,695,383,143,000 |
515d795bae6b706719cc301655957c94d621d3d5 | [XLA:GPU] Re-add logging of Triton IR.
cl/560999729 replaced logging with dumps but dumping is not enabled on autotuning variants and thus does not replace logging for cases when autotuned variants crash.
PiperOrigin-RevId: 567594816 | Ilia Sergachev | sergachev@google.com | 1,695,385,883,000 |
fad56f6ce87d25e05b43bae1c32d1fff86a95dd0 | Remove clang16 toolchain and rename clang17 toolchain
The upgrade to LLVM-17 is finished, so we can remove the old clang16 based
toolchain and rename the LLVM17 toolchain.
PiperOrigin-RevId: 567598888 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,387,394,000 |
16c23ebd42c6487544881fd361fc0cd45d2abc25 | Add fusion pattern for `jax.nn.log_softmax`
PiperOrigin-RevId: 567604558 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,389,138,000 |
8c87d6780b91eeb010a611e37599ac3f01f32f08 | [XLA:GPU][NFC] Separate TritonAutotuner from other autotuning passes
This is a preparation for a later CL.
PiperOrigin-RevId: 567605682 | Tamás Danyluk | tdanyluk@google.com | 1,695,389,519,000 |
ff8fd53f5cf138c582fdfa906570cce66ba615c2 | [XLA:GPU] Fix and cover with tests type conversions in Triton emitters.
PiperOrigin-RevId: 567619889 | Ilia Sergachev | sergachev@google.com | 1,695,393,958,000 |
be8bc81b566e89b6f9958d2ca0d84bb250ebcfeb | Added a few missing members to jax_jit.JitState
PiperOrigin-RevId: 567621107 | Sergei Lebedev | slebedev@google.com | 1,695,394,317,000 |
b698c540108c00752a4668765bb5b2af762519cc | Integrate LLVM at llvm/llvm-project@058222b23166
Updates LLVM usage to match
[058222b23166](https://github.com/llvm/llvm-project/commit/058222b23166)
PiperOrigin-RevId: 567628854 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,396,492,000 |
655db6e4db11c3bc310e46145d7bb2292ea3063c | Clarify that PjRtClient and PjRtDevice memory_spaces are not in particular order
PiperOrigin-RevId: 567630629 | Hyeontaek Lim | hyeontaek@google.com | 1,695,396,992,000 |
5b7a8b07e6ead1db358fa0a5cbff930a83bbedfd | Generate CUDA stubs using implib.so, rather than by writing C++ stubs.
We load CUDA libraries lazily using dlopen()/dlsym() primarily to comply with the manylinux rules for Python wheels, which require that libraries in a wheel only link directly against an allowlist of libraries. In order to access CUDA using dlopen(... | Peter Hawkins | phawkins@google.com | 1,695,398,476,000 |
04c49a0785d25d0f1878605a719748d6e22d0bcc | Cache buffer sizes in `BufferAssignment` when they are accessed.
In some cases, the buffer-size computation involves a non-trivial amount of work, and executing it multiple times per value can be very expensive.
PiperOrigin-RevId: 567636486 | Derek Murray | mrry@google.com | 1,695,398,604,000 |
6b81dd15bd52110374f287b9d54ab7660fb0ce20 | Update TFRT dependency to use revision
http://github.com/tensorflow/runtime/commit/3b7012849fde349a679c495f5bd421935d6599f9.
PiperOrigin-RevId: 567638210 | A. Unique TensorFlower | gardener@tensorflow.org | 1,695,399,054,000 |
f697bba3b44fbb5230714acaf2a8412bcae4ec2a | Remove allow_multiple_exports param from tf_export.
PiperOrigin-RevId: 567651676 | Ramesh Sampath | rameshsampath@google.com | 1,695,402,385,000 |
6cfebd9843e5944f49eb937ead055bec0708c080 | Improve shardings if the users of a const broadcast have conflict shardings.
PiperOrigin-RevId: 567655772 | Shibo Wang | shibow@google.com | 1,695,403,296,000 |
ac8da081f7c9af3ce661a4e84dc699762d6830f7 | Internal Code Change
PiperOrigin-RevId: 567656609 | Gabriel Rasskin | grasskin@google.com | 1,695,403,477,000 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.