hash stringlengths 40 40 | msg stringlengths 1 131k | author stringlengths 1 33 | email stringlengths 0 57 | date int64 1,447B 1,698B |
|---|---|---|---|---|
3711ad6481f661748ab9b1236ba2f60edcc79714 | Updating CONTRIBUTING.md. changing links to tensorflow devoloper docker file | ArkadebMisra | misra.arkadeb@gmail.com | 1,696,864,231,000 |
9101aa36a498f39f179e640f1201a2a37691e19d | Add int4 support to IrArray.
In LLVM, an int4 IrArray is represented as an array of i8 values, where each i8 value stores two packed 4-bit integers. The reason an array of i4 values is not used is because i4 values are not packed in LLVM IR (the size of any type is a multiple of 1 byte).
Also add tests that check the... | Reed Wanderman-Milne | reedwm@google.com | 1,696,916,485,000 |
d5e2784cb89f83c8ac06c03bdf41ee741b0f4e00 | Consider reverse device mesh permutations as well when generating strategies for reduce ops based on the strategies of its operand as the operand could be the result of a reverse HLO op.
PiperOrigin-RevId: 572139124 | A. Unique TensorFlower | gardener@tensorflow.org | 1,696,916,550,000 |
115b2dbfbe97457c116ecd188972e00e484a1d00 | Internal Code Change
PiperOrigin-RevId: 572144502 | A. Unique TensorFlower | gardener@tensorflow.org | 1,696,918,426,000 |
55d9b5cbab0c6384b1434067b6a7d311b9a46a7d | [gpu] NFC: Do not use StreamExecutor internal API to access underlying Gpu stream
PiperOrigin-RevId: 572155889 | Eugene Zhulenev | ezhulenev@google.com | 1,696,922,012,000 |
298a6ddacc5c9548bb33554f4a588976fbade29c | Integrate StableHLO at openxla/stablehlo@78f57e34
PiperOrigin-RevId: 572161714 | Gunhyun Park | gunhyun@google.com | 1,696,923,992,000 |
3effef22b1c4a965548a8ccec611c0e2d59c2eba | [gpu] Do not use StreamExecutor internal API to access underlying Gpu stream
PiperOrigin-RevId: 572167716 | Eugene Zhulenev | ezhulenev@google.com | 1,696,925,975,000 |
c66f8f6a0fa5a8d03379979366f71c29607966df | [TSL] Remove `TSL_STATIC_THREAD_LOCAL_POD` macro.
PiperOrigin-RevId: 572168960 | Christian Sigg | csigg@google.com | 1,696,926,381,000 |
6e688d2537e75fe40a94d195d34591086a9bffd2 | Update GraphDef version to 1645.
PiperOrigin-RevId: 572175807 | A. Unique TensorFlower | gardener@tensorflow.org | 1,696,928,525,000 |
8070a54e566bdc9cba94bdb85e6074b4a49dfc36 | compat: Update forward compatibility horizon to 2023-10-10
PiperOrigin-RevId: 572175821 | A. Unique TensorFlower | gardener@tensorflow.org | 1,696,928,528,000 |
049dc105e36339f451472c31fb7f62b9ad6e220b | Internal Code Change
PiperOrigin-RevId: 572177328 | A. Unique TensorFlower | gardener@tensorflow.org | 1,696,928,946,000 |
e713a8a89da4c078bd0c6a24aa62ea1f983cccab | Integrate LLVM at llvm/llvm-project@2e8269690909
Updates LLVM usage to match
[2e8269690909](https://github.com/llvm/llvm-project/commit/2e8269690909)
PiperOrigin-RevId: 572181129 | A. Unique TensorFlower | gardener@tensorflow.org | 1,696,930,065,000 |
389ec92505dd25c1a9ed553766e54a5ec2e8cbf4 | [stream_executor] NFC: Remove redundant overrides
PiperOrigin-RevId: 572181451 | Eugene Zhulenev | ezhulenev@google.com | 1,696,930,167,000 |
325aed2106d41100059f0c8596fd4f6b383ae57c | [XLA/GPU] Create empty GpuCompilationEnvironment.
- Add helper functions to use this proto as part of xla::CompilationEnvironmentsProto.
PiperOrigin-RevId: 572199997 | A. Unique TensorFlower | gardener@tensorflow.org | 1,696,935,792,000 |
d93de85aa7b08f7783d90afae98f78b13f2175e0 | [XLA:GPU] NFC: Cost model variable type and name cleanup.
PiperOrigin-RevId: 572214509 | Christian Sigg | csigg@google.com | 1,696,940,581,000 |
19c648a1ccf0887c91a58b5d4ba09a1372b7cc10 | [XLA:GPU] Set name for priority fusion pass.
PiperOrigin-RevId: 572224618 | Oleg Shyshkov | shyshkov@google.com | 1,696,943,444,000 |
1f4e6e7a011b926527ce28e1f6946161e97fdea3 | PR #6171: [NVIDIA XLA GPU] add helper to hlo computation to identify to_apply regions
Imported from GitHub PR https://github.com/openxla/xla/pull/6171
add helper to hlo computation to identify to_apply regions of collectives
Copybara import of the project:
--
c005b83a27ff42dd0701a1c0898784a24d88f333 by TJ <tjx@nvidi... | TJ Xu | tjx@nvidia.com | 1,696,943,826,000 |
03d8df0f8d9c51cacea632053ce2c387c41d5aaa | Generalize the logic of following values up and down through nested blocks.
PiperOrigin-RevId: 572240865 | Matthias Kramm | kramm@google.com | 1,696,947,813,000 |
6044fe278e1f440a995fc92412627be1e12d14c8 | [XLA:GPU] Move TF_RET_CHECK to TSL
PiperOrigin-RevId: 572245598 | George Karpenkov | cheshire@google.com | 1,696,948,954,000 |
66e87d589edfd18021b104d71af1bfcc9b17dfbc | [JAX] Remove old RBE images formerly used by JAX.
PiperOrigin-RevId: 572246411 | Peter Hawkins | phawkins@google.com | 1,696,949,148,000 |
761827ba101878316c294c5e17e88c2c8c46594b | Re-enable layering_check for uniform_quant_ops package.
PiperOrigin-RevId: 572250814 | A. Unique TensorFlower | gardener@tensorflow.org | 1,696,950,123,000 |
cd8ec9b28db7deef0becb603de9b4f5a28edb855 | Add flag protected dynamically quantized Fully Connected operator to XNNPACK delegate
PiperOrigin-RevId: 572253490 | Alan Kelly | alankelly@google.com | 1,696,950,673,000 |
78345ba4a706c9bc96368e57006f0d5fac478773 | Combine all-gathers with different dimensions
Combine all-gathers as a tuple-based representation when the gather dimensions
differ. Insert bitcasts to align the operands' gather dimensions. This is
supported by NCCL as we always gather the major-most layout dimension. We want
to use this for GPU.
PiperOrigin-RevId: ... | Frederik Gossen | frgossen@google.com | 1,696,953,992,000 |
22141485c2fa68d01a2ce3b567fe874346ffdc54 | Redirect references from the framework target to the new single-source-file targets.
PiperOrigin-RevId: 572278124 | Juan Martinez Castellanos | juanantoniomc@google.com | 1,696,956,064,000 |
f80d8e3941cc880013e0141168669d04a0429635 | Remove take_ownership support in DLPack.
When take_ownership is true, the original buffer is marked as deleted and enforced that JAX won't attempt to read or write the buffer. This provides better error checking but at the cost of one more C++ API and two more C APIs. The same semantic can be achieved by not using tak... | Jieying Luo | jieying@google.com | 1,696,956,136,000 |
ff19a2fb08f64be66dd49fab33d023d5ac466cb0 | [xla:gpu] Support multiple fusion instructions in command buffer scheduling
#5756
PiperOrigin-RevId: 572281792 | Anlun Xu | anlunx@google.com | 1,696,956,794,000 |
c4af513208400b159a390a587a5f8427a91def30 | Remove AddGraphExportLoweringPasses as users should migrate to tf_to_dialect_executor
PiperOrigin-RevId: 572285093 | Mason Chang | masonchang@google.com | 1,696,957,352,000 |
bfc08694e04476ffa2b36e72decb85f9306b2a5e | Re-enable layering check for package.
PiperOrigin-RevId: 572288230 | A. Unique TensorFlower | gardener@tensorflow.org | 1,696,957,852,000 |
c54a7953a97ea11c08e01071fefa0e1cf65de412 | Pass bytes limit in allocator to compile options.
PiperOrigin-RevId: 572288596 | Tao Wang | wangtao@google.com | 1,696,957,912,000 |
0bd26e7c3709a92d3b8ed6448692e508943d1e5e | Import keras directly in distributed_table_test.
PiperOrigin-RevId: 572298063 | Fiona Lang | flang@google.com | 1,696,959,346,000 |
afd5e797622575c60faa85679e3a3b2299035898 | [XLA:GPU] Shard topk_kernel.cu.cc.
I've seen this file take over 5 minutes to build. Shard it by type.
PiperOrigin-RevId: 572298298 | Peter Hawkins | phawkins@google.com | 1,696,959,390,000 |
36f2c87e0892e3dfc56d5fcb868cc7765ebc910e | Log the compilation results on success
PiperOrigin-RevId: 572299330 | A. Unique TensorFlower | gardener@tensorflow.org | 1,696,959,586,000 |
ae9030d10b7b63207d9a6d7adfa071983122c674 | Only call Eigen::div_ceil / divup with integer types. A future update of Eigen will enforce this with a static assert.
PiperOrigin-RevId: 572303623 | A. Unique TensorFlower | gardener@tensorflow.org | 1,696,960,288,000 |
5fdf7027057299f5c34eb7ccba2feaf72836a9c3 | Remove calling configure_library_path during jax import and get libtpu path from libtpu_module.get_library_path().
PiperOrigin-RevId: 572306461 | Jieying Luo | jieying@google.com | 1,696,960,736,000 |
d8f781766c347fdcb30daab206f5de12f0e65f5e | Fix host stall calculation for megascale stats.
PiperOrigin-RevId: 572311423 | Clive Verghese | cliveverghese@google.com | 1,696,961,639,000 |
746aa247987ae2139c7dfc898af092aff2d99dab | Implement Freeze() in pathways/tfrt serving.
PiperOrigin-RevId: 572312322 | Kuangyuan Chen | chky@google.com | 1,696,961,814,000 |
29dc2a31b9b35688d3d21fcad08d5f89db460819 | [xla:pjrt] Do not use StreamExecutor internal API to access underlying Gpu stream
PiperOrigin-RevId: 572328836 | Eugene Zhulenev | ezhulenev@google.com | 1,696,965,023,000 |
e3ea4bb9806b952ea60ee245e44bc72411f05c06 | [XLA:GPU] Move TF_RET_CHECK to TSL
PiperOrigin-RevId: 572340491 | Robert David | lrdx@google.com | 1,696,967,670,000 |
bcc3511e0585b2ea80e1909193ddf8a7f7823a8e | [PJRT C API] Add a changelog for PJRT C API.
PiperOrigin-RevId: 572351942 | Jieying Luo | jieying@google.com | 1,696,970,135,000 |
3f881db832520a0b903063e10f4ade32191023f3 | Integrate LLVM at llvm/llvm-project@49cb1595c1b3
Updates LLVM usage to match
[49cb1595c1b3](https://github.com/llvm/llvm-project/commit/49cb1595c1b3)
PiperOrigin-RevId: 572354636 | A. Unique TensorFlower | gardener@tensorflow.org | 1,696,970,705,000 |
b13ee9352eb7c0568661b14a27780bf44a39c5de | Move EmbeddingProgramKeyPass after ReplicateInvariantOpHoistingPass.
PiperOrigin-RevId: 572357064 | Matthias Kramm | kramm@google.com | 1,696,971,202,000 |
31fc45d77d99161403ebe4cb7ca31790117a5918 | #tf-data-service Split provider restores repetition index correctly.
If the dispatcher has not written the next split file, the restored
repetition index is always 0.
PiperOrigin-RevId: 572362508 | Yang Chen | yangchen@google.com | 1,696,972,207,000 |
f841af26fad9a37a8a24a4d0b1a16a3084572b64 | Merge branch 'master' into gaurides/placeholder2constant | Gauri1 Deshpande | gauri1.deshpande@intel.com | 1,696,974,139,000 |
7506bd2d43a0cd21a57e72812f67bc0021e7d68b | PR #5079: [NVIDIA XLA GPU] Introduce an analytical latency estimator for LHS.
Imported from GitHub PR https://github.com/openxla/xla/pull/5079
This uses GPUPerformanceModel to estimate cost for compute kernels adds mechanism to compute collective cost using GpuHloAnalysis This initial version deals with single-node d... | TJ Xu | tjx@nvidia.com | 1,696,979,353,000 |
83c43b0716c3809f8c460406ac94db1458f9ba36 | Preserves the operand index of an instruction and uses this to retrieve the appropriate default strategy from sharding propagation.
PiperOrigin-RevId: 572395302 | A. Unique TensorFlower | gardener@tensorflow.org | 1,696,979,545,000 |
6729179be91aebfd887585f3b6a95a024104cbac | Fix format of changelog.
PiperOrigin-RevId: 572396142 | Jieying Luo | jieying@google.com | 1,696,979,757,000 |
99d017cd6e57b5ff2f8428d35f6b86176cc47055 | Support INT64 indices in C++ GatherNdGrad
PiperOrigin-RevId: 572397938 | A. Unique TensorFlower | gardener@tensorflow.org | 1,696,980,199,000 |
b5f1ce7b83b5157d01a0f33073fae820ba9c42e0 | [PJRT C API] Add a markdown version of the PJRT integration guide.
PiperOrigin-RevId: 572410177 | Jieying Luo | jieying@google.com | 1,696,983,173,000 |
95e508e6d795cb734506914936363e91ab7c85c3 | Add TPU to CPU passes to QAT pipeline
PiperOrigin-RevId: 572417036 | Thai Nguyen | thaink@google.com | 1,696,985,148,000 |
0b18e1e48150cf7b92b87ffcd6e65b2def923652 | #tf-data-service Add a test for sources with different repetitions.
PiperOrigin-RevId: 572418080 | Yang Chen | yangchen@google.com | 1,696,985,510,000 |
a6a725d0f4ce9b0185badfab98dd0d5f583244e7 | Removes unnecessary semicolons as these generate errors when
-Wextra-semi is used.
PiperOrigin-RevId: 572419061 | A. Unique TensorFlower | gardener@tensorflow.org | 1,696,985,839,000 |
c40e4e1fd98c744f43ee55e168b057ab1c5962e7 | Fix dialect_test for upstream 479057887fbc8bfef17c86694f78496c54550f21
This test previously got the IRUnit overload. But with the new overload
set it is getting the Operation overload, which is much more limited.
Change it to use the previously-used Region overload.
PiperOrigin-RevId: 572419259 | A. Unique TensorFlower | gardener@tensorflow.org | 1,696,985,900,000 |
b5a8736d62da4f69507abbc663bd07855a0786df | Fix failed Waymo test.
PiperOrigin-RevId: 572431104 | A. Unique TensorFlower | gardener@tensorflow.org | 1,696,990,199,000 |
15caee50bad207ae04513ac1bef48ca7ea422ccb | #tf-data-service Use absl::Status in snapshot_split_provider.
PiperOrigin-RevId: 572434287 | Yang Chen | yangchen@google.com | 1,696,991,514,000 |
e90e58c4451a9665389731fb6f3673dad0482c32 | Preserves the tuple shape index of an instruction and uses this to retrieve the appropriate default strategy from sharding propagation.
PiperOrigin-RevId: 572441800 | A. Unique TensorFlower | gardener@tensorflow.org | 1,696,993,981,000 |
ecd2f1a9ffec1eed097417e6fc3dd44eabe3b28c | [XLA:Runtime] Moved the cholesky thunk to a new folder, as part of a thunk clean up, and updated the necessary directories pointing to this thunk. #5758
PiperOrigin-RevId: 572446850 | A. Unique TensorFlower | gardener@tensorflow.org | 1,696,995,921,000 |
fd98e6b49295b2cd0792a1cda661a7041941b984 | Fixes build issue for Windows with Google build system.
PiperOrigin-RevId: 572451662 | Jake Harmon | jakeharmon@google.com | 1,696,997,817,000 |
410aac901358b3f74c86e321521e2f2a469ccf3d | Internal Code Change
PiperOrigin-RevId: 572458740 | A. Unique TensorFlower | gardener@tensorflow.org | 1,697,000,674,000 |
39f5c002eb70b7e16793015d2710b6691cefb217 | Only ignore iota when the comparator allows for it.
Untangling or simplifying comparators is fairly complex. For now,just fallback
to the default path.
PiperOrigin-RevId: 572465711 | A. Unique TensorFlower | gardener@tensorflow.org | 1,697,003,011,000 |
1bd1b7f4b0311575ead19f14e74488a97b524555 | PR #6199: [ROCm] Hip graph exec kernel node set params
Imported from GitHub PR https://github.com/openxla/xla/pull/6199
Fixed ROCm build error due to https://github.com/openxla/xla/commit/4f94b3290f8b69e35930410bcc77674c1d10afbe
@akuegel @ezhulenev Thanks in advance!
Copybara import of the project:
--
9ea7cbda574... | Chao | cchen104@amd.com | 1,697,003,367,000 |
a3563726c277ea032f4f2c291e947887d75ef8a5 | Merge pull request #62074 from ArkadebMisra:feature-update-contributing-md
PiperOrigin-RevId: 572475692 | TensorFlower Gardener | gardener@tensorflow.org | 1,697,006,498,000 |
35a367283020d90d7f8e80776dbd95e980ede2e4 | [XLA] Fix XlaZeta Euler-Maclaurin formula.
This change fixes the imprecision in XlaZeta function, by correcting the terms in Euler-Maclauring formula.
See https://arxiv.org/pdf/1309.2877.pdf (5th formula) for reference. The corrected term is "T".
PiperOrigin-RevId: 572504242 | A. Unique TensorFlower | gardener@tensorflow.org | 1,697,014,269,000 |
e2d54885ea50f7b3926da9bf10f28d84b9eb5f5d | [XLA:GPU] Don't verify Triton and GPU Runtime LLVM modules by default
Please set --xla_gpu_llvm_verification_level=1 to verify.
The verification was slow for big HLOs, taking up about 50% of the runtime of TritonWrapper.
The verification always happens in debug mode (when NDEBUG is not set).
This is not autotuning-s... | Tamás Danyluk | tdanyluk@google.com | 1,697,014,936,000 |
f202d6d46877c4831fe64961936155b5cf543a5f | compat: Update forward compatibility horizon to 2023-10-11
PiperOrigin-RevId: 572506757 | A. Unique TensorFlower | gardener@tensorflow.org | 1,697,014,978,000 |
7b7e067474d0b7998217c30f87f81ac8dc6c884b | Update GraphDef version to 1646.
PiperOrigin-RevId: 572506848 | A. Unique TensorFlower | gardener@tensorflow.org | 1,697,014,992,000 |
9f9c8010f900bd70a5bc936ba19e9b10e9da36b9 | [XLA] Rewrite XlaZeta op for readability.
In this change additional variables are present (as constexpr), and some are renamed to match the whitepaper.
The biggest change is the decomposition of the final result for readability (result = S + I + T).
PiperOrigin-RevId: 572514262 | A. Unique TensorFlower | gardener@tensorflow.org | 1,697,017,214,000 |
0aecf379b7bbdbe93be91643825f0ae94171d509 | Fix StatusOr invariant violation.
Straight-line code is accessing a StatusOr value after checking that the status is non-OK.
Includes incidental changes to code formatting requested by clang-format.
PiperOrigin-RevId: 572515414 | A. Unique TensorFlower | gardener@tensorflow.org | 1,697,017,542,000 |
c3af75d1e63ddfe5b9b0fd355cf7185d06b82b89 | [XLA:GPU] Clean up and sort dependencies of gpu_compiler.cc
PiperOrigin-RevId: 572539272 | Tamás Danyluk | tdanyluk@google.com | 1,697,025,294,000 |
e138f54c85a986442009ef24b84f2de067b149ee | Integrate LLVM at llvm/llvm-project@ebea9308379d
Updates LLVM usage to match
[ebea9308379d](https://github.com/llvm/llvm-project/commit/ebea9308379d)
PiperOrigin-RevId: 572542275 | A. Unique TensorFlower | gardener@tensorflow.org | 1,697,026,358,000 |
9d7bdcd21d01db72249670cb67182525762468d8 | Update clip_ops.py
change tf.cast to math_ops.cast | Surya | 116063290+SuryanarayanaY@users.noreply.github.com | 1,697,027,774,000 |
8a5d02a320d042490a3e2d471aa1e5ef3d5e0d79 | Fix null pointer dereference in DirectSession::RunInternal | Darya Parygina | Parygina.Darya.2001@yandex.ru | 1,680,618,091,000 |
eb8b34be5122311e397f821b9bc6b5b55b181725 | Change tsl::errors::Internal to absl::InternalError | Darya Parygina | pa_darochek@ispras.ru | 1,686,660,343,000 |
8a30d0ae844142624135a3d8c77818f27c32bb29 | Add guards on each subsequent use of run_metadata | Darya Parygina | pa_darochek@ispras.ru | 1,697,031,070,000 |
cab8e07ef16cbe32fdf4e98e036caae4e90700ba | Change the default python used for commands in arm64 experimental container to 3.10.
Modify new scripts to support aarch64 auditwheel verification
Create experimental arm64 nightly envs for python 3.9 and 3.11
PiperOrigin-RevId: 572560764 | Michael Hudgins | michaelhudgins@google.com | 1,697,032,162,000 |
eddae73fa63e53f2ebb17f5c30ff5bc06ca060e1 | Register all StableHLO quantization passes to stablehlo-quant-opt.
This change registers all StableHLO quantization passes to `stablehlo-quant-opt` by invoking the registration functions generated by `GEN_PASS_REGISTRATION`.
PiperOrigin-RevId: 572567939 | Dan Suh | dansuh@google.com | 1,697,034,185,000 |
95585e9d25942d4015ec04f7eced82c40536a1b7 | HloFusionAnalysis: Don't peek outside the fusion boundary.
Currently, we use the FindNonTrivialHero overload that doesn't take a boundary
function. This isn't always correct.
PiperOrigin-RevId: 572576780 | Johannes Reifferscheid | jreiffers@google.com | 1,697,036,574,000 |
4963eac6fbefcf68c8b9f1aaa0d53f3430e93ddf | Add flag protected Batch Matrix Multiply operator to XNNPACK delegate
PiperOrigin-RevId: 572580753 | Alan Kelly | alankelly@google.com | 1,697,037,517,000 |
b3acbcd0da289e04bdc9d24cb731ef698d146ec7 | Extract TPU Bridge out to Bridge API v/2 Cluster TF
PiperOrigin-RevId: 572589861 | Mason Chang | masonchang@google.com | 1,697,039,810,000 |
790d7846f8bc9a2f2e5f6be2914e927b09e7c1ff | [XLA:SPMD] Improve tupled shape manual sharding propagation.
1. Allow default forward manual sharding propagation for tuple-shaped custom-call.
2. Restrict tuple-shaped manual sharding propagation to parallel ops only in backward sharding propagation, for example:
(s32[], s32[]) %x = some-op(...)
s32[] %y = get-... | Tongfei Guo | tongfei@google.com | 1,697,041,507,000 |
3570ca735ff4f6e64ea7ac0068e0604c4438c024 | Add HLO Verifier support to hlo-expand tool
PiperOrigin-RevId: 572602542 | Kevin Gleason | gleasonk@google.com | 1,697,042,679,000 |
358a7a4bd8127e61c1b664e7c903df772195f23b | Split client target into single-source targets.
PiperOrigin-RevId: 572611109 | Juan Martinez Castellanos | juanantoniomc@google.com | 1,697,044,424,000 |
882094101f80eee20206f9cc08be5731b24ef778 | Pass profiler option to constructor of PluginTracer.
PiperOrigin-RevId: 572613934 | Jieying Luo | jieying@google.com | 1,697,045,067,000 |
fe323cf13eedd34c6650954cf1acf56a8eebcf5d | AOT compile a TF function to PjRtExecutable.
PiperOrigin-RevId: 572616383 | Shixin Li | shixinli@google.com | 1,697,045,563,000 |
486f4db7eb4cf142780e148cabff449bca33202c | minor change | Gauri1 Deshpande | gauri1.deshpande@intel.com | 1,697,047,802,000 |
4258f33e80631959df0b04562d9f25a9acd75c41 | Add tf_export decorator for API generation for TPUEmbeddingV2
PiperOrigin-RevId: 572632286 | Hye Soo Yang | hyey@google.com | 1,697,048,460,000 |
a2f2d3f34a5df9d502419be1c645bd8250c4a25b | Merge pull request #61242 from tensorflow:SuryanarayanaY-patch-3
PiperOrigin-RevId: 572635308 | TensorFlower Gardener | gardener@tensorflow.org | 1,697,049,532,000 |
fe31d9dec360b2f4da4bfa04649d03df7b265444 | Mark some bridge entry points as deprecated
PiperOrigin-RevId: 572637424 | Mason Chang | masonchang@google.com | 1,697,049,324,000 |
a4c0d4ce6eb7159ab8fe5f1611f891eb3c61e947 | Combine reduce-scatters with different dimensions
Combine reduce-scatters as a tuple-based representation when their dimensions
differ. Insert bitcasts to align the operands' dimensions. This is
supported by NCCL as we always gather the major-most layout dimension. We want
to use this for GPU.
PiperOrigin-RevId: 5726... | Frederik Gossen | frgossen@google.com | 1,697,049,341,000 |
49fe0c8a6ea4237458fe63223f0be7894aeec515 | Preserves tuple element indices in a few extra places.
PiperOrigin-RevId: 572643045 | A. Unique TensorFlower | gardener@tensorflow.org | 1,697,050,375,000 |
bdf4fca6c1236daee6f55cbf86256d26108edac7 | Lift done callback out of cuda context scope
This change fortifies the codebase against random deadlocks.
The TF Executor may inline another kernel in the successful done callback,
which expects the original cuda context and hangs if the finishing kernel does not
restore the context.
Note that some of the use of the ... | Yu Feng | feyu@google.com | 1,697,051,749,000 |
1fdb8e028e46d68a02ba1f90a0230a1ee67e248b | Delete xla/rpc/* as it has no users
PiperOrigin-RevId: 572650647 | David Dunleavy | ddunleavy@google.com | 1,697,051,953,000 |
bbe424eb8a280c7730c2c8816b77da6e40e5972e | Use library detailed logging rather than copy/pasted version
PiperOrigin-RevId: 572653141 | Mason Chang | masonchang@google.com | 1,697,052,560,000 |
15232520bba5793ed15e174a737c0d9296dbec30 | Add cuda-nvml-dev-11-8 to Dockerfile
PiperOrigin-RevId: 572654434 | David Dunleavy | ddunleavy@google.com | 1,697,052,847,000 |
683c54c1d8192590f0a93e9be1afe6db9edd4e9f | Integrate FDO into the new host preprocessing ops
PiperOrigin-RevId: 572662088 | Pat Notz | patn@google.com | 1,697,054,505,000 |
bdb14431fb8623df8301db9a5b70e933d487f748 | [xla:gpu] Revert the default gpu_graph_level back to 1
PiperOrigin-RevId: 572662254 | Anlun Xu | anlunx@google.com | 1,697,054,527,000 |
bd7efa358aa86de30372238d8c345492acddbaaa | Re-enable layering checks for package.
PiperOrigin-RevId: 572672005 | A. Unique TensorFlower | gardener@tensorflow.org | 1,697,056,515,000 |
602f7c18bc5c0f5368922872eaed03cb72bfbf18 | Initialize the output tensor to zeros to avoid data leaking when input indices do not cover all portions.
PiperOrigin-RevId: 572679649 | Nan Wei | nanwei@google.com | 1,697,058,088,000 |
30133b25a5fc4db255fd351ce0faea160ca7188d | Updated oneDNN version to v3.3 | Bhavani Subramanian | bhavani1.subramanian@intel.com | 1,697,052,961,000 |
e0779fff8341d8e59c626eb296d37207acf76516 | Remove unnecessary source files from the deprecation_test target.
PiperOrigin-RevId: 572681740 | Juan Martinez Castellanos | juanantoniomc@google.com | 1,697,058,491,000 |
f7d817d8075bc25fc1413192a11a63b4ef4d0ebb | PR #6126: [NVIDIA XLA GPU] Fix allreduce/reducescatter combiner bug
Imported from GitHub PR https://github.com/openxla/xla/pull/6126
AR/RS combiners are run after fusion and float normalization passes. This will make them fail to combine anything because:
1. float normalization will promote to_apply regions of ARs an... | TJ Xu | tjx@nvidia.com | 1,697,060,117,000 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.