hash stringlengths 40 40 | msg stringlengths 1 131k | author stringlengths 1 33 | email stringlengths 0 57 | date int64 1,447B 1,698B |
|---|---|---|---|---|
a9d44e0001d6feba6b21c14a9fc670864c50d89f | Fix GitHub export for quantized variable ops tests
PiperOrigin-RevId: 558283652 | Marat Dukhan | maratek@google.com | 1,692,403,159,000 |
36f5660069fa4535c3ee5db4c93abfeb5927b2a8 | [xla:gpu] NFC: Better error message for GPU graph update
PiperOrigin-RevId: 558291877 | Anlun Xu | anlunx@google.com | 1,692,405,901,000 |
9db029dfd4d0def37c2297aca057c40a0ab5dcac | Add a transform for cross-replica-sum to all_reduce
At the moment, StableHLO does not have the spec and the interpreter for `cross-replica-sum`. `cross-replica-sum` ops are are essentially `all_reduce` ops with a summation computation, and this pass does a `cross-replica-sum -> all_reduce` conversion.
PiperOrigin-Rev... | Gunhyun Park | gunhyun@google.com | 1,692,407,173,000 |
c9fafed9bc8cb0238a775fd4a0680e648c06b5b6 | Copy tensor content through the API
PiperOrigin-RevId: 558308631 | Haibo Huang | hhb@google.com | 1,692,413,475,000 |
0f65075055bb7d226c80a09942fb72681a21873a | compat: Update forward compatibility horizon to 2023-08-19
PiperOrigin-RevId: 558366456 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,435,835,000 |
176ffcef28e9199cac244b25eaa78b24e291fd8c | Update GraphDef version to 1593.
PiperOrigin-RevId: 558366460 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,435,837,000 |
6a79834856576fbcbe3323b8e3bd8efb17394a95 | Update GraphDef version to 1594.
PiperOrigin-RevId: 558532383 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,522,180,000 |
2bc270a640c17fffb865b225c9fc5d69e80c1dc9 | compat: Update forward compatibility horizon to 2023-08-20
PiperOrigin-RevId: 558532396 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,522,183,000 |
c054efe23a6f7fbac401b4ebf9b6353ecc18e0a0 | Add `#else` to fix broken test.
PiperOrigin-RevId: 558671703 | Jieying Luo | jieying@google.com | 1,692,593,964,000 |
6552f8ab3a2f5139c26f885ce366240df6a5a940 | Merge pull request #61619 from angerson:master
PiperOrigin-RevId: 558691018 | TensorFlower Gardener | gardener@tensorflow.org | 1,692,600,444,000 |
732f44799e3297dcff4c3c3cc6652b7fbfcf75c0 | Internal Code Change
PiperOrigin-RevId: 558691301 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,600,238,000 |
a0b91ddf11027cc09007539f70bcaa62cfc8d4b8 | Merge pull request #61638 from ROCmSoftwarePlatform:fix_ambiguous_min_error
PiperOrigin-RevId: 558698327 | TensorFlower Gardener | gardener@tensorflow.org | 1,692,602,655,000 |
7a024da3da7f12eb98b8a8c13fb6b6f8d033cf93 | Wrap remaining unfused instructions in fusions before conversion to LHLO.
After this, `HloToLhloModule` no longer creates additional fusions, which
means we can easily produce a mapping from MLIR operation to corresponding
HLO instruction (and their types will match). After this, we can remove
the MLIR->HLO conversion... | Johannes Reifferscheid | jreiffers@google.com | 1,692,602,677,000 |
ceb4d4137bfa020a3e04b4a50b78127f2f9a1d38 | compat: Update forward compatibility horizon to 2023-08-21
PiperOrigin-RevId: 558718662 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,608,534,000 |
aa8104b7a9469a13e2512d444e8095c34de7e542 | Update GraphDef version to 1595.
PiperOrigin-RevId: 558718663 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,608,534,000 |
f192e07982b880ca344d0234e1672d2e906cb5ee | The test cases regarding CalibrationOptions were simplified.
PiperOrigin-RevId: 558721358 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,609,196,000 |
2e4a6d0f2b37d03d31bfabd3c79e4958a0f508ee | [XLA:GPU] Triton GEMM: fix handling of degenerate fragments of split tensor dimensions.
PiperOrigin-RevId: 558743824 | Ilia Sergachev | sergachev@google.com | 1,692,616,077,000 |
953443fedcafbd4d2d0f7a9891bcf4e2a7e466f4 | [DelegatePerformance] Move `finish()` into the accuracy activity impl class to prevent the application from finishing early.
PiperOrigin-RevId: 558756398 | Shiqing Yan | shiqing@google.com | 1,692,620,019,000 |
c177dbe4c2f6b277e0ba439a5106aa39c7a2540a | Use the original HLO ops in ir_emitter_unnested.
This lets us skip the MLIR->HLO conversion. NFC.
PiperOrigin-RevId: 558767519 | Johannes Reifferscheid | jreiffers@google.com | 1,692,623,593,000 |
6064691f03d71d25d554472055a04f990e6d74c3 | #tf-data Some pipelines can be very deep, built using many `map` transformations. Such pipelines can sometimes cause stack overflow. This CL moves temporary iterator context of `ParallelMap` iterator initialization from the stack space to the heap to reduce the stack space usage.
PiperOrigin-RevId: 558774496 | Wilsin Gosti | wilsin@google.com | 1,692,625,803,000 |
aa6995feb4ca9559122f4950d01f352212482a20 | [XLA:GPU][NFC] Triton GEMM: log computations in shorter format.
PiperOrigin-RevId: 558774941 | Ilia Sergachev | sergachev@google.com | 1,692,625,930,000 |
fd6347be708dc955ef01f7db40b08e4ff6942ad2 | Add APIs to configure delegate-created tensors.
PiperOrigin-RevId: 558776662 | Matt Kreileder | mattalexander@google.com | 1,692,626,455,000 |
bbb8c4f5e52e0cbec4f442be28316c5d45561e24 | Remove logic for handling LaunchFunc ops.
These can't actually exist at this stage of the pipeline. Previously
they could, but the functionality was removed in an earlier cleanup.
PiperOrigin-RevId: 558784555 | Johannes Reifferscheid | jreiffers@google.com | 1,692,628,722,000 |
1887af01e86861dcc8b5c827547e7a3d7873f151 | Change fuzzing annotations from custom type to Annotated.
PiperOrigin-RevId: 558788270 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,629,738,000 |
48043f5993743a9bba3a7b4adc9155cb66bbe6b2 | Update targets from aliased targets to real targets
PiperOrigin-RevId: 558788608 | Mason Chang | masonchang@google.com | 1,692,629,828,000 |
222eb2a82ac3a1fbbbba75a9b2bb89a19ff2276b | Implementation for resize + tests.
PiperOrigin-RevId: 558798076 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,632,118,000 |
5871524a8f4f4ca3364b566e191a063be11b4b5a | GlobalIterId and GlobalFrameId ops are not Pure
The Pure side effect was leading to incorrect hoisting.
PiperOrigin-RevId: 558812661 | Pat Notz | patn@google.com | 1,692,635,407,000 |
75d692bb9def905f2d84970d50e9edb137fbf1a5 | Remove deprecated "distributions" package from exports.
It has been deprecated since 2019, and is now triggering warnings on
TensorFlow startup. It does, however, still need to be included in
`python/__init__.py` so that it can be found by `tf.compat.v1.distributions`,
which several projects depend on.
We will remov... | Antonio Sanchez | cantonios@google.com | 1,692,636,316,000 |
11b4f12886b5db44b4a2b34e65a0be56b27974fd | Make all Python targets under mlir/quantization/tensorflow/python/ have strict dependencies.
PiperOrigin-RevId: 558816973 | Juan Martinez Castellanos | juanantoniomc@google.com | 1,692,636,332,000 |
c82530710acef5d250913ccb72d1fb7bb49bad19 | Add new toolchains for Clang-17 testing
PiperOrigin-RevId: 558820276 | Kanglan Tang | kanglan@google.com | 1,692,637,014,000 |
881e416ae5a00db4bda5289df948ff5f5049fe2e | Pass tensor name through API
PiperOrigin-RevId: 558820894 | Haibo Huang | hhb@google.com | 1,692,637,140,000 |
1f03579f68a74fe8d71d159e1a09430d0fe099c7 | Transition away from deprecated perftools::gputools namespace to stream_executor namespace.
PiperOrigin-RevId: 558826606 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,638,182,000 |
ad9c85163d9fe412e3f9a77eb73f36ca18cae2d6 | Add the newly exposed TF Quantization to the list of target patterns for the new API generator.
PiperOrigin-RevId: 558829045 | Brian Wieder | bwieder@google.com | 1,692,638,665,000 |
117e2b85bc869d3e45addc01f650113967d7aa7f | Sort the while body function before converting to async while function body to a first ready first run order.
PiperOrigin-RevId: 558829222 | Deqiang Chen | deqiangc@google.com | 1,692,638,700,000 |
4e6d567dc7e01c61c8cae4146a64cc1444957599 | Transition away from deprecated perftools::gputools namespace to stream_executor namespace.
PiperOrigin-RevId: 558831566 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,639,155,000 |
5a04bcd59c64bb1a4a47dd7dd22a5dc9714a9daf | [JAX] Add xla_client.Client.device_from_local_hardware_id method.
This is exposing PjRtClient::LookupAddressableDevice in Python.
PiperOrigin-RevId: 558837006 | Skye Wanderman-Milne | skyewm@google.com | 1,692,640,167,000 |
54d905febbea201821f21a14cd133c22e3746e82 | Make all targets under tensorflow/lite/toco/python/ have strict dependencies.
PiperOrigin-RevId: 558850986 | Juan Martinez Castellanos | juanantoniomc@google.com | 1,692,642,692,000 |
f45af4dfc0c4e8a1764c84c85f3b6dbee89ff834 | Copy tensor through C API for rendezvous
PiperOrigin-RevId: 558851634 | Haibo Huang | hhb@google.com | 1,692,642,812,000 |
9af342313607927ba349e601383f0ef9258d2ca3 | Do not do decomposition passes when avoiding lowering to xla in the mlir bridge.
Do not attempt to populate IO info into compilation result if not lowering.
PiperOrigin-RevId: 558851812 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,642,849,000 |
aea17717b2f5a5efcc582522da44e9da37bbca0b | Refactor rbe configs in TF's .bazelrc files
Changes include the following:
- Merge `rbe_linux_py3` and `rbe_linux_base` into `rbe_linux`
- Rename the `rbe` config defined in the root bazelrc to `rbe_base`. This avoids loops when the `rbe` config in cpu/cuda.bazelrc depends on rbe configs in the root bazelrc. There wi... | Kanglan Tang | kanglan@google.com | 1,692,645,138,000 |
c6f438ec935a4a88547be6aa3a56105654edaf06 | Allow MSA to reduce size of scoped memory if input/output of the instruction has been assigned in alternative memory.
PiperOrigin-RevId: 558863961 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,645,178,000 |
f4315bbba7f1a974a27ec5bd8c36bae1fa49a1a8 | #tf-data Add logging for creating IndexSplitProviders.
PiperOrigin-RevId: 558867250 | Yang Chen | yangchen@google.com | 1,692,645,887,000 |
02f9aecc13950c41a740c239e2f02c0279f657ff | Append \0 to the end of strncpy dest
PiperOrigin-RevId: 558867821 | Haibo Huang | hhb@google.com | 1,692,646,014,000 |
a242d74e61d9e43923ab9010e504d4ad5ecc290d | [Memories][PJRT:C] Add `PJRT_Memory*` to `PJRT_Client_BufferFromHostBuffer_Args` to support the variant of `PjRtCApiClient::BufferFromHostBuffer` which takes `PjRtMemorySpace*` as an arg.
PiperOrigin-RevId: 558873596 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,647,341,000 |
184c410fa1c0f5eb20765c2dd6d8ba34cd719330 | [xla:gpu] Fallback to op-by-op mode if graph instantiation fails from OOM error
PiperOrigin-RevId: 558876961 | Anlun Xu | anlunx@google.com | 1,692,648,118,000 |
81eae5489956612985c8359ef6a8f8815eb94801 | Fix use-after-free in Rendezvous API
PiperOrigin-RevId: 558880510 | Haibo Huang | hhb@google.com | 1,692,648,859,000 |
8a0fee00855a0e806bd5c9cc1ad6c0175a985922 | [XLA] Don't use isnan on values which can't have NaN.
While we are here, don't upcast to double just to check if something is NaN.
PiperOrigin-RevId: 558885664 | David Majnemer | majnemer@google.com | 1,692,649,935,000 |
142e0d68ffd60b4313a621b87780369ad8b96195 | Change fuzzing annotations from custom type to Annotated.
PiperOrigin-RevId: 558895820 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,652,079,000 |
088425456ef26287048aedab61dd8ed7457749f1 | [XLA] Improve memory pressure heuristics in latency hiding scheduler
A few improvements:
1) Use Depth as a way to explore the graph for better memory pressure reduction nodes.
Currently there are cases where we can get stuck in local minimas where the scheduler
picks the less memory pressure increasing node, in... | Marcello Maggioni | maggioni@google.com | 1,692,652,115,000 |
37d5069ef9fe0f4410f3a280c777e715c223b0bf | Make all Python targets under tensorflow/lite/g3doc/**/ have strict dependencies.
PiperOrigin-RevId: 558896107 | Juan Martinez Castellanos | juanantoniomc@google.com | 1,692,652,146,000 |
f23e87cee05d6dc38bfe2767c1914185a51390a2 | [PJRT C API] Remove LoadTpuLibraryAndInitializeTpuStructFns in xla.cc.
TPU profiler C API initialization has been moved to tpu_tracer.cc (https://github.com/openxla/xla/commit/9f70d6df0f4d323406ea572bee0e44a9ac2c109f). Currently TPU embedding C APIs initialization is done in tensorflow code (jax_tpu_embedding triggers... | Jieying Luo | jieying@google.com | 1,692,653,009,000 |
4682b81cc4ac3782b3f33ee22f80c54db065a4b5 | Add bazel macro to include the TF_Tensor implementation for Tensorflow but only headers when building LibTPU
PiperOrigin-RevId: 558907077 | David Silverstone | silverstone@google.com | 1,692,654,658,000 |
5bab8e92b27c799cf2ca0eb3cda7d9e1dd946d41 | [XLA:Python] Move contents of :xla_extension py_extension() rule into a cc_library().
Replace the py_extension() rule's contents with a shim.
This change works around linking problems that mean py::module_local() does not work correctly when built in Google's environment. It should have no negative effects to others.... | Peter Hawkins | phawkins@google.com | 1,692,655,799,000 |
c079067f8f177c7cac7e36d3119c51b13d55bb78 | Fix uniform attributes utils and generated quantized functions.
For SRQ UniformQuantizedConvolution, output's axis will be determined by weight instead of hard-coded -1. For UniformRequantize, dynamically set output axis based on output scale rank.
For operations where the inputs are qint8 and the output is qint32, c... | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,656,173,000 |
a5c35aa20982d5a3805640a855354347e55e41a7 | #tf-data Fix missing includes.
PiperOrigin-RevId: 558918116 | Yang Chen | yangchen@google.com | 1,692,657,155,000 |
26333798e85d0ff2cf80d9262c83ef8ece29ab3d | [XLA:Python] Mark XLA classes as pybind11::module_local().
Fixes duplicate registration of XlaBuilder between JAX and PyTorch/XLA.
PiperOrigin-RevId: 558928043 | Peter Hawkins | phawkins@google.com | 1,692,659,505,000 |
39f8bbfbeebab3ec2130fbfcdb7f1dbd1b70b355 | Move `CompileToPjRtLoadedExecutable` to a util file.
Also adding a new version of the function which does not take `OpKernelContext`.
PiperOrigin-RevId: 558929210 | Changhui Lin | changhuilin@google.com | 1,692,659,820,000 |
7397533b82a60298bdf35cb6ddcfcd2c388aaab1 | Releases TfCThunkDeviceContext to prevent memory leak
PiperOrigin-RevId: 558932800 | Haibo Huang | hhb@google.com | 1,692,660,756,000 |
6f00e33d3e574225f7d2613482085d5884ca222c | protect libtpu single-user multi-loading with guarded lockf removal atexit
PiperOrigin-RevId: 558934747 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,661,248,000 |
4bbbb545e657769770dd07a1d2d2baf9933919fa | [XLA] Drop sharding annotations (if any), on instructions modified by the reshape mover pass. The auto-sharding pass, when turned on, runs before reshape mover. Reshape mover however, can change the shape of elementwise ops, leading to inconsistent sharding specifications. This CL drops any such shardings, if present. ... | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,662,097,000 |
a29c0ca864bfada67bf28cf5375c5daee469192b | Register TF UniformQuantized Ops as MlirXlaOpKernel
PiperOrigin-RevId: 558940608 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,662,734,000 |
0374961a42575447718514dc038d4e86ba878205 | Revert: [XLA:Python] Mark XLA classes as pybind11::module_local().
Fixes duplicate registration of XlaBuilder between JAX and PyTorch/XLA.
PiperOrigin-RevId: 558943318 | Peter Hawkins | phawkins@google.com | 1,692,663,486,000 |
3fa0dc17c1f70767be6fc99256835c6b0cd332e2 | Export errors from `Executor`
PiperOrigin-RevId: 558946514 | Juanli Shen | juanlishen@google.com | 1,692,664,446,000 |
b49318cd03aff87fed1c625a08fc27d0cbc5b9e3 | [xla:gpu2] NFC: Build a more efficient IR for importing tensors
1. Use iree_input.tensor.subspan to import memref views as tensors. Buffer views should be almost never constructed in input programs.
2. Cache optimization barriers between different compiled ops lowerings and make sure they are in the parent function en... | Eugene Zhulenev | ezhulenev@google.com | 1,692,668,277,000 |
9af64a6dc28269114240939fa2954623da95c2b3 | Update `HloBufferDonorConfig::Verify` and `HloVerifier`. The overlap between buffer_donor_config and input_output_alias_config is not allowed.
PiperOrigin-RevId: 558957704 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,668,302,000 |
e56206755ff0c98269eb3e50c98fccbaadb6884d | Support quantized i64 during flatbuffer import
Sometimes tflite flatbuffer represent `i32` quantized values as `i64`s. In these cases we should truncate down to a lower bit width to avoid creating illegal types.
We check the bitwidth of the type at load time and pick a power-of-2 bit
width where the value can be safe... | Robert Suderman | suderman@google.com | 1,692,668,542,000 |
30b5dfe2bda36a9cf46d50fff47cc29fb347dee8 | resolved rsync failure on Windows | Raunak | mayank.kumar.raunak@intel.com | 1,692,670,241,000 |
a62e6c5970f88f374a71cc8fc4ae2f51e6f805e8 | Indentation of comments | mraunak | 83710963+mraunak@users.noreply.github.com | 1,692,670,647,000 |
6c761bf7e13978fb2973ae45f3e95da591059000 | Deprecate old graph export pipeline in generic pipeline
We migrate to graph export pipeline v1 in generic pipeline. In the v2 version, the control dependencies are only added in the end of the pipeline and are more concise.
PiperOrigin-RevId: 558968297 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,672,395,000 |
7ac9d9826da24c1c2682c64bf03c56d130d8c657 | Create Uniform Quantized opset numerical tests for conv2d.
PiperOrigin-RevId: 558980724 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,677,895,000 |
a9f79e089af7c1c166cac45dd0e33e4e9285884f | Internal Code Change
PiperOrigin-RevId: 558990409 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,681,022,000 |
06ba93d0968814d48975d980ec0479c92b8a37f6 | Remove unused, legacy forwarding libraries
PiperOrigin-RevId: 559021838 | Thomas Köppe | tkoeppe@google.com | 1,692,689,220,000 |
e5035bcd43fe7346f20dcd581672066ea159a5fa | [XLA:GPU][NFC] Refactor Triton GEMM rewriter.
Change the tensor iteration spec storage type to support arbitrary number of tensor dimensions; make dimension order map merging safer; update print function.
PiperOrigin-RevId: 559035669 | Ilia Sergachev | sergachev@google.com | 1,692,693,288,000 |
76474187d60e8db2db1f5f6342d43975e3519ab4 | [XLA:GPU] Manually tune the number of warps launched depending on the length of
the reduction dimension in Triton Softmax emitter.
PiperOrigin-RevId: 559041031 | Benjamin Chetioui | bchetioui@google.com | 1,692,694,879,000 |
b2283e64ccb95922b11d9f9fe412f4e749f37022 | compat: Update forward compatibility horizon to 2023-08-22
PiperOrigin-RevId: 559041277 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,694,939,000 |
64126d8f6af16bac526675c8d90865a291f3c38a | Update GraphDef version to 1596.
PiperOrigin-RevId: 559041801 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,695,062,000 |
fe0e7551b067029674db2d88ebfc560d6dc70510 | [XLA:GPU] Triton GEMM: refactor the emitter.
- Enable loading of both block pointers and legacy pointers (for scalars).
- Move softmax tile mask generation into the emission of reduction where it's used.
- Further decouple properties of different tensors in GEMM fusions.
- Make advancing tensor pointers optional w... | Ilia Sergachev | sergachev@google.com | 1,692,695,979,000 |
c47f90ab02344f68921e9f3e9922a722541d0972 | Internal Code Change
PiperOrigin-RevId: 559046662 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,696,486,000 |
25ac2a5882f1f225a33671e76a209ff99e2bdf4f | Add a helper for finding the parameters of a fusion.
Also remove a hack that was previously needed because in codegen we would be
dealing with invalid HLO. This is no longer the case.
PiperOrigin-RevId: 559047400 | Johannes Reifferscheid | jreiffers@google.com | 1,692,696,691,000 |
ce838b191d17108e71a18a256ebf9dbc9c2aaf15 | Ensure that pywrap_quantize_model.so has correct RPATH
The new shared object pywrap_quantize_model.so needs to have its
RPATH updated to find all dependent other shared objects. | Andrew Goodbody | andrew.goodbody@linaro.org | 1,692,699,593,000 |
d9d9d5c134d3324ca01448d49ca15ff302b5244e | [XLA:GPU] Add pass to collect cost model stats.
PiperOrigin-RevId: 559059193 | Oleg Shyshkov | shyshkov@google.com | 1,692,700,443,000 |
243eadb20533d898caf7c42885d1b7365cda0274 | Integrate LLVM at llvm/llvm-project@0e17372b3804
Updates LLVM usage to match
[0e17372b3804](https://github.com/llvm/llvm-project/commit/0e17372b3804)
PiperOrigin-RevId: 559068190 | Krasimir Georgiev | krasimir@google.com | 1,692,703,364,000 |
2770dd15b6320ceae7d327d51c81a81177875e14 | Move kernel gen jit outlining to the whole function level.
Before, the outlining would match a compatible operation and produce a
corresponding jit region. However, we end up always outlining the whole bode of
the function anyway. With this change, we unconditionally do this independent of
the contained operations. Th... | Stephan Herhut | herhut@google.com | 1,692,703,530,000 |
bfbf57a2e7064a25c422496a56cb0b4ec2c8c169 | [XLA:GPU] Put split-K autotuning behind the flag (enabled by default).
Disabling split-K autotuning is often used during debugging and having a flag which disables it is much more convenient, than setting a parameter to 1 in the code manually each time.
PiperOrigin-RevId: 559072593 | Aliia Khasanova | aliia@google.com | 1,692,704,933,000 |
31ec3439977ac587d5b6c3fa614d1fa228a4060a | Update TFRT dependency to use revision
http://github.com/tensorflow/runtime/commit/c167ef1ccb7f8b7a234c8bebdcab269933b87164.
PiperOrigin-RevId: 559075080 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,705,698,000 |
c5ccdbe0228c2d3f4f226c61b4505c6b27669e39 | Remove unused, legacy forwarding libraries
PiperOrigin-RevId: 559079686 | Thomas Köppe | tkoeppe@google.com | 1,692,707,168,000 |
c778745f5fe7ab086b6e17290e329caeab070cd8 | [XLA] Include HLO opcode to pipeline in CollectivePipeliner config.
- Specify the HLO opcode the pipeliner config, and use that to derive
more descriptive pass name
- Some minor code cleanup.
PiperOrigin-RevId: 559108348 | Rahul Joshi | jurahul@google.com | 1,692,715,115,000 |
343e10970c7c8f3b96d0fc44f5be8e43b34f4eb1 | Remove I64 from MeanOp axis schema declaration
PiperOrigin-RevId: 559110599 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,715,763,000 |
616aa8414f58f5477e6350bf53372c9a47042765 | Updated the file with necessary changes | Harshavardhan Bellamkonda | harshavardhan_b@srmap.edu.in | 1,692,717,452,000 |
54c45bfbe89efc12fdb68483e6d7488ca5c74408 | Merge branch 'tensorflow:master' into master | halseycamilla | 111537937+halseycamilla@users.noreply.github.com | 1,692,718,800,000 |
de1deaa841f85e2a5ff3c20678855185bfc137e0 | Change fuzzing annotations from custom type to Annotated.
PiperOrigin-RevId: 559124488 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,719,327,000 |
dee7bb553858422fc89b0c4a5d2ce9d614a2ef95 | Add control dependencies for input output buffer alias. If needed, we add control dependency to enforce output-related instructions after the users of input buffers.
PiperOrigin-RevId: 559127871 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,720,131,000 |
b68d701a6f21abcbc623ffb4d8961b95aa9ed7e1 | [XLA:GPU] Drop RecordEstimatedRunTime in multi output fusion.
Logging cost model data is now handled by GpuCostModelStatsCollection pass.
PiperOrigin-RevId: 559130362 | Oleg Shyshkov | shyshkov@google.com | 1,692,720,675,000 |
903ae956995d6f59b22d8b3813a690936ec24a79 | [XLA/GPU] Enable LHS GPU async tracker by default.
PiperOrigin-RevId: 559135036 | Rahul Joshi | jurahul@google.com | 1,692,721,701,000 |
fef811c6e576e07ce45e74da21705abe73ac8d17 | Removes the default implementation of the non-pure virtual method: `ifrt::Array::FullyReplicatedShard` and there by turns the `ifrt::Array` into an abstract class that describes only the API (and not the implementation).
PiperOrigin-RevId: 559137803 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,722,314,000 |
3cb4df1164a52d7eec500b48500d8f69f3f9c88a | Update args description of crop_to_bounding_box & pad_to_bounding_box
Currently the documentation of the API tf.image.crop_to_bounding_box doesn't mentioning the acceptable dtypes of the arguments offset_height, offset_width, target_height, target_width. This API accepts only tf.int32 Tensor or a integer.For others it... | Surya | 116063290+SuryanarayanaY@users.noreply.github.com | 1,692,722,751,000 |
7061d3eded44e956317fe4f297c4ef70553b1242 | Recursively do mesh propagation inside while region.
(1) Use walk instead of getOps to recursively do mesh propagation.
(2) Rewrite ExtractMeshFromBlockArgument to support both function and while region.
(3) Add a unit test with GPU and CPU mesh
PiperOrigin-RevId: 559143426 | Tianrun Li | tianrun@google.com | 1,692,723,515,000 |
12c90718b7b5712b9de41be7fd127e32f0cdba99 | GlobalIterId has a side effect, but has no side effect edges. The side effect prevents constant folding and hoisting.
PiperOrigin-RevId: 559146333 | Michael Delorimier | mdel@google.com | 1,692,724,050,000 |
a5dc7279354160b7051742ab2a2f419b142544fc | Check in generated pyi files for some py_extension targets.
PiperOrigin-RevId: 559157112 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,726,085,000 |
1bf3b3c9d0c98d262a97ac302e69d238bca21a36 | Check in generated pyi files for some py_extension targets.
PiperOrigin-RevId: 559157532 | A. Unique TensorFlower | gardener@tensorflow.org | 1,692,726,162,000 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.