hash stringlengths 40 40 | msg stringlengths 1 131k | author stringlengths 1 33 | email stringlengths 0 57 | date int64 1,447B 1,698B |
|---|---|---|---|---|
71778cae789e9ef483af6ca59c97abc96f5a4b69 | Add type hints for ops.Operation in cudnn_rnn_grad.py.
PiperOrigin-RevId: 561208020 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,364,545,000 |
115bbd9a0b1a620c69ea59dc7e79c638f90f4164 | [tflite] Add stablehlo.scatter kernel to tflite runtime.
PiperOrigin-RevId: 561220243 | Majid Dadashi | majiddadashi@google.com | 1,693,368,991,000 |
2755f08a0e42f12bdfcd9d791cb42e2b07b09db0 | Implement whole model verify mode using DumpTensorOp
DumpTensorOp saves entire value of quantizable layer during inference.
Whole model verify creates two models:
1. Unquantized model with DumpTensorOp - saved at unquantized_dump_model_path
2. Quantized model with DumpTensorOp - saved at output_directory of quantized... | Kyuyeun Kim | kyuyeunk@google.com | 1,693,369,963,000 |
d32365d27709198391a73573cac5a1e97e2d4a9d | Implement per-layer verify mode using DumpTensorOp
Per-layer verify creates a model with both quantized and original unquantized layer.
Unquantized layer's input comes from previous layer's quantized layer.
Outputs of quantized and unquantized layers are saved via DumpTensorOp
PiperOrigin-RevId: 561233115 | Kyuyeun Kim | kyuyeunk@google.com | 1,693,373,420,000 |
512f4b101762dba010e75f93b84cb096b075992c | Update `CalibrationWrapper` to support models larger than 2GB using offset buffers.
Add implementations to handle the case where the calibration wrapper accepts a TFLite flatbuffer model that uses offset buffers.
Offset buffers are enabled by the TFLite converter's `_experimental_use_buffer_offset` flag.
Whenever the ... | Dan Suh | dansuh@google.com | 1,693,379,120,000 |
4067c19364347404e6b2b444bb53ff4f2b439015 | Arm NEON int4->float32 cast
PiperOrigin-RevId: 561263745 | Alan Kelly | alankelly@google.com | 1,693,383,332,000 |
75e5c93acff9803f337cb6f0794073ab9ca1143d | Update GraphDef version to 1604.
PiperOrigin-RevId: 561272265 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,386,107,000 |
8709187ca14ead8b3a70a1cf58aec7b8434cdb02 | Merge branch 'tensorflow:master' into toupstream/int6x8_32_accum | SaoirseARM | 44364573+SaoirseARM@users.noreply.github.com | 1,693,386,547,000 |
c4f212a0623e1a303d61ca935de3dc2b2597b670 | compat: Update forward compatibility horizon to 2023-08-30
PiperOrigin-RevId: 561272373 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,386,133,000 |
ffd3326d98663f6b203f29e909ea9aed8e25a441 | Cast Int4->Float32 bench
PiperOrigin-RevId: 561274977 | Alan Kelly | alankelly@google.com | 1,693,386,861,000 |
5b5acfe3289ac9169d4d1d0692a4c405e9db7a2d | Internal Code Change
PiperOrigin-RevId: 561283397 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,389,712,000 |
c2888c7f6fa6149aaf167a807aa584dbdd9b197b | Try to eliminate more copies after scheduling.
After scheduling, it is possible to eliminate more copies that have been added
by the CopyInsertion pass, because now it is clear in which order the
operations are executed.
PiperOrigin-RevId: 561300243 | Adrian Kuegel | akuegel@google.com | 1,693,395,136,000 |
ccf19eb10b6ceb6395322322920e2a652d174699 | Update OutputType in image_segmenter.md
The line
```
segmentation_options = processor.SegmentationOptions(
output_type=processor.SegmentationOptions.OutputType.CATEGORY_MASK)
```
is changed to
```
segmentation_options = processor.SegmentationOptions(
output_type=processor.SegmentationOptions.outp... | Pratik Joshi | 118897289+pjpratik@users.noreply.github.com | 1,693,396,547,000 |
e6a43eb8b85e8651d2be1d2d75192597f4f1579c | Integrate LLVM at llvm/llvm-project@d77ae428e086
Updates LLVM usage to match
[d77ae428e086](https://github.com/llvm/llvm-project/commit/d77ae428e086)
PiperOrigin-RevId: 561318672 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,400,738,000 |
9133f87cde3bcc37bd5afe270c2c815365bd5f96 | Cleanup aliased targets and point to real targets
PiperOrigin-RevId: 561321137 | Mason Chang | masonchang@google.com | 1,693,401,470,000 |
dd96aed9b59dfe680dd0a9e65365ce6b4a76ae7c | Fix argument naming in `UndoBroadcastFullyConnectedBiasAdd` and `UndoBroadcastConvBiasAdd`
PiperOrigin-RevId: 561329337 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,404,087,000 |
09625f2ae8303c96de21bee63c55a475cef3fa77 | [xla:gpu] Don't pass both GpuDeviceInfo and GpuVersion (or CudaComputeCapability).
Pass compute capability to priority based fusion (which got dropped in cl/544028940, presumably by accident).
PiperOrigin-RevId: 561331858 | Christian Sigg | csigg@google.com | 1,693,404,855,000 |
949b63fa788411e4ddba84bbd70385aedd4ed4a2 | Fix broken build by using correct intrinsics
PiperOrigin-RevId: 561337516 | Alan Kelly | alankelly@google.com | 1,693,406,505,000 |
256a18372389714caf1328f9545a7a8d17dbd374 | Rollback because of a broken test
PiperOrigin-RevId: 561347423 | Eugene Zhulenev | ezhulenev@google.com | 1,693,409,136,000 |
dd6f2218a657c7d91c516a7fc9dc34efce040370 | Remove release_gpu_windows config and deprecated Windows GPU jobs
TensorFlow no longer supports GPU on native-Windows since 2.11 (https://github.com/tensorflow/tensorflow/releases/tag/v2.11.1).
PiperOrigin-RevId: 561366344 | Kanglan Tang | kanglan@google.com | 1,693,413,609,000 |
bf8a20b6bc6b244947bf173bfb00793a67d2949f | Applies additive salt to the terms in the Auto Sharding objective function.
PiperOrigin-RevId: 561368806 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,414,173,000 |
570ae0b638c47907e0d8e72c3d605303338dec38 | [xla:gpu] Remove redundant Cuda/RocmComputeCapability arguments in `CompileModuleToLlvmIr` and `IrEmitterContext`.
PiperOrigin-RevId: 561369540 | Christian Sigg | csigg@google.com | 1,693,414,331,000 |
a7badba3e9609616dd4b4f5042500f460d5fc98d | Implementation for RGB to YUV conversion + tests.
PiperOrigin-RevId: 561371282 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,414,713,000 |
68fb0f073765b3a0d00b3c07d1948c2780d153c9 | Remove ops/nn_impl.py's dependency on distribute/distribute_lib.py.
Make a new file, ops/nn_impl_distribute.py, and move the functions in nn_impl.py that depend on distribute_lib.py into it.
PiperOrigin-RevId: 561371645 | Fiona Lang | flang@google.com | 1,693,414,796,000 |
f8aa319e43a4edde2b0fa4eccdaa85cfcabf59d4 | Do not run argminmax_test in asan config due to timeouts.
PiperOrigin-RevId: 561372883 | Xin Zhou | zhouxin@google.com | 1,693,415,011,000 |
f80729bbccdedc37125d14eaf3b60f87df1fe206 | Factor out quant/dequant in QuantizeWeightsPass
This does not affect the end results but wraps current uniform
quantization scheme under "dequantize" function so that
1) passes after this pass can just use "dequantize" to propagate quantized
values instead of specifying "dequantize_i8"/"dequantize_to_float"/ etc.
2) a... | Songyi Han | syhan@google.com | 1,693,415,248,000 |
da0a6924f5f8306d31b3641973342aa7df1104d8 | Improve FFTN ops
1. Add argument "norm" for *fftn ops.
2. Raise an error when axes is not proper.
Note that the Gradient support for FFTN is still missing.
Submitting on behalf of sunrj@.
PiperOrigin-RevId: 561376369 | Yu Feng | feyu@google.com | 1,693,415,679,000 |
4b713c048527764f6d684fd32a995ea8da01ef1b | Add IWYU (include-what-you-use) "export" pragmas to some headers in lite/ that are part of the "framework_stable" target,
to suppress lint warnings that currently encourage the direct use of the corresponding headers in
lite/core/api. It seems easier for clients to refer to these headers in lite/ and thus have fewer
b... | Fergus Henderson | fergus@google.com | 1,693,416,854,000 |
20bdcf23e1b68c5a3f10ccb592be7439d7f1b27e | Make RaggedTensor responsible for registering ragged ops.
PiperOrigin-RevId: 561385595 | Fiona Lang | flang@google.com | 1,693,417,399,000 |
2a7efd891d3b16ef82b462d76fd9e61d111bf901 | [tflite] Add stablehlo.scatter serialization.
This adds the MLIR to Flatbuffers logic for stablehlo.scatter.
PiperOrigin-RevId: 561396330 | Majid Dadashi | majiddadashi@google.com | 1,693,419,437,000 |
ae85400566790d3bd2f3df5688c6d482768260bb | [XLA] Modify XLA tests to conform to the new pattern matching format.
-Add no operand support for XLA_VARIADIC_OP_PATTERN. For example, fusion can take zero or more operands.
-Add WithMinorToMajor implementation for Layout to match with the minor to major dimension ordering.
-Add WithLayout implementation for Shape to... | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,421,381,000 |
d1333f4bf52d31bc3a0da008ef39b9e52c85ac6a | Updata atan2 xla doc | Alexander Pivovarov | apivovarov@gmail.com | 1,693,268,590,000 |
91a3294f0a27a3c6f70e7a291347a6ed951ea884 | Optionally enable `tf-tpu-variable-runtime-reformatting` pass in bridge, enabled by default.
To disable set flag `tf-tpu-variable-runtime-reformatting` to false in `TF_XLA_FLAGS`.
For example,
```
TF_XLA_FLAGS = 'tf-tpu-variable-runtime-reformatting=false'
```
PiperOrigin-RevId: 561408990 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,422,077,000 |
8d124ea3a2e75a740c15512e051b9c79da262597 | Update references to NoneTensorSpec in its new location.
PiperOrigin-RevId: 561411978 | Fiona Lang | flang@google.com | 1,693,422,684,000 |
e3b1f16096c65ba9fc470c7e12531fe9a50bf983 | Implementation for YUV to RGB conversion + tests.
PiperOrigin-RevId: 561413193 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,422,967,000 |
8587860d4aad2d0f0a774a9c74b77dfe63a78e58 | Allow `tensorflow::ConfigProto` to determine whether we allow parallel execution of remote functions in the eager context.
PiperOrigin-RevId: 561421703 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,424,857,000 |
adda263d8c6ff84da45294d8d4539b9488a538cb | [tflite] Add stablehlo.scatter de-serialization.
This adds support for converting `stablehlo.scatter` in the tflite Flatbuffers file to MLIR.
PiperOrigin-RevId: 561434324 | Majid Dadashi | majiddadashi@google.com | 1,693,427,514,000 |
04a0c8ad7dbe98054ffee159d30839eb520b5d6e | Add build:release_base configuration option back.
PiperOrigin-RevId: 561438023 | Juan Martinez Castellanos | juanantoniomc@google.com | 1,693,428,344,000 |
4f42eb374acd6f5485065c7d8b1fbdbe5df44a83 | Add bfloat16 registration for StatelessMultinomial op.
Motivation: bfloat16 is supported via XLA bridge / MLIR, but not as a vanilla TensorFlow op. This registration will avoid user confusion when running this op in a context that does not invoke XLA.
PiperOrigin-RevId: 561459041 | RJ Skerry-Ryan | rjryan@google.com | 1,693,432,659,000 |
970f1bd9562a7de3b46545db4d5a4e5fd6d9b672 | #tf-data Fix docs for Options.autotune_options
PiperOrigin-RevId: 561460845 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,433,058,000 |
5b8e70c145d9496d2ad3630b02ed00f84d724684 | Support names other than `__init__.py` as the root file for the generated API in the new API generator.
PiperOrigin-RevId: 561464914 | Brian Wieder | bwieder@google.com | 1,693,434,010,000 |
4aa0659cac381691817b02cc5a33ac16ccc6c433 | Support for select in generic pattern matching for FP8 GEMMs. | Philipp Hack | phack@nvidia.com | 1,693,435,395,000 |
1f3db3d2be511bc97af1b8925c1391a52dc62633 | Update TopkRewriter to handle tensors with rank > 2
Just add reshapes before and after if necessary.
PiperOrigin-RevId: 561479066 | Kevin Chen | kevinch@google.com | 1,693,437,254,000 |
7f522660d4f4c97aff969f15a9ee75c1bee5adb9 | Extend matmul ptq test cases.
PiperOrigin-RevId: 561480889 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,437,719,000 |
e189198ecbfc7d11a94b0b1eb078da4c56485693 | Preparation to move Servo TFRT integration code out of google folder to oss
PiperOrigin-RevId: 561481469 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,437,888,000 |
3dd1d205908fdb87df81d6f1c764f83b8ef75f05 | Add a link to go/tf_mirror_md in workspace2.bzl where tf_http_archives are being used.
PiperOrigin-RevId: 561483912 | Camilla Halsey | halseycamilla@google.com | 1,693,438,546,000 |
b036baa47330c0f4e1cec981e12e9d2e32322070 | #tf-data When calling from `standalone.cc`, the tf.data model could have been created even when autotuning is off. This CL resets it back to `nullptr` when autotuning is off.
PiperOrigin-RevId: 561486415 | Wilsin Gosti | wilsin@google.com | 1,693,439,218,000 |
a81534844e4d9749c6d9b0f787b17dfd13e0620f | Implementation for crop to bounding box + tests.
PiperOrigin-RevId: 561491781 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,440,700,000 |
1f4fb0f73255182afa5d96a8fdd6acaf4f63e6b8 | No public description
PiperOrigin-RevId: 561492053 | Fiona Lang | flang@google.com | 1,693,440,780,000 |
27bc79a94808725bffa24972501d1a4197790729 | Set up sample presubmit configurations, and adjust copybara invocation
PiperOrigin-RevId: 561492256 | Austin Anderson | angerson@google.com | 1,693,440,845,000 |
83e3f0675b287e00e942fd78fe8a244cad7ec0c0 | Add type hints for ops.Operation.
PiperOrigin-RevId: 561492951 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,441,030,000 |
a395830b79b4224a66b0d718d14a5eff799cb616 | Port DTensor to new DEBUG_DATA_DUMP() system.
Some slight tweaks in the new dumper config base class to plumb through the
operation being dumped, such that we can extract the operation name from
the module.
Example:
```
> TF_DUMP_GRAPH_PREFIX=/tmp
> TF_DUMP_GRAPH_NAME_FILTER=*
> TF_DUMP_GRAPH_GROUPS=dtensor.mlir,dte... | Yu Feng | feyu@google.com | 1,693,442,177,000 |
88ba2208aff790f5aa32f5c779ec32b0511b29e7 | Support for select in generic pattern matching for FP8 GEMMs. | Philipp Hack | phack@nvidia.com | 1,693,443,915,000 |
bde78038df9a2320e8954a4a194c53d298d09594 | Integrate LLVM at llvm/llvm-project@92023b150990
Updates LLVM usage to match
[92023b150990](https://github.com/llvm/llvm-project/commit/92023b150990)
PiperOrigin-RevId: 561508721 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,445,743,000 |
b1b51da6500abf3d6e404a9cae35e194aa815f59 | [xla:gpu] Disable multistreaming on non-Ampere architectures
According to Nvidia Ampere architecture whitepaper, there is performance regression from repeated-fork-join-style topology in CUDA graphs in non-Ampere architecture.
Source: Section "Task Graph Acceleration on NVIDIA Ampere Architecture GPUs" in https://ima... | Anlun Xu | anlunx@google.com | 1,693,450,255,000 |
48e400d57f48b96c069fbf1d3adcec97c34e140b | #tf.data Log `tf.data.Options` used to ease debugging.
PiperOrigin-RevId: 561527499 | Wilsin Gosti | wilsin@google.com | 1,693,452,786,000 |
168320e0b525e0e616bd4164bef9b5211f87a54c | Merge pull request #61636 from christian-steinmeyer:christian-steinmeyer-patch-1
PiperOrigin-RevId: 561565011 | TensorFlower Gardener | gardener@tensorflow.org | 1,693,465,161,000 |
39446a83336f1509478af789a4751d0b2bab0ca4 | Merge pull request #61752 from linux-on-ibm-z:update_rules_python
PiperOrigin-RevId: 561567003 | TensorFlower Gardener | gardener@tensorflow.org | 1,693,465,692,000 |
daaee3d19d0aebc34f8d5bff614e71041e0a5a4e | Internal Code Change
PiperOrigin-RevId: 561570826 | Richard Levasseur | rlevasseur@google.com | 1,693,466,378,000 |
70709bf587fd4fad6fe2fb6aaedbaa7dfcaceae7 | Internal Code Change
PiperOrigin-RevId: 561571359 | Richard Levasseur | rlevasseur@google.com | 1,693,466,555,000 |
3ab35edfc8cf01b38ad3b5e3fbc101960ae6d7b2 | Internal Code Change
PiperOrigin-RevId: 561571366 | Richard Levasseur | rlevasseur@google.com | 1,693,466,557,000 |
af1b00d71dd2c1dd0223d2fbc194ae8ca2ac24c2 | Add PropagateQuantizeTypePass
Propagation of dequantize op is added in PropagateQuantizeTypePass.
To allow flexible shapes, tensor shapes in the dequantize function is
changed to unranked shape. Gather op quantization is added in the
QuantizeWeightsPass to meet the feature parity with the existing weight-
only quantiz... | Songyi Han | syhan@google.com | 1,693,467,260,000 |
95b325db3a643f0599084eb375d4c99f7f8d4d35 | Implements CalibrationAlgorithm class to calculate min and max values.
PiperOrigin-RevId: 561584692 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,470,854,000 |
a0a4937afb4b4a6f710473d7bac0fb92c03c419f | compat: Update forward compatibility horizon to 2023-08-31
PiperOrigin-RevId: 561589874 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,472,517,000 |
602463ba2751825d739793cd14781357e9ccc875 | Update GraphDef version to 1605.
PiperOrigin-RevId: 561589924 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,472,531,000 |
ba3f8c8e82aa626ce85395afa9ee7d42cd9a4ddf | Internal Code Change
PiperOrigin-RevId: 561600015 | Richard Levasseur | rlevasseur@google.com | 1,693,475,881,000 |
45eb60b6e779d9e30b9b89d863d35bb273a63ef4 | Add test cases for vectorization/unrolling of column reductions.
Specifically, the parts of the heuristic that depend on graph traversal.
PiperOrigin-RevId: 561603191 | Johannes Reifferscheid | jreiffers@google.com | 1,693,476,908,000 |
2f3976722f44d614018c394b49d0c71f2d6b23fc | Integrate LLVM at llvm/llvm-project@82e851a407c5
Updates LLVM usage to match
[82e851a407c5](https://github.com/llvm/llvm-project/commit/82e851a407c5)
PiperOrigin-RevId: 561604863 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,477,485,000 |
e16b168f375ac8356ec62e1f95e2ac3f68f9cdf8 | [XLA:GPU][NFC] Split up Triton Softmax rewriter logic to enable reuse.
This is in preparation to landing autotuning-driven fusion choices for Softmax.
We expose the newly created utils in order to cleanly separate matcher tests
from autotuning-related tests in the upcoming changes.
PiperOrigin-RevId: 561606077 | Benjamin Chetioui | bchetioui@google.com | 1,693,477,958,000 |
a3f564ce5bec36c41bc26177133b78e9124177d1 | Split CalibrationStatisticsCollector to small collectors.
PiperOrigin-RevId: 561614946 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,480,990,000 |
414b887af47e4793d0efa1a8aae45565255dffff | Do not depend on fusion->fused_parameters in HLO fusion analysis.
This should be the last step before we can pass the fusion roots and a boundary
function to the analysis creation function.
PiperOrigin-RevId: 561620111 | Johannes Reifferscheid | jreiffers@google.com | 1,693,482,882,000 |
4c8d1c1969c79689aa7310ea6c9a7e2e32ae872b | [XLA:GPU][NFC] Minor stylistic fixes to Triton Softmax rewriter.
PiperOrigin-RevId: 561632169 | Benjamin Chetioui | bchetioui@google.com | 1,693,487,031,000 |
c55a9012e440673edaea1241c85c05af78d5a1e5 | Ensure snapshot files are finished writing to fs
Snapshot files may take a little time to all be written out to
the filesystem on a heavily loaded system so need to allow for
this before asserting that they are all there. | Andrew Goodbody | andrew.goodbody@linaro.org | 1,693,487,556,000 |
89289d1d5868f3a52bc0423f1801061f94867fcc | Add missing headers to fix build.
PiperOrigin-RevId: 561667701 | Weiyi Wang | weiyiw@google.com | 1,693,497,277,000 |
4f1f0cf693f1ed70c31cbd3ed7d5e8932b9d8612 | [tflite] Replace absl::string_view constant with char[]
PiperOrigin-RevId: 561672211 | Majid Dadashi | majiddadashi@google.com | 1,693,498,268,000 |
c05758079f9607fbde4351b9056cf88f66ad8b82 | [XLA:GPU] Create cost analysis inside priority fusion queue.
Right now we use std::optional to reset cost analysis for each run of the pass. In reality we don't even need to share the same cost analysis between two computation in one pass run. Reusing one instance of cost analysis only unnecessarily grows hash maps.
... | Oleg Shyshkov | shyshkov@google.com | 1,693,499,096,000 |
dac356fe24de28260750441a2b072dbbc458fb10 | Fix for lint and addressing comments. | Kulin Seth | kulin_seth@apple.com | 1,693,501,068,000 |
0fb5b606f3657a86a4acb3f326f79de544d3ee74 | Adds an optional 'overbudget coefficient' to penalize solutions that eclipse the memory budget.
PiperOrigin-RevId: 561690874 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,502,334,000 |
7b3b60cd73f10899e5980c14cd63616bdc5c5ef2 | Support for select in generic pattern matching for FP8 GEMMs. | Philipp Hack | phack@nvidia.com | 1,693,502,629,000 |
f419316ea948e2f6922a8a0146603c6ef65baabb | Handle degenerate shapes for Broadcast.
PiperOrigin-RevId: 561693715 | Jared Junyoung Lim | jaredlim@google.com | 1,693,502,895,000 |
2de6822138d78c769ecb2b594ab5b3aab302ab97 | Preparation to move Servo TFRT integration code out of google folder to oss
PiperOrigin-RevId: 561694550 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,503,080,000 |
7547372f6157c5dcf023d377cc6534b13f0fc135 | [xla:gpu] Cache all memref.view operations with same physical shape
This will generate smaller IR with fewer memref.view operations.
PiperOrigin-RevId: 561700272 | Eugene Zhulenev | ezhulenev@google.com | 1,693,504,231,000 |
7884cf3a95170d418bfe351c0c30dcf5d277261e | Make default AtomicFunction call structured
PiperOrigin-RevId: 561702586 | Faizan Muhammad | fmuham@google.com | 1,693,504,690,000 |
e77390da4c276287fdf653b94527ddf5e58bfbea | [tflite] Create Eager API for square
PiperOrigin-RevId: 561705833 | Grant Jensen | grantjensen@google.com | 1,693,505,305,000 |
2289daf62fa657f080c4e3dd0b8c3b4fada52d64 | ## Wheel Test
This section is dedicated to tests that require a built TensorFlow wheel file,
and that's why this is not a part of TensorFlow package.
### test_import_api_packages
This Python test verifies whether the API v2 packages can be imported from the
current build. It utilizes the `_api/v2/api_packages.txt` l... | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,505,552,000 |
253ac98d39c628734b9021fcefba15a62e8c5446 | Add XlaCallModule test for multi-platform lowering.
Specifically, test the case when the code for one platform contains a custom call whose target is unknown to other platforms.
PiperOrigin-RevId: 561707531 | George Necula | necula@google.com | 1,693,505,629,000 |
6480d941710c7c10f2b8048736e4201d6d95d36c | Fix return type of DeviceTopology.serialize (Py bind).
The Py bind serialize function return type is expected to be
bytes while the C++ PjRtTopologyDescription::Serialize function
returns a string. Apply py::bytes in the Py binding.
PiperOrigin-RevId: 561715539 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,507,293,000 |
ae366f7245ba1e4b66338edf5a06807d7436bccf | [SPMD] Support shard like interface in XLA and JAX (0/N).
Define sharding annotation with key words
1. shard_as <num>
2. shard_like <num>
to group shardings.
PiperOrigin-RevId: 561733657 | Tongfei Guo | tongfei@google.com | 1,693,511,382,000 |
6bffcecb465773ae0b00c99cff6e73a3d8863197 | Merge pull request #61758 from philipphack:u_fp8_select_xla
PiperOrigin-RevId: 561735390 | TensorFlower Gardener | gardener@tensorflow.org | 1,693,511,992,000 |
f99d884e55acafb71102a266b362d6e1d4d52dde | Integrate LLVM at llvm/llvm-project@c7dce0283b5b
Updates LLVM usage to match
[c7dce0283b5b](https://github.com/llvm/llvm-project/commit/c7dce0283b5b)
PiperOrigin-RevId: 561736250 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,511,974,000 |
0788925f3ae2af11c8bdbaf183774b9c53fcb6e6 | Test more edge cases for f32-scatter.
We recently added a "special" lowering for f32-scatter-max. This PR adds some
additional tests to make sure that we got the edge cases correct.
In order to write this test, we have to make two additional changes.
1. Fix PjRt to forward the XLA debug options from the HLO module,... | Justin Lebar | jlebar@google.com | 1,693,512,187,000 |
b6e86679444bdec838911fa7fbf4f40db5f92aa8 | Remove unused Platform::RegisterTraceListener and ::UnregisterTraceListener.
PiperOrigin-RevId: 561737623 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,512,246,000 |
964eb83814e7cb840556da672ca1404c1c255bdb | Add type hints for ops.Operation.
PiperOrigin-RevId: 561741675 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,513,132,000 |
429b183ffd71f8ae2623f81455e16961a27de8e3 | [SPMD] Prevent DCE on sharding custom-call to allow a dangling sharding custom-call attached to an op.(1/N)
1. The removal of sharding custom-call will be handled in sharding propagation preprocess where "Sharding" custom-calls are converted to copies and DCE-ed after.
2. By allowing a dangling sharding custom-call, t... | Tongfei Guo | tongfei@google.com | 1,693,515,131,000 |
1f569eaee963298b3958c2ad2e061d6dc46fd62a | Change numpy version in requirements.in from 1.23.6 to back to 1.23.5.
PiperOrigin-RevId: 561751813 | Camilla Halsey | halseycamilla@google.com | 1,693,515,271,000 |
9f7df82b4e4ba5c8ca2bf58bee3537602f9cd56d | Create a bridge target package in preparation to make it an official API
PiperOrigin-RevId: 561751918 | Mason Chang | masonchang@google.com | 1,693,515,294,000 |
967e0c170319bd01048458ed5419d79108ec8ffa | Align reshape and squeeze integer array types with flatbuffer data type
The integer-based arrays in the TFLite flatbuffer schema for ReshapeOptions and SqueezeOptions result in a generated `flatbuffers::Vector<int32_t>` for C++. Their BuiltInOpData counterparts of `TfLiteReshapeParams` and `TfLiteSqueezeParams` have `... | RJ Ascani | rjascani@google.com | 1,693,515,853,000 |
74961f378b578705859471276eebd999738c7fee | Speed up StochasticCastOpTest
Reduce the number of values sampled.
PiperOrigin-RevId: 561760868 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,517,128,000 |
5806388e36466692ad1074247be9583f0c2b121e | Add GpuRunner to compile and execute a program.
PiperOrigin-RevId: 561760953 | Changhui Lin | changhuilin@google.com | 1,693,517,149,000 |
ea85f5df0f0d112b397f5014b4cc463afa1f06cc | Implementation for RGB to grayscale + tests.
PiperOrigin-RevId: 561764764 | A. Unique TensorFlower | gardener@tensorflow.org | 1,693,518,010,000 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.