PiperOrigin-RevId: 340033160
Change-Id: Ia7007ddb1bcd128c684712eafa284af79fedac05

PiperOrigin-RevId: 340026157
Change-Id: Ibf94e17f0cd2cf88738d1d08be06cbe68a848ecf

tf.parallel_stack is implemented as a graph rewrite and therefore has no support in eager mode. Added documentation to parallel_stack and raising an error now to reflect that.

PiperOrigin-RevId: 340016749
Change-Id: I7c695dfb98549bbc56105592b1b1c6b1ea06b231

PiperOrigin-RevId: 340005180
Change-Id: Id2be30ca6dc141ef6c51113e0ef915d03568ba3b

PiperOrigin-RevId: 340005178
Change-Id: Ic63905f93733940c57d64156c502b95f067bba4f

PiperOrigin-RevId: 339989040
Change-Id: I46eaffb4e4b28c168b5f0182c4c0bb946d6eefe8

PiperOrigin-RevId: 339983364
Change-Id: I877f2394f13b899ace0bb2893e6cb5f073b03458

Created a CMakeLists.txt under the lite/examples/minimal

Usage)

$ git clone https://github.com/tensorflow/tensorflow.git tensorflow_src
$ mkdir minimal_build
$ cd miniaml_build
$ cmake ../tensorflow_src/tensorflow/lite/examples/minimal
$ cmake --build . -j

PiperOrigin-RevId: 339977906
Change-Id: I85cada5430407e72354d0a31a17584c48407db2f

When processing buffers that are placed in tuples, we create (possibly
redundant) additional tuple and gte instructions, which we later simplify. We
don't need to create these additional instructions for buffers that were
unchanged after memory space assignment. We were facing compile time and OOM
issues without this change.

PiperOrigin-RevId: 339972236
Change-Id: Ieb82767ed52d39260141f3e3179387fbf5783920

ConvConstants extended for better mapping on thin dst tensors.

PiperOrigin-RevId: 339971199
Change-Id: Ia14be12361ec81c70c0522baa9516c8cb28ba7ff

ConvolutionTransposed3x3 generation changed to use storage type properties instead of specific storage types.

PiperOrigin-RevId: 339967056
Change-Id: I5f6e192e21cbabf33bd152f91f4ab664b139fd14

Multi-worker tutorial: Add the workflow of MWMS+CTL example that is going to be added in the tutorial in multi_worker_tutorial_test.

Fix the flakiness of the test and re-enable in TAP.

PiperOrigin-RevId: 339966024
Change-Id: Icb866f8a7054fa88f2e474c02960982a57c542b3

nest.flatten was mangling some attribute lists

Fixes #40895.

PiperOrigin-RevId: 339965951
Change-Id: Ibdbbb5a12b2d06a077b62902764ca4849a2b8d2a

DepthwiseConvolution generation changed to use storage type properties instead of specific storage types.

PiperOrigin-RevId: 339964553
Change-Id: I2cded9c306a40b136002c08e610daca2d75e1758

Added A14 GPU to enum.

PiperOrigin-RevId: 339964411
Change-Id: I13439c83c6f3d0f75a166b950a5f088843dd2023

PiperOrigin-RevId: 339959753
Change-Id: I492d1f750e7a6fd97d4fdccbb838492e699c56b6

PiperOrigin-RevId: 339956677
Change-Id: Id4cea5f63b3f257170c121d74e9cb2995a10b629

PiperOrigin-RevId: 339954904
Change-Id: Id9f6717da5c32bff185a10a37e9682be64cc6501

PiperOrigin-RevId: 339953940
Change-Id: I568589fc09fa6e5aae137e7e8533e1dab86190d3

PiperOrigin-RevId: 339951415
Change-Id: Ic3bbb2e474a4fb287fb48f24509beb44625ee2e4

PiperOrigin-RevId: 339951343
Change-Id: I02b02143721f7c656af8c1d19176dc44f3bffa20

PiperOrigin-RevId: 339948109
Change-Id: Ia2612efa66b48ac5784d2a53cdeb0ce9001f4025

PiperOrigin-RevId: 339946530
Change-Id: I9ff4d4a380b6f2c2181edfc76cbb41af09abf0e2

PiperOrigin-RevId: 339944681
Change-Id: Ifd0f43a2616328fe54819fae18ee626a44556e99

Dataset cardinality op should be placed on the same host device as the dataset, otherwise TF will try to encode/decode the dataset to another host device, which is not possible.

PiperOrigin-RevId: 339942622
Change-Id: Ifbeb0aabdfc35f6c0e1248d69a8d22727e19c4a2

This further simplifies ShapedBuffer object as itself doesn't have any
logic to use the platform field.

Notice that this cl also removed some sanity check in allocation
tracker. we can add that sanity check back if need --- just keep track
of `platform` inside of allocation tracker as a side map.

PiperOrigin-RevId: 339938197
Change-Id: I090e603927ed3fccdb51254f972b3af2e1ec1470

PiperOrigin-RevId: 339937575
Change-Id: I1d80b244adcb27a429bc09c63434847107d23bee

PiperOrigin-RevId: 339929158
Change-Id: I8592aa6e2cec32a2ba97743a6f022f263f0f65e2

TFRT batch fallback kernel to crash in high QPS load. The bug makes it only
create an object of base class BatchTask even when splitting a task of the
derived class FallbackBatchTask (used only in TFRT), and put it into a batch
with other FallbackBatchTask objects. When this batch of mixed types of tasks is
processed, it crashes.

PiperOrigin-RevId: 339927837
Change-Id: Ie52bd11c61c9ddbe6ab803cd90208419d4b2dba6

PiperOrigin-RevId: 339922525
Change-Id: I88c0d616154484ddd353d262f2fe7408cfc058c4

#KERAS_PRIVATE_API_CLEANUP

PiperOrigin-RevId: 339917969
Change-Id: Iadd13a8d23a941528e0384e090dc07ddc9d63da6

With this change, it is now possible to mix-and-match tf.keras.Layers and
tf.Modules inside a tf.keras.Model and everything will be tracked properly.

- Variables in tf.Modules that are set as attributes of custom Layers and
  Models now show up properly in properties such as Layer.trainable_variables
  and Model.trainable_variables.
- tf.Modules do not show up in Model.layers. Instead, a new method
  Layer._flatten_modules is added that iterates over tf.Modules and Layers in
  the order that Keras expects. The existing method Layer.submodules (inherited
  from tf.Module) can still be used to iterate over tf.Modules and Layer with the
  tf.Module ordering. Layer._flatten_layers is built on top of
  Layer._flatten_modules
- Layer._layers is renamed to Layer._self_tracked_trackables to avoid naming
  conflicts with user-defined attributes (and to reflect that this attr
  contains Layers, Modules, and TrackableDataStructures)
- A new property is added to tf.Module to enable this, namely
  tf.Module.non_trainable_variables

PiperOrigin-RevId: 339917644
Change-Id: I96a7302745280a6261de8c4295c5cbf5f4d7dd5c

PiperOrigin-RevId: 339915378
Change-Id: I5675fdf910a3d8fd550d28a5382f0da20e450281

PiperOrigin-RevId: 339913277
Change-Id: If4a92b3fafe922b5abf00bc8dc04deef1ef6ca6d

PiperOrigin-RevId: 339911119
Change-Id: Iddf765ff3e58ea67c0cdf3b1ffe3065a92e8dcfc

This change avoids an implicit conversion from enum BuiltinOperator to int8_t
for calls to CreateOperatorCodeDirect. This should mean that once the
BuiltinOperator enum is larger than a byte, the TFLM code does not get
tripped.

PiperOrigin-RevId: 339910848
Change-Id: Ief978f2bb9ea818f8211ade3927baa35dc7a8512

Add property that allows layers to specify that the input_spec can also be used as the layer call function's input_signature.

By default,all Keras exported layers will have this property return True, since Keras more rigidly defines the input_spec shape (compared to user-defined models, which may only have `ndims` set).

PiperOrigin-RevId: 339909813
Change-Id: I1b486747aa1e413e6f24f6809fb88846ad4712ab

PiperOrigin-RevId: 339909803
Change-Id: I7cbcc1d74c0de2cd7f86e10e2b44f0d4b81274a0

PiperOrigin-RevId: 339909494
Change-Id: I1a9352342e9b79b55b5ba1f324fecf16a34b7647

PiperOrigin-RevId: 339906858
Change-Id: I0ad0e7fef95d634a75ddb22a5b08dee9b66c2ec1