tensorflow / tensorflow

…loading resources that do not override _gather_saveables_for_checkpoint. This allows us to support simple resources like StaticHashTable. Note that additional logic related to checkpoints will need to be added to support df6b21c.

PiperOrigin-RevId: 333221978
Change-Id: Ib724b6cee9a57bf1b3ee98d2a8eaf9f394a8d64d

… can rewrite ModelDataset.

PiperOrigin-RevId: 333211278
Change-Id: Ic83a8f2295fab3ea17aaf563ca022ad4d5c62646

PiperOrigin-RevId: 333208193
Change-Id: Ic9adc699a11ffcc23a0ae518b54ee29cce8569ce

PiperOrigin-RevId: 333207731
Change-Id: Ia8d9dcda8e0ff28f3e23b8562e07e60468a032ac

PiperOrigin-RevId: 333206274
Change-Id: I9c041c707986bc64a696e4fc4d9a43771c249388

…late

PiperOrigin-RevId: 333206121
Change-Id: I28e05304b0b1daf2445b5fb1a180e6b7d06f4a14

This change also moves TF2 tests to kernel_tests/. The plan to keep ops/collective_ops_test TF1, and add/move TF2 tests to the one under kernel_tests

PiperOrigin-RevId: 333205878
Change-Id: Ifc3681c63cf7908bed5a8bc2517b6ce1cc3746cb

…_level is 0, i.e. disabled.

PiperOrigin-RevId: 333205135
Change-Id: Ibf40852b9afe8ce532528b14cbdfc3d060d29a04

PiperOrigin-RevId: 333204746
Change-Id: Icef8422fda806d798b95839faa759bb7cec1e54b

…ocessing mode.

PiperOrigin-RevId: 333201582
Change-Id: I43477fc620ae427e9c187063836dc8dba28dd183

…cases

We do this several other places for tf.function gradients. Currently we can't manage them in C++.

Fixes #43449.

PiperOrigin-RevId: 333200606
Change-Id: Ibd941f1bb6fc39c778037e79a29205697f3fe6fe

PiperOrigin-RevId: 333195785
Change-Id: Ie51da1485a45946b5bdd404ce7e9cc07d34e310d

GitHub does not insert automatic links and smart code snippets in these files, so we have to do it manually.

PiperOrigin-RevId: 333195707
Change-Id: I1e2fed8ff207fbfce6eb8fb2b910d12bcab4100c

Helps with TSAN timeouts.

PiperOrigin-RevId: 333195144
Change-Id: Iceb62706c9664ef14cb78f59919631af81e30389

While at it, added changed `SP_Device` argument to host_callback to be const.

PiperOrigin-RevId: 333195011
Change-Id: I5224812674c00332336a5c0fcf68ed28dd33d0aa

PiperOrigin-RevId: 333192867
Change-Id: Ib5670619b681080c08057c037858228b107bae33

We need to repack the return of the if/else branches to ensure any
CompositeTensors are returned with the correct type. For functional
while we need to be sure to unpack and repack when handling the body.

PiperOrigin-RevId: 333190256
Change-Id: If9a9449a10616afa3dd79e39b1e66ee0ee571a9e

…flu_softnax_int16_ref

PiperOrigin-RevId: 333185684
Change-Id: I4d8a2845286ba905b16cebda204ef006c1a8e535

…e it tries to save EagerTensors in a TF 1.x tf.Graph (used in TF 1.x style serving APIs).

The error message was:
"RuntimeError: Attempting to capture an EagerTensor without building a function."

PiperOrigin-RevId: 333185295
Change-Id: I05b65d2bc8c446032300faf31696316898b45ef4

PiperOrigin-RevId: 333177675
Change-Id: I15eb3f422f1c6139741d681223710c5cc741a8f0

Also, move the tests to experimental/kernel_tests where they belong.

PiperOrigin-RevId: 333177431
Change-Id: I04f415c08268c0f9113d9bbb342bd00bdf911956

PiperOrigin-RevId: 333177305
Change-Id: I20e789cafd03e01ddf34759f436cf1cf4d39ccac

PiperOrigin-RevId: 333174134
Change-Id: I7d8b5cb91603ab77a85cf201a0df860fd404c5ce

…shards when trying to auto shard by FILE.

Also improve documentation in the input guide along those lines.

PiperOrigin-RevId: 333172619
Change-Id: I4cfe88ac7a78991d6b768a9d3403864c3378ae59

PiperOrigin-RevId: 333170739
Change-Id: I3511b3184eb93abc4d10b7688c8be96db6b7a94c

We need to repack the return of the if/else branches to ensure any
CompositeTensors are returned with the correct type. For functional
while we need to be sure to unpack and repack when handling the body.

PiperOrigin-RevId: 333169759
Change-Id: I45cd86467dba58f28d8fd7abd2faf2b6700132ec

Updates LLVM usage to match
[f21212215031](llvm/llvm-project@f212122)

PiperOrigin-RevId: 333168406
Change-Id: Ifd4da56082b2c8ab9f253e7d16969c64c9f9f407

- Use AbstractContext to create Tensor to support creating Tensor when turning TFRT on.

PiperOrigin-RevId: 333159868
Change-Id: Ic0092fe20190bc1003d4827e59d6cd0e3626237f

Only attempt legalization of tf.Min, tf.Max, tf.Prod, and other reduction ops
for int, float, and complex element types, to avoid crashing in the case of
unsupported types such as quantized integer types.

PiperOrigin-RevId: 333159797
Change-Id: I48f58c7f7e3dab196d2a2a1e728531a65fbf5872

…ffer_sizes`, which will inject autotuned PrefetchDatasets after non-prefetched asynchronous datasets. The optimization will also rewrite those existing non-autotuned PrefetchDatasets into autotuned with fixed start value and minimal value for the tunable parameter `buffer_size`.

PiperOrigin-RevId: 333155780
Change-Id: I5bdd061100ed7bd564f5ad6f7d9cc8c9a723a9ed

PiperOrigin-RevId: 333154326
Change-Id: I83ed06d13d1c223cc6d72ca27bb8b4279a14669a

…tal.h.

This provides adds a new callback-based op registration C API.
This new function would be required in order to implement the TF Tasks Library, or any other C++ API that uses OpResolver parameters, using the TF Lite C API.

PiperOrigin-RevId: 333154110
Change-Id: I6e53edc91ff1271f41c355bb9ad1a48d32bf3e41

… TensorHandle captures associated with it. This helps further decouple Function lifetimes from other objects in the SavedModel. (Eg: Imagine a case where a RestoredResource must invoke a "destroy_resource" function in its destructor, and where a destroy_resource function may capture other state. This allows orderly deconstruction.)

PiperOrigin-RevId: 333153620
Change-Id: Ibd4dcee6d40eebb43c6c6f95283ac242c38c04e8

…conflict.

PiperOrigin-RevId: 333151375
Change-Id: I84a4370ed1e0e0c25ff9348679b2462acf295fb0

… value, instead of the SaveableObject. This helps pave the way for future changes where the SavedModel class stores all revived objects in a per-type map instead of a single erased map.

PiperOrigin-RevId: 333144175
Change-Id: I3e4727e6bf5610772a8b4b1414b1e0c39acf88a7