Add workaround for the latest toolchain repository not supporting older bazel
versions; only load it conditionally.

PiperOrigin-RevId: 224965872

PiperOrigin-RevId: 224956744

Rewrite Adam and LazyAdam optimizer to take global step for computing beta1 and beta2 accumulators, instead of having the optimizer instance to keep its own independent beta1 and beta2 accumulators as non-slot variables.

PiperOrigin-RevId: 224948020

PiperOrigin-RevId: 224937131

PiperOrigin-RevId: 224936924

PiperOrigin-RevId: 224933587

in the specific case of:
* Eager execution enabled
* Inside a FuncGraph, inside a graph
* In a replica context (such as in a call to
  `tf.distribute.Strategy.call_for_each_replica()`).

PiperOrigin-RevId: 224930182

Previously we only had a flag for disabling specific passes.  But being able to
disable all passes is helpful if you have some already-optimized HLO that you
just want to run.

PiperOrigin-RevId: 224928095

When you pass XLA_FLAGS to replay_computation, you very likely want that only
to apply to the actual computation(s) being run, not to the XLA computations
that replay_computation synthesizes to generate fake data for the "real" ones'
arguments.

PiperOrigin-RevId: 224927003

This gets a DebugOptions struct with all the defaults filled in as though
XLA_FLAGS were empty.  This is useful when you want to run an XLA computation
and explicitly ignore any XLA_FLAGS passed to the binary.

PiperOrigin-RevId: 224925335

Expose a `tensorflow.autograph` namespace, with a minimal core API under experimental. Clean the documentation for public symbols.

PiperOrigin-RevId: 224921147

Enable mixing value tensors (eager tensors or numpy arrays) and Keras symbolic tensors when building Keras graphs-of-layers in an eager scope.
In these cases, the value tensors are treated as symbolic constants.

This enables the following pattern to work in the same way in both V1 and V2:

```
lstm = LSTM(2)
inputs = keras.Input((None, 3))
outputs = lstm(inputs, initial_state=tf.ones(shape))
```

(without this change, the above code works in V1 but fails in V2 with an artificial exception).

Known issue: in case a random tensor is used, there is a (usually harmless) behavior discrepancy remaining between V1 and V2, which is that in V2 we'd be using the same random value every time, whereas in V1 we'd be drawing new random values (since the tensor would be treated as a random op and not as a constant). We think this is not a problem because in V2 users should have the mental model "tensors are values" and thus would be expecting a random tensor to behave like a constant value and not like a random generator.

PiperOrigin-RevId: 224915621

This bypasses a nullptr error that appears downstream due to referencing a length-zero array in a temporary buffer.

PiperOrigin-RevId: 224915050

PiperOrigin-RevId: 224914276

PiperOrigin-RevId: 224913339

This disables the keepalive watchdog for TF/GRPC channels.  The watchdog ping timer is intended to monitor channels in case they have gone "stale".  If this occurs, any pending RPCs are marked failed.  This interacts poorly with large TF models, where we can saturate the network exchanging tensors, causing the watchdog ping to be delayed.

The timer is not essential (normal deadline processing and socket termination is still respected), so we can disable it with minimal risk here.

PiperOrigin-RevId: 224913045

it is a bash test, and language filters do not work properly on windows.

PiperOrigin-RevId: 224912071

PiperOrigin-RevId: 224905468

Also stop exporting CuDNNLSTM since its all covered by unified LSTM.

PiperOrigin-RevId: 224900214

Returning void is more common than returning Status so pick the longer name for
the less common variant.

PiperOrigin-RevId: 224897169

PiperOrigin-RevId: 224897071

PiperOrigin-RevId: 224894987

PiperOrigin-RevId: 224894043

PiperOrigin-RevId: 224893836

PiperOrigin-RevId: 224891138

`keras.backend.function`.

PiperOrigin-RevId: 224886577

PiperOrigin-RevId: 224877586

Allow completely stateless(i.e., with no outputs) loops. Simplify the handling of stateless conditionals.
This change will still not support stateless loops pre-v2 until we add auto deps. However, it works properly in tf.function.

PiperOrigin-RevId: 224876064

PiperOrigin-RevId: 224875931

PiperOrigin-RevId: 224874845

PiperOrigin-RevId: 224872228

PiperOrigin-RevId: 224870669

This fixes a bug where Adam beta*_power variables were always created as RefVars
even if the optimizer acts on ResourceVars.  This broke certain defun + Adam
use cases.

Also fixed the unit tests, which *always* created ResourceVariables
(ever since variables.Variable() constructor became aliased to ResourceVariables).

PiperOrigin-RevId: 224869338

PiperOrigin-RevId: 224869113

PiperOrigin-RevId: 224865488

PiperOrigin-RevId: 224863771

I believe for historical reason, the activation function for LSTM is hard_sigmoid because it is faster compare to sigmoid. With the new LSTM, the performance issue should be fixed with grappler swapping the backend.

PiperOrigin-RevId: 224863406

The new toggle replaces ENABLE_COND_V2, ENABLE_WHILE_V2, and
ENABLE_TENSOR_ARRAY_V2. This means that these can't be toggled
independently anymore, notably that v1 TensorArrays can only be run
with v1 loops, and v2 TensorArrays with v2 loops.

This also introduces a corresponding environment variable
TF_ENABLE_CONTROL_FLOW_V2. I kept the old env vars as well in case
people are using them. They all flip the new single toggle now.

In addition, this change removes some while_v2 code for dealing with
v1 TensorArrays, since this is no longer a supported configuration.

PiperOrigin-RevId: 224862245

Add attribute to Keras model which generates an exportable tf.function. SaveModel save now looks for this attribute when searching for a function to export.

PiperOrigin-RevId: 224861089