Renames to prepare for the TF2.0 API. HashTable -> StaticHashTable and MutableDenseHashTable -> DenseHashTable. Also insert -> insert_or_assign and remove -> erase for DenseHashTable.

PiperOrigin-RevId: 235206233

PiperOrigin-RevId: 235201217

PiperOrigin-RevId: 235197577

PiperOrigin-RevId: 235197341

PiperOrigin-RevId: 235192783

PiperOrigin-RevId: 235184105

This function previously stated that it could accept a rank-0 Tensor as its `num_or_size_splits` argument. This is in fact not correct: TF graphs must have a static number of output tensors for any operation, so the number of splits has to be known statically or we cannot validate the correctness of the graph.

This method will now raise a ValueError if called with a rank-0 Tensor. I also make the documentation more explicit. Strangely, the test was testing for the incorrect behaviour, so I also fixed that.

PiperOrigin-RevId: 235157900

PiperOrigin-RevId: 235149900

PiperOrigin-RevId: 235144234

PiperOrigin-RevId: 235138912

Previously, when in a GPU device context, the logic attempted to create the default empty metadata tensor on the GPU, which fails since it's a string tensor.

PiperOrigin-RevId: 235135723

Add implementations for CUDA library functions which dynamically load the CUDA library DSO and forward the call to the actual function. Most of the code is auto-generated.

Libraries implemented: CUDA driver API, cuRAND, cuFFT, cuBLAS, and cuDNN. The CUDA runtime is not implemented because it's not clear whether it should just be linked statically.

Older versions of CUDA (starting with 8.0) and cuDNN (starting with 7.0) are handled with two approaches:
- if the API didn't change and only new functions were added, declare types that were introduced afterwards. It is acceptable to implement APIs that were only introduced later: if they are called despite not being declared (in older headers), they will not be able to dlsym the corresponding function and return an error.
- if the API changed, different contextual headers are included.

PiperOrigin-RevId: 235135165

Replace angle bracket CUDA kernel launches ('<<<>>>') with API function calls ('cudaLaunchKernel'), through a helper function which also checks the return code in debug builds.

PiperOrigin-RevId: 235135115

PiperOrigin-RevId: 235128104

Handle the case when tf.function's target is a tf_decorator method that was hot-swapped after tf.function was applied.

PiperOrigin-RevId: 235124593

Address TODO to support materialization of Fill nodes that are fed from constants stored in raw serialized format.

PiperOrigin-RevId: 235123971

maximum_iterations no longer needs to be a TF-graph-compile-time constant but should be a compile-time constant when building the XLA graph (when values of placeholders are also available).

There are broadly 2 parts of this change.
1. while_v2: This removes the _maximum_iterations attr from the functional While op generated by while_v2 and makes it an input instead. This frees maximum_iterations from being a graph-building time constant. Its index in the list of op inputs is preserved in both the forward and backwards graph so it should be ok for n'th order derivatives.

2. The XLA changes are needed to propagate constants inside the loop body e.g. when building the gradient of a While inside a while_loop. They are 2-fold:
a. This updates the const_analysis pass to inspect the cond and body functions of While for compile time constant requirements.
b. When compiling the XLA While op in while_op.cc if there are inputs that are compile time constants and the corresponding loop variables are loop invariants, the constants get propagated into the loop body.

Makes XlaCompiler::FindFunctionBody public to allow accessing the body of the While in while_op.cc. If the body has been specialized using the PropagateConstIntoFunctionalNodes pass the rewritten function only exists in XlaCompiler::local_flib_def_.

PiperOrigin-RevId: 235122516

it doesn't take a table as input and takes an initializer. Also oov_buckets
should always be > 0 for this table.

PiperOrigin-RevId: 235119219

PiperOrigin-RevId: 235117320

PiperOrigin-RevId: 235114715

PiperOrigin-RevId: 235112772

resources will now be passed as 'data' instead of on the other attributes.
END_PUBLIC

For more information on the resources cleanup and rules, see:
  go/apple-rules:resources

BEGIN_PUBLIC
The "resources", "asset_catalogs", "bundles", "datamodels", "storyboards",
"strings", and "xibs" attributes are directly moved to the "data" attribute.

Any "structured_resources" are moved to a new apple_resource_group instance,
and that is added to "data".

PiperOrigin-RevId: 235112154

Change the Keras model_iteration() to correctly infer the size of validation datasets when validation_steps is None.

PiperOrigin-RevId: 235112144

PiperOrigin-RevId: 235107020

PiperOrigin-RevId: 235106933

Pow(x, y) acts like std::pow which complicates trying to use Pow(x, 0.5) to implement sqrt.

Pow(-0, 0.5) is +0 but Sqrt(-0) should be -0.

As such, it makes sense to provide an explicit Sqrt and Rsqrt.

PiperOrigin-RevId: 235105846

PiperOrigin-RevId: 235104911

PiperOrigin-RevId: 235102987

Remove per-op code in hlo_verifier and hlo_instruction.cc that checks instruction arities for non-variadic ops, use generic code instead.

Refactoring only, NFC intended. This is a small step towards reducing the amount of boilerplate needed to define a new op.

PiperOrigin-RevId: 235100117

Fixes test.

PiperOrigin-RevId: 235096096

PiperOrigin-RevId: 235089008

PiperOrigin-RevId: 235087375

PiperOrigin-RevId: 235086165

PiperOrigin-RevId: 235085572

PiperOrigin-RevId: 235085131

PiperOrigin-RevId: 235084042

PiperOrigin-RevId: 235082791

PiperOrigin-RevId: 235080500

PiperOrigin-RevId: 235080484

PiperOrigin-RevId: 235078474