These utilities and MLIR types are required to enable PAX
model checkpointing involving the new TF int4 type.

PiperOrigin-RevId: 563915180

PiperOrigin-RevId: 563892946

Updates LLVM usage to match
[f8a1c8b7c1d0](https://github.com/llvm/llvm-project/commit/f8a1c8b7c1d0)

PiperOrigin-RevId: 563871269

PiperOrigin-RevId: 563858734

This is required to prevent a circular dependency in https://github.com/tensorflow/tensorflow/pull/60161

PiperOrigin-RevId: 563836979

PiperOrigin-RevId: 563836927

These types are used by JAX/XLA.  Expanding coverage in TF to match.

PiperOrigin-RevId: 563831916

Update CMake build rules to adopt new TSL path

PiperOrigin-RevId: 563829912

PiperOrigin-RevId: 563829150

This does not need to be disabled anymore now that TensorFlow IO has Apple Silicon wheels available.

See https://pypi.org/project/tensorflow-io-gcs-filesystem/0.34.0/#files

PiperOrigin-RevId: 563826666

PiperOrigin-RevId: 563822172

The current toolchain doesn't support it.

PiperOrigin-RevId: 563796297

PiperOrigin-RevId: 563787003

This method is no longer used by JAX, and its existence is misleading (what it returns does not correspond to any particular default device assignment that JAX uses). Remove it.

PiperOrigin-RevId: 563784500

PiperOrigin-RevId: 563784409

PiperOrigin-RevId: 563776249

PiperOrigin-RevId: 563773691

Check iterators before dereference if the validity is not trivial.
(I think there are invariants which ensure that they are valid currently, but this could break accidentally in a future change.)
Change an algorithm to use iterators more safely.

PiperOrigin-RevId: 563765952

PiperOrigin-RevId: 563753463

PiperOrigin-RevId: 563746345

This can be used to supply it with a graph cut instead of a fusion computation

PiperOrigin-RevId: 563739866

PiperOrigin-RevId: 563735974

PiperOrigin-RevId: 563712154

PiperOrigin-RevId: 563705867

PiperOrigin-RevId: 563691979

PiperOrigin-RevId: 563686010

PiperOrigin-RevId: 563686005

PiperOrigin-RevId: 563648249

PiperOrigin-RevId: 563641616

Two problems:
* 1L is not int64 sized on Windows, leading to a template argument deduction problem.
* MSVC limits string literals to 16K characters. However, you can use string literal concatenation to concatenate shorter literals into a long literal. Do that instead.

PiperOrigin-RevId: 563641264

This change introduces new function lifting mode in the TPU quantizer to lift
the StableHLO subgraph with TF::XlaCallModuleOp.

PiperOrigin-RevId: 563640156

PiperOrigin-RevId: 563639265

We want a consistent definition for all int4 usages.  Currently,
JAX/TF use the underlying type defined in `ml_dtypes`.

PiperOrigin-RevId: 563633360

~Argument() names the destructor of the superclass, and MSVC rightly complains about it in the event that (*self)[i] does not return an Argument, which it does in some cases.

PiperOrigin-RevId: 563615070

This is a prerequisite to adding proper dtype support.  Exposes
`tsl::int4` and `tsl::uint4` (with corresponding `tensorflow::` versions)
into the platform list of types.

PiperOrigin-RevId: 563607886

PiperOrigin-RevId: 563596107

Spec: https://github.com/openxla/stablehlo/blob/main/docs/spec.md#batch_norm_training

Lowering the op with some TF ops (tf.Mean, tf.SquaredDifference)
which will be eventually lowered to TFL ops.

Tested with tensorflow/compiler/mlir/lite/stablehlo/tests:unfuse_mhlo_batch_norm.mlir.test

PiperOrigin-RevId: 563582603

BAZEL_CURRENT_REPOSITORY is given a value like
```
BAZEL_CURRENT_REPOSITORY=\"xla\"
```
but nvcc is unable to parse the quotes in that argument. It's a value produced by Bazel and probably not used, so just skip it.
PiperOrigin-RevId: 563577393

PiperOrigin-RevId: 563575970

PiperOrigin-RevId: 563570251