The owned PjRtBuffers in `owned_executable_args` need to live until execution is complete. Currently this is achieved by blocking until all the executable outputs are ready. However, this seemed to cause performance overheads, see b/299683272 and b/300102691.

With this change, we don't block until execution is complete. The ownership of `owned_executable_args` is moved to a lambda which is executed as a callback when the PjRtFuture returned by ExecutePortable is ready (which happens when the execution is complete).

PiperOrigin-RevId: 565169152

PiperOrigin-RevId: 565168980

Such that users can act on the classes, adding the override methods there.

PiperOrigin-RevId: 565168942

[XLA] Add WithReplicaGroups in pattern matcher and modify tests to conform to the new pattern matching format

-Add WithReplicaGroups implementation for HloInstructionPattern to match with the collective instruction's replica groups.

PiperOrigin-RevId: 565160403

PiperOrigin-RevId: 565158613

PiperOrigin-RevId: 565157773

Open source sparse_core_ops_utils*

PiperOrigin-RevId: 565156105

PiperOrigin-RevId: 565152217

PiperOrigin-RevId: 565140913

PiperOrigin-RevId: 565134913

PiperOrigin-RevId: 565132363

PiperOrigin-RevId: 565121864

PiperOrigin-RevId: 565119660

PiperOrigin-RevId: 565113938

PiperOrigin-RevId: 565111478

PiperOrigin-RevId: 565106502

PiperOrigin-RevId: 565097655

PiperOrigin-RevId: 565094585

PiperOrigin-RevId: 565092976

[XLA] Rework dot() sharding propagation to lookahead instructions sharding to choose a sharding for dot() that agrees with the users if possible.

PiperOrigin-RevId: 565086052

PiperOrigin-RevId: 565083641

Small collectives might be better off when sinked and there are other potnential use cases
Also fix a bug, where we were accepting reuse of the data that we were storing and changing the tests using that pattern to match the fix.

PiperOrigin-RevId: 565080772

instructions through control dependence.

This is because the generated HLO program is correct even without the control
dependence chaining. The purpose of the control dependence chaining is to
support a scheduler, such as the latency hiding scheduler, and thus will be
added to the latency hiding scheduler preparation pass. Not producing the
control dependence chaining while decomposing collective-permute can also
simplify the implementation of collective-pipeliner in pipelining Send and
Recv instructions.

PiperOrigin-RevId: 565073772

Updates LLVM usage to match
[8ebe1d1cc1e4](https://github.com/llvm/llvm-project/commit/8ebe1d1cc1e4)

PiperOrigin-RevId: 565069541

Without this additional bazelrc file TSL will be built with the system compiler which happens to be GCC.

PiperOrigin-RevId: 565060297

This change was merged upstream in https://github.com/openai/triton/pull/2068 but hasn't made it to the OpenXLA fork yet.

Improves build time in OSS by not building a number of MLIR dialects/passes that are not needed.

PiperOrigin-RevId: 565058458

PiperOrigin-RevId: 565025664

PiperOrigin-RevId: 565024582

Unify iterator names to *_it;
Use auto* instead of auto if the type is a pointer type.

PiperOrigin-RevId: 565016627

This is another prefactoring to make Triton fusions compatible with
FusionInterface and partially fused HLO. For the former, we need
the LaunchDimension computation to be a separate function. For the
latter, we change the launch dimension function signatures to no longer
take an HloComputation, because we don't yet have that during fusion (at
least not a complete one). For now, this change is a no-op, since we do
not yet have any boundary functions for non-fusion ops.

PiperOrigin-RevId: 565015567

that the corresponding FlatBuffer C API type can also be used instead of
the FlatBuffer C++ API type for the TFLiteSettings FlatBuffer parameter.

PiperOrigin-RevId: 565008668

From the original code at commit 03d304b, we are only supposed to scale the reduction
tile size of dimX by the unroll factor for column reductions, so the check when creating
`ReductionCodegenInfo` is only valid for column reductions.

PiperOrigin-RevId: 564991695

Add a comparison operator for fragments, improve encapsulation, fix comments.

PiperOrigin-RevId: 564990063

LaunchDimensions are being decoupled from codegen. Shared memory requirements
are only known after codegen.

Currently, this feature is only used for Triton fusions. All other
fusions that use shared memory allocate it within the kernel.

PiperOrigin-RevId: 564981685

PiperOrigin-RevId: 564978233

PiperOrigin-RevId: 564978209

The upcoming CUDA-12 upgrades requires TensorRT 8.6 and this version
has a new set of headers which requires an update to the bazel configure script.

I also had to change `find_cuda_config.py` because previously it has only
been reporting the major version of TensorRT back to the configure script. But
we will also need the minor version to distinguish between TensorRT 8.5 and below,
and TensorRT 8.6+.

PiperOrigin-RevId: 564927510

PiperOrigin-RevId: 564914384

PiperOrigin-RevId: 564909053

GCC is missing some intrinsics, expand them by hand.

PiperOrigin-RevId: 564899448