The line 
```
segmentation_options = processor.SegmentationOptions(
    output_type=processor.SegmentationOptions.OutputType.CATEGORY_MASK) 
 ```
is changed to 
```
segmentation_options = processor.SegmentationOptions(
    output_type=processor.SegmentationOptions.output_type.CATEGORY_MASK) 
```

After scheduling, it is possible to eliminate more copies that have been added
by the CopyInsertion pass, because now it is clear in which order the
operations are executed.

PiperOrigin-RevId: 561300243

PiperOrigin-RevId: 561283397

PiperOrigin-RevId: 561274977

PiperOrigin-RevId: 561272373

PiperOrigin-RevId: 561272265

PiperOrigin-RevId: 561263745

Add implementations to handle the case where the calibration wrapper accepts a TFLite flatbuffer model that uses offset buffers.
Offset buffers are enabled by the TFLite converter's `_experimental_use_buffer_offset` flag.
Whenever the model's base buffer (that contains the serialized flatbuffer) is modified, the offset values to the offset buffer of the tensor buffers should be updated.

PiperOrigin-RevId: 561251067

Per-layer verify creates a model with both quantized and original unquantized layer.
Unquantized layer's input comes from previous layer's quantized layer.
Outputs of quantized and unquantized layers are saved via DumpTensorOp

PiperOrigin-RevId: 561233115

DumpTensorOp saves entire value of quantizable layer during inference.

Whole model verify creates two models:
1. Unquantized model with DumpTensorOp - saved at unquantized_dump_model_path
2. Quantized model with DumpTensorOp - saved at output_directory of quantized model

PiperOrigin-RevId: 561222579

PiperOrigin-RevId: 561220243

PiperOrigin-RevId: 561208020

PiperOrigin-RevId: 561206751

PiperOrigin-RevId: 561205731

PiperOrigin-RevId: 561195486

Updates LLVM usage to match
[96e83d3705c7](https://github.com/llvm/llvm-project/commit/96e83d3705c7)

PiperOrigin-RevId: 561194584

PiperOrigin-RevId: 561192100

PiperOrigin-RevId: 561188743

PiperOrigin-RevId: 561184976

PiperOrigin-RevId: 561183522

PiperOrigin-RevId: 561183306

DataRef implementation to wrap tflite tensors. This lets us interop cleanly with the tflite memory system within out functions.

PiperOrigin-RevId: 561180210

PiperOrigin-RevId: 561178670

See https://dmlc.github.io/dlpack/latest/python_spec.html.

This is the import path. The export path was implemented in
https://github.com/openxla/xla/commit/0b3cbfe4bc7cd68dc20924bd878cdfb2faa1a169.

This allows for creating jax.Arrays from external GPU arrays
asynchronously.

PiperOrigin-RevId: 561172624

PiperOrigin-RevId: 561168251

* Rearrange PJRT_Api struct before major version bump
* Return unimplemented instead of nullptr

PiperOrigin-RevId: 561165318

Temporarily disable LaxBackedNumpyTests.testSortKeyValue_(2,3,4)_float32_bool_1 to pass TF nightly build

PiperOrigin-RevId: 561163175

PiperOrigin-RevId: 561161226

PiperOrigin-RevId: 561150086

PiperOrigin-RevId: 561148975

This changes removes `PJRT_Buffer_OnDeviceTrimmedShape` and related
functionality, since passing around xla::Shapes is expensive and often
includes more information than is necessary or even meaningful, and we
now have more specific getters that should be used instead.

This means that `PjRtBuffer::on_device_shape` and
`PjRtBuffer:logical_on_device_shape` no longer work with the
PjRtCApiClient. Instead, C++ callers should use the new more-specific
getters:

* `PjRtBuffer::element_type()`
* `PjRtBuffer::dimensions()`
* `PjRtBuffer::layout()`
* `PjRtBuffer::has_dynamic_dimensions()`
* `PjRtBuffer::is_dynamic_dimension()`
* `PjRtBuffer::logical_dimensions()`

These all have corresponding C APIs.

PiperOrigin-RevId: 561148407

PiperOrigin-RevId: 561141352

PiperOrigin-RevId: 561137986

Broke argsort on gpu.

PiperOrigin-RevId: 561136138

PiperOrigin-RevId: 561134835

PiperOrigin-RevId: 561134253

PiperOrigin-RevId: 561126894

JAX CI started failing with:

```
external/xla/xla/service/gpu/fusion_pipeline.cc:85:10: error: use of deleted function 'xla::HloPassPipeline::HloPassPipeline(const xla::HloPassPipeline&)'
   85 |   return fusion;
      |          ^~~~~~
In file included from external/xla/xla/service/gpu/fusion_pipeline.h:21,
                 from external/xla/xla/service/gpu/fusion_pipeline.cc:16:
external/xla/xla/service/hlo_pass_pipeline.h:36:7: note: 'xla::HloPassPipeline::HloPassPipeline(const xla::HloPassPipeline&)' is implicitly deleted because the default definition would be ill-formed:
   36 | class HloPassPipeline : public HloPassInterface {
      |       ^~~~~~~~~~~~~~~
external/xla/xla/service/hlo_pass_pipeline.h:36:7: error: use of deleted function 'std::unique_ptr<_Tp, _Dp>::unique_ptr(const std::unique_ptr<_Tp, _Dp>&) [with _Tp = xla::CompilationStats; _Dp = std::default_delete<xla::CompilationStats>]'
In file included from /dt9/usr/bin/../lib/gcc/x86_64-pc-linux-gnu/9/../../../../include/c++/9/memory:80,
                 from external/xla/xla/stream_executor/device_description.h:25,
                 from external/xla/xla/service/gpu/gpu_types.h:21,
                 from external/xla/xla/service/gpu/gpu_device_info.h:22,
                 from external/xla/xla/service/gpu/fusion_pipeline.h:19,
                 from external/xla/xla/service/gpu/fusion_pipeline.cc:16:
/dt9/usr/bin/../lib/gcc/x86_64-pc-linux-gnu/9/../../../../include/c++/9/bits/unique_ptr.h:414:7: note: declared here
  414 |       unique_ptr(const unique_ptr&) = delete;
      |       ^~~~~~~~~~
external/xla/xla/service/gpu/fusion_pipeline.cc: In function 'xla::HloPassPipeline xla::gpu::HorizontalFusionPipeline(const xla::gpu::GpuDeviceInfo&)':
external/xla/xla/service/gpu/fusion_pipeline.cc:96:10: error: use of deleted function 'xla::HloPassPipeline::HloPassPipeline(const xla::HloPassPipeline&)'
   96 |   return horizontal_fusion;
      |          ^~~~~~~~~~~~~~~~~
```

I'm guessing this is just GCC having a slightly different interpretation of whether this needs a copy.

PiperOrigin-RevId: 561123362

PiperOrigin-RevId: 561114720

PiperOrigin-RevId: 561111836