// type.
absl::StatusOr<DeviceType> GetCompilationDeviceType(
    const DeviceType& platform_device_type);

// Builds a DeviceCompiler that uses xla::LocalClient using `platform_info` and
// `compilation_device_type` (in non-TPU case) and sets *xla_device_compiler to
// point to it. Uses flags from `MarkForCompilationPassFlags` for configuring
// the persistor used in the DeviceCompiler. The platform ID from
// `platform_info` must not be null in CPU case.

  if (!is_calibration_required) return absl::OkStatus();

  // `duplicate_shape_determining_constants = false` because the
  // resulting graph of this step is not expected to be loaded on TPU.
  const ExportOptions export_opts = {
      /*duplicate_shape_determining_constants=*/false,
      /*unfreeze_constants=*/false, checkpoint_dir,
      /*debug_name=*/absl::StrCat(kName, kExportStepSuffix)};

        `.predict` is available for Cloud TPUs, Cloud TPU, for all types of
        Keras models (sequential, functional and subclassing models).
    *   Automatic outside compilation is now enabled for Cloud TPUs. This allows
        `tf.summary` to be used more conveniently with Cloud TPUs.
    *   Dynamic batch sizes with DistributionStrategy and Keras are supported on
        Cloud TPUs.

                       const FunctionLibraryDefinition* function_library) {
  auto predicate = [](const Graph& graph) {
    for (const Node* node : graph.nodes()) {
      // _tpu_replicate is used in replicated TPU graphs. It will be converted
      // to_replication_info and _xla_compile_device_type in phase 1 pipelines.
      if (node->attrs().FindByString(std::string(kTpuReplicateAttr))) {
        return true;
      }
    }

#include "absl/strings/strip.h"
#include "tensorflow/compiler/mlir/tensorflow/utils/dump_graph.h"
#include "xla/parse_flags_from_env.h"
#include "tensorflow/core/platform/macros.h"
#include "tensorflow/core/tpu/kernels/sparse_core_xla_flags_defaults.h"
#include "tensorflow/core/util/command_line_flags.h"

namespace tensorflow {
namespace {

BuildXlaOpsPassFlags* build_ops_flags;

extern const char kXlaLiftedArgOutsideCompilationAttrName[];

// Attribute indicating that this is an IdentityN node receiving inputs for a
// outside compilation Placeholder node (the original outside compilation node
// is moved out of TPU computation, and we left a Placeholder node there).
// Attribute value will be a string, which is the outside compilation cluster
// name for the outside compilation Placeholder node.

// RUN: tf-opt %s -split-input-file -verify-diagnostics -prepare-tpu-computation-for-tf-export | FileCheck %s

// CHECK-LABEL: @ShardingAttr
func.func @ShardingAttr(%arg0: tensor<128x10xf32> {mhlo.sharding = "\08\03\1A\02\01\02\22\02\00\01"}, %arg1: tensor<10x1024xf32> {mhlo.sharding = "\08\01\1A\01\01\22\01\00"}, %arg2: tensor<128x1024xf32> {mhlo.sharding = ""}) -> (tensor<128x10xf32>, tensor<10x1024xf32>, tensor<128x1024xf32>) {

-   Different Python versions
-   Linux, MacOS, and Windows machines (these pool definitions are internal)
-   x86 and arm64
-   CPU-only, or with NVIDIA CUDA support (Linux only), or with TPUs

## How to Test Your Changes to TensorFlow

You may check how your changes will affect TensorFlow by:

1. Creating a PR and observing the presubmit test results
2. Running the CI scripts locally, as explained below

		return
	}
	tp0 := s.typePointersOfType(typ, addr)
	tp1 := s.typePointersOf(addr, size)
	failed := false
	for {
		var addr0, addr1 uintptr
		tp0, addr0 = tp0.next(addr + size)
		tp1, addr1 = tp1.next(addr + size)
		if addr0 != addr1 {
			failed = true
			break
		}
		if addr0 == 0 {
			break
		}
	}
	if failed {
		tp0 := s.typePointersOfType(typ, addr)

// RUN: tf-tfrt-opt -tf-to-tfrt %s | FileCheck %s --dump-input=fail --dump-input-filter=all
// RUN: tf-tfrt-opt -pass-pipeline='builtin.module(tf-to-tfrt{target-tpurt=true tpu-use-core-selector=false})' %s | FileCheck %s --dump-input=fail --dump-input-filter=all

// CHECK-LABEL: func @_tfrt_fallback_init
// CHECK-SAME: {{.*}} !tfrt.chain
// CHECK: tfrt_fallback_async.createop(%arg0) key(0) device("/device:CPU:0") "tf.ParseExampleV2"()