import os
import math
from flask import Flask, request
app = Flask(__name__)


@app.route('/hello')
def hello():
    version = os.environ.get('SERVICE_VERSION')

    # do some cpu intensive computation
    x = 0.0001
    for i in range(0, 1000000):
        x = x + math.sqrt(x)

    return 'Hello version: %s, instance: %s\n' % (version, os.environ.get('HOSTNAME'))


@app.route('/health')
def health():

   *   <li>To get uninterruptibility and remove checked exceptions, use {@link
   *       Futures#getUnchecked}.
   * </ul>
   *
   * @throws ExecutionException if the computation threw an exception
   * @throws CancellationException if the computation was cancelled
   */
  @CanIgnoreReturnValue
  @ParametricNullness
  public static <V extends @Nullable Object> V getUninterruptibly(Future<V> future)

// For case 3, we need to create a PjRtBuffer from the raw device mem pointer,
// and we need to ensure the PjRtBuffer persists till XLA computation is
// complete. Therefore we put the newly created PjRtBuffer into `owned_args`.
// Caller is responsible to ensure `owned_args` lives till the end of XLA
// computation.
Status PreparePjRtExecutableArguments(
    int num_missing_prefix_ctx_inputs, const std::vector<int>& input_mapping,

    }
  }
}

// Since we have the outputs from host and device computation after moving
// outside compiled ops, we can create the actual parallel_execute regions.
// Still, one region is for the host computation for outside compilation and
// the other one is for the original Device cluster computation.
mlir::tf_device::ParallelExecuteOp CreateFinalParallelExecuteOp(

  }

 private:
  std::unique_ptr<DeviceMgr> device_mgr_;
  std::unique_ptr<ProcessFunctionLibraryRuntime> pflr_;
};

TEST_F(ExtractOutsideCompilationForFunctionTest, Basic) {
  // Build the XLA computation func.
  // "const0"
  // "identity0" = "const0" (outside compilation cluster "0")
  // "identity1" = "identity0" (outside compilation cluster "1")
  // "identity2" = "identity1"
  FunctionDefLibrary fdl;
  {

should be used with caution when working with untrusted models.

### Saved graphs and checkpoints

When loading untrusted serialized computation graphs (in form of a `GraphDef`,
`SavedModel`, or equivalent on-disk format), the set of computation primitives
available to TensorFlow is powerful enough that you should assume that the
TensorFlow process effectively executes arbitrary code.

// CHECK-NEXT:  for node {{[{][{]node Add[}][}]}}

func.func @main() {
  tf_executor.graph {
    %0 = tf_executor.island wraps "tf._TPUReplicate"() {computation = @foo, Tinputs = [], Tbroadcast_inputs = [], NumVariables = 0, Tguaranteed_constants = [], output_types = []} : () -> () loc("_TPUReplicate")
    tf_executor.fetch
  }
  func.return
}

func.func @foo() {

  %cst_3 = "tf.Const"() {value = dense<4> : tensor<1xi32>} : () -> tensor<1xi32>
  %0 = "tf.XlaReduceWindow"(%arg0, %arg1, %cst_0, %cst_1, %cst_2, %cst_3, %cst) {computation = @sum_reducer} : (tensor<7xf32>, tensor<f32>, tensor<1xi32>, tensor<1xi32>, tensor<1xi32>, tensor<1xi32>, tensor<1x2xi32>) -> tensor<10xf32>
  func.return %0 : tensor<10xf32>
}

* Introducing `tf.types.experimental.AtomicFunction` as the fastest way to perform TF computations in Python.

    * Can be accessed through `inference_fn` property of `ConcreteFunction`s
    * Does not support gradients.
    * See `tf.types.experimental.AtomicFunction` documentation for how to call and use it.

XlaCompiler::CompileOptions GetCompileOptions(bool for_pjrt = false) {
  XlaCompiler::CompileOptions compile_options;
  compile_options.is_entry_computation = true;
  // Optimization: where possible, have the computation return a naked array
  // rather than a one-element tuple.
  compile_options.always_return_tuple = false;
  if (for_pjrt) {
    compile_options.use_tuple_arg = false;
    compile_options.always_return_tuple = true;
  }