// Assign an initial value to the variable, mirroring it to each component
  // device.
  {
    TensorHandlePtr initial_value_cpu = FloatTensorHandle(20., status.get());
    ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
    std::array<TFE_TensorHandle*, 2> components{initial_value_cpu.get(),
                                                initial_value_cpu.get()};

// Return a tensor handle containing a 3x2 matrix of floats
TFE_TensorHandle* TestMatrixTensorHandle3X2(TFE_Context* ctx);

// Return a variable handle referring to a variable with the given initial value
// on the given device.
TFE_TensorHandle* TestVariable(TFE_Context* ctx, float value,
                               const tensorflow::string& device_name = "");

// Return an add op multiplying `a` by `b`.

  if (TF_GetCode(status) != TF_OK) {
    return nullptr;
  }

  // Dump the op creation stacktraces for debugging purpose.
  DEBUG_DATA_DUMPER()->DumpOpCreationStackTraces(
      fn_name, kDebugGroupOpStacktrace, "initial", &fn_body->graph);

  tensorflow::StackTracesMap stack_traces;
  for (const Node* n : fn_body->graph.nodes()) {
    stack_traces[n->name()] = n->GetStackTrace();
  }

  // for created operations.
  const char* name;
} TF_WhileParams;

// Creates a TF_WhileParams for creating a while loop in `g`. `inputs` are
// outputs that already exist in `g` used as initial values for the loop
// variables.
//
// The returned TF_WhileParams will have all fields initialized except
// `cond_output`, `body_outputs`, and `name`. The `body_outputs` buffer will be

-  char ut_line[UT_LINESIZE];	/* Devicename.  */
-  char ut_id[4];		/* Inittab ID.  */
-  char ut_user[UT_NAMESIZE];	/* Username.  */
-  char ut_host[UT_HOSTSIZE];	/* Hostname for remote login.  */
+  char ut_line[UT_LINESIZE]
+    __attribute_nonstring__;	/* Devicename.  */
+  char ut_id[4]
+    __attribute_nonstring__;		/* Inittab ID.  */
+  char ut_user[UT_NAMESIZE]
+    __attribute_nonstring__;	/* Username.  */

  }

  std::unordered_map<int64_t, int64_t> gradients_size;
  // TODO(apassos) multiple threads could be dequeuing from op_stack at the same
  // time, for better CPU backprop performance.
  VLOG(1) << "Initial stack:";
  if (VLOG_IS_ON(1)) {
    for (auto t : op_stack) {
      VLOG(1) << "  " << t;
    }
  }
  while (!op_stack.empty()) {
    const int64_t op = op_stack.back();
    VLOG(1) << "Popped " << op;

    int64_t init_timeout_in_ms, TF_Status* status,
    bool clear_existing_contexts);

// Set server def with retries and timeout. This is helpful for fault-tolerant
// initial connection in high-preemption environments, such as
// ParameterServerStrategy training.
// This API is for experimental usage and may be subject to change.
TF_CAPI_EXPORT extern void TFE_ContextSetServerDefWithTimeoutAndRetries(

      converts them into `NoOp`s instead.

*   `tf.data`

    * `tf.data` now has an `autotune_options.initial_parallelism` option to
      control the initial parallelism setting used by autotune before the data
      pipeline has started running. The default is 16. A lower value reduces
      initial memory usage, while a higher value improves startup time.

## Keras

*  `keras.layers.experimental.DynamicEmbedding`

#include <stddef.h>
#include <stdint.h>

#include "tensorflow/c/tf_file_statistics.h"
#include "tensorflow/c/tf_status.h"

/// This is the interop header between core TensorFlow and modular filesystem
/// plugins (see initial RFC https://github.com/tensorflow/community/pull/101).
///
/// Both core TensorFlow and every plugin will use this header. The associated
/// `.cc` file is only used by core TensorFlow to implement checking needed for