//
  // -1 indicates an unknown axis length; this is unreachable for most standard
  // ImmediateExecutionTensorHandles, but comes up for example when computing
  // the shape of a parallel tensor with component shapes differing across
  // devices.
  virtual absl::Status Dim(int dim_index, int64_t* dim) const = 0;

  // Returns the device which created the handle.

    TF_ShapeAndTypeList** shape_list_array, int num_items);

// Infer shapes for the given `op`. The arguments mimic the arguments of the
// `shape_inference::InferenceContext` constructor. Note the following:
//   - The inputs of the `op` are not used for shape inference. So, it is
//     OK to not have the inputs properly set in `op`. See `input_tensors`
//     if you want shape inference to consider the input tensors of the
//     op for shape inference.

  dltensor_in->device = {kDLCPU, 0};
  dltensor_in->ndim = static_cast<int32_t>(shape.size());
  dltensor_in->dtype = {kDLFloat, 32, 1};
  dltensor_in->shape = shape.data();
  dltensor_in->strides = strides.data();
  TFE_TensorHandle* handle = TFE_HandleFromDLPack(&dlm_in, status, ctx);
  ASSERT_NE(handle, nullptr)
      << TF_Message(status) << " (shape=[" << absl::StrJoin(shape, ",")
      << "], strides=[" << absl::StrJoin(strides, ",") << "])";

Anis Elleuch <******@****.***> 1676280398 +0100

TFE_Op* MatMulOp(TFE_Context* ctx, TFE_TensorHandle* a, TFE_TensorHandle* b);

// Return an identity op.
TFE_Op* IdentityOp(TFE_Context* ctx, TFE_TensorHandle* a);

// Return a shape op fetching the shape of `a`.
TFE_Op* ShapeOp(TFE_Context* ctx, TFE_TensorHandle* a);

// Return an allreduce op adding up input tensor `in` from `group_size` workers.
TFE_Op* AllReduceOp(TFE_Context* ctx, TFE_TensorHandle* in, int group_size);

  virtual absl::Status TensorHandleStatus() const;

  // Returns tensor shape. If tensor has unknown rank, shape remains untouched.
  virtual absl::Status Shape(tensorflow::PartialTensorShape* shape) const = 0;

  // Returns tensor (full) type.
  // While there is no immediate plan to deprecate dtype and shape in favor
  // of only using full type type information, this is a future possibility.
  //

    return nullptr;
  }
  tensorflow::PartialTensorShape partial_shape;
  if (shape.num_dims != -1) {
    DCHECK(shape.dim_sizes != nullptr);
    Status status = tensorflow::PartialTensorShape::MakePartialShape(
        reinterpret_cast<int64_t*>(shape.dim_sizes), shape.num_dims,
        &partial_shape);
    if (!status.ok()) {
      tsl::Set_TF_Status_from_Status(s, status);

*   Tracing/timeline support for distributed runtime (no GPU profiler yet).
*   C API gives access to inferred shapes with `TF_GraphGetTensorNumDims` and
    `TF_GraphGetTensorShape`.
*   Shape functions for core ops have moved to C++ via
    `REGISTER_OP(...).SetShapeFn(...)`. Python shape inference RegisterShape
    calls use the C++ shape functions with `common_shapes.call_cpp_shape_fn`. A
    future release will remove `RegisterShape` from python.

    ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());

    TFE_Op* shape_op = ShapeOp(ctx, hgpu);
    TFE_OpSetDevice(shape_op, gpu_device_name.c_str(), status.get());
    ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
    TFE_TensorHandle* retvals[1];
    int num_retvals = 1;
    TFE_Execute(shape_op, &retvals[0], &num_retvals, status.get());

// number of dimensions of a parallel tensor.
int ParallelTensorNumDims(void* data, TF_Status* status) {
  const std::vector<int64_t>* shape;
  absl::Status s = reinterpret_cast<ParallelTensor*>(data)->Shape(&shape);
  if (!s.ok()) {
    tsl::Set_TF_Status_from_Status(status, s);
    return -1;
  }
  return shape->size();
}

// Used as an argument to TFE_NewCustomDeviceTensorHandle, for computing a
// dimension of a parallel tensor.