}
  }
  attr {
    key: "dtype"
    value {
      type: DT_INT32
    }
  }
}
node {
  name  : "y_reshape"
  op    : "Reshape"
  input : "y_const"
  input : "y_shape"
  attr { key: "T" value { type: DT_INT32 } }
  # Attribute TShape not specified; needs to be set to its default
  # by tfcompile.
}
node {
  name  : "x_y_sum"
  op    : "Add"
  input : "x_const"

  auto x_shape = Shape(scope, x);
  auto output_shape = Shape(scope, op.output(0));

  // Reduce away broadcasted leading dims.
  auto reduce_x = internal::BroadcastGradientArgs(scope, x_shape, output_shape);
  auto gx_sum =
      ReduceSum(scope, gx, /*axis=*/reduce_x.r0, ReduceSum::KeepDims(true));
  auto gx_sum_reshape = Reshape(scope, gx_sum, x_shape);

  // this can fail if we're not running on a TPU-enabled node.
  // TODO(kramm): Move this into a separate pass. See b/184944903
  xla::Shape old_shape = xla::TypeToShape(t);
  XLA_Shape old_shape_c = {};
  XLA_Shape new_shape_c = {};
  TfTpu_ExecutorApiFn *executor = stream_executor::tpu::ExecutorApiFn();
  if (!stream_executor::tpu::IsInitialized(executor)) {
    return failure();
  }

#include "tensorflow/c/experimental/saved_model/public/signature_def_param_list.h"
#include "tensorflow/c/experimental/saved_model/public/tensor_spec.h"
#include "tensorflow/c/tf_datatype.h"
#include "tensorflow/c/tf_shape.h"
#include "tensorflow/c/tf_status.h"
#include "tensorflow/c/tf_tensor.h"
#include "tensorflow/core/lib/io/path.h"
#include "tensorflow/core/platform/status.h"
#include "tensorflow/core/platform/stringpiece.h"

        "@local_xla//xla/tsl/c:tsl_status_internal_headers",
    ],
    visibility = [
        "//tensorflow/python:__subpackages__",
    ],
)

cc_library(
    name = "tf_shape",
    srcs = ["tf_shape.cc"],
    hdrs = ["tf_shape.h"],
    copts = tf_copts(),
    visibility = ["//visibility:public"],
    deps = [
        ":c_api_macros",
        ":tf_shape_internal",
        "//tensorflow/core:framework",

  // https://www.tensorflow.org/api_docs/cc/class/tensorflow/ops/batch-mat-mul
  int64_t x_row_dim = x_shape[x_shape.size() - 2];
  int64_t x_col_dim = x_shape[x_shape.size() - 1];
  int64_t y_row_dim = y_shape[y_shape.size() - 2];
  int64_t y_col_dim = y_shape[y_shape.size() - 1];
  int64_t out_row_dim = output_shape[output_shape.size() - 2];
  int64_t out_col_dim = output_shape[output_shape.size() - 1];

  if (op.num_inputs() != 2) {
    return errors::InvalidArgument("BroadcastTo requires 2 inputs");
  }

  auto x_shape = Shape(scope, op.input(0));
  auto args = internal::BroadcastGradientArgs(scope, x_shape, op.input(1));
  auto sum_gx = Sum(scope, grad_inputs[0], args.r0);
  grad_outputs->push_back(Reshape(scope, sum_gx, x_shape));
  grad_outputs->push_back(NoGradient());
  return scope.status();
}

    TF_RETURN_IF_ERROR(tensorflow::ConvertToTensor(
        llvm::dyn_cast<TF::ConstOp>(fold_results[0].getDefiningOp()).getValue(),
        &tensor));
    xla::Shape xla_shape;
    TF_RETURN_IF_ERROR(tensorflow::TensorShapeToXLAShape(
        tensor.dtype(), tensor.shape(), &xla_shape));
    xla::PjRtClient::HostBufferSemantics host_buffer_semantics =
        xla::PjRtClient::HostBufferSemantics::kImmutableUntilTransferCompletes;

  ArrayRef<int64_t> in_shape = ranked_type.getShape();
  if (in_shape.empty() || in_shape[0] < 0) {
    return context_op->emitOpError()
           << "A map_outside_compilation op's input and output shapes must "
              "have rank at least one and the first dimension must be known.";
  }
  int64_t split_size = in_shape[0] / num_cores_per_replica;
  if (in_shape[0] % num_cores_per_replica != 0) {

  )
  def test_einsum_ptq_model(
      self,
      equation: str,
  ):
    _, y_shape, bias_shape, x_signature, y_signature = (
        self._prepare_sample_einsum_datashapes(equation, use_bias=True)
    )

    model = self._create_einsum_model(
        self._input_saved_model_path,
        equation,
        y_shape,
        x_signature,
        y_signature,
        bias_shape,
    )