TFE_TensorHandle* m = TestMatrixTensorHandle(ctx);
  TFE_Op* matmul = TFE_NewOp(ctx, "MatMul", status);
  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TFE_TensorHandle* retvals[1];
  int num_retvals = 1;
  for (auto s : state) {
    TFE_OpReset(matmul, "MatMul", nullptr, status);
    CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
    TFE_OpAddInput(matmul, m, status);

    TF_Operation* const1 = FloatConst2x2(graph_, s_, const1_val, "Const_1");
    TF_Operation* matmul = MatMul(graph_, s_, const0, const1, "MatMul");
    inputs[0] = TF_Output{const0, 0};
    inputs[1] = TF_Output{const1, 0};
    outputs[0] = TF_Output{matmul, 0};
    EXPECT_EQ(TF_OK, TF_GetCode(s_)) << TF_Message(s_);
  }

  void BuildExpectedGraph(bool grad_inputs_provided,

  TFE_TensorHandle* h0_task0 = TestMatrixTensorHandle(ctx);

  TFE_Op* matmul = MatMulOp(ctx, h0_task0, h0_task0);
  TFE_OpSetDevice(matmul, remote_device_name, status);
  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  TFE_TensorHandle* retvals[1];
  int num_retvals = 1;
  TFE_Execute(matmul, &retvals[0], &num_retvals, status);
  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  TFE_TensorHandle* m = TestMatrixTensorHandle(ctx);
  TFE_Op* matmul = MatMulOp(ctx, m, m);
  TFE_TensorHandle* retvals[2] = {nullptr, nullptr};
  int num_retvals = 2;
  TFE_Execute(matmul, &retvals[0], &num_retvals, status);
  EXPECT_EQ(1, num_retvals);
  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TFE_DeleteOp(matmul);
  TFE_DeleteTensorHandle(m);

  TFE_ContextOptions* tfe_context_options_;
  TFE_Context* tfe_context_;
};

TEST_F(ShapeInferenceTest, InfersShapesFromInputShapes) {
  TFE_Op* matmul_op;
  matmul_op = TFE_NewOp(tfe_context_, "MatMul", status_);
  CHECK_EQ(TF_OK, TF_GetCode(status_)) << TF_Message(status_);

  // Infer shape when everything is known.
  CheckOutputShapes(matmul_op,

  return op;
}

TFE_Op* MatMulOp(TFE_Context* ctx, TFE_TensorHandle* a, TFE_TensorHandle* b) {
  TF_Status* status = TF_NewStatus();

  TFE_Op* op = TFE_NewOp(ctx, "MatMul", status);
  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TFE_OpAddInput(op, a, status);
  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TFE_OpAddInput(op, b, status);

* `tf.config.experimental.enable_tensor_float_32_execution`
    * Disabling TensorFloat-32 execution now causes TPUs to use float32 precision for float32 matmuls and other ops. TPUs have always used bfloat16 precision for certain ops, like matmul, when such ops had float32 inputs. Now, disabling TensorFloat-32 by calling `tf.config.experimental.enable_tensor_float_32_execution(False)` will cause TPUs to use float32 precision for such ops instead of bfloat16.