if ver < 35:
            print('WARNING: XLA does not support CUDA compute capabilities '
                  'lower than sm_35. Disable XLA when running on older GPUs.')
      else:
        ver = float(m.group(0))
        if ver < 3.0:
          print('ERROR: TensorFlow only supports CUDA compute capabilities 3.0 '
                'and higher. Please re-specify the list of compute '

};

// Template instantiations here

inline bool IsDtypeTrainable(DataType dtype) {
  switch (dtype) {
    case DT_HALF:
    case DT_BFLOAT16:
    case DT_FLOAT:
    case DT_DOUBLE:
    case DT_COMPLEX64:
    case DT_COMPLEX128:
    case DT_RESOURCE:
    case DT_VARIANT:
      return true;
    case DT_QINT8:
    case DT_QINT16:
    case DT_QINT32:

# END TF TEST SUITE OPTIONS

# START CROSS-COMPILE CONFIGS
# Set execution platform to Linux x86
# Note: Lot of the "host_" flags such as "host_cpu" and "host_crosstool_top"
# flags seem to be actually used to specify the execution platform details. It
# seems it is this way because these flags are old and predate the distinction

    ASSERT_EQ(errors::OK, s.code()) << s.message();
    ctx.reset(ctx_raw);
  }

  AbstractTensorHandlePtr t;
  {
    AbstractTensorHandle* x_raw = nullptr;
    Status s = TestScalarTensorHandle<float, TF_FLOAT>(ctx.get(), 1.0f, &x_raw);
    ASSERT_EQ(errors::OK, s.code()) << s.message();
    t.reset(x_raw);
  }

  AbstractOperationPtr check_numerics_op(ctx->CreateOperation());
  ForwardOperation forward_op;

  GetDims(theta_tensor, theta_dims.data());

  // Initialize auxilary data structures.
  vector<float> thetaPlus_data(num_elems);
  vector<float> thetaMinus_data(num_elems);
  AbstractTensorHandle* f_outputs[1];

  // Numerical Grad Check
  for (int i = 0; i < num_elems; i++) {
    // Get relative epsilon value
    float epsilon = theta_data[i] == 0 ? 1e-4 : std::abs(theta_data[i] * 1e-4);
    AbstractTensorHandlePtr two_eps;

  delete[] static_cast<float*>(data);
}

// Creates a TFE_TensorHandle with value `v`.
TensorHandlePtr FloatTensorHandle(float v, TF_Status* status) {
  const int num_bytes = sizeof(float);
  float* values = new float[1];
  values[0] = v;
  std::unique_ptr<TF_Tensor, decltype(&TF_DeleteTensor)> tensor(
      TF_NewTensor(TF_FLOAT, nullptr, 0, values, num_bytes, &FloatDeallocator,

TFE_TensorHandle* TestMatrixTensorHandleWithInput(TFE_Context* ctx,
                                                  float data[], int64_t dims[],
                                                  int num_dims);

// Get a Matrix TensorHandle with given float values and dimensions
TFE_TensorHandle* TestTensorHandleWithDimsFloat(TFE_Context* ctx, float data[],
                                                int64_t dims[], int num_dims);

}

TEST(TensorPjRtBufferUtilTest, GetPjRtCBufferFromTensorNoBuffer) {
  auto allocator = std::make_unique<AsyncValueAllocator>();
  tensorflow::Tensor tensor(allocator.get(), DT_FLOAT, {1});

  EXPECT_THAT(
      GetPjRtCBufferFromTensor(&tensor),
      StatusIs(error::INTERNAL, HasSubstr(absl::StrCat(
                                    "Input tensor does not have PjRtBuffer"))));
}

      "        name: 'sum'"
      "        type: DT_FLOAT"
      "      }"
      "    }"
      "    node_def {"
      "      name: 'read0'"
      "      op: 'ReadVariableOp'"
      "      input: 'var'"
      "      device: '/job:localhost/replica:0/task:0/device:CPU:0'"
      "      attr {"
      "        key: 'dtype'"
      "        value {"
      "          type: DT_FLOAT"
      "        }"
      "      }"
      "    }"

                         status.get());
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());

  // Create two vectors with different lengths
  std::vector<float> size_two_value{1., 2.};
  std::vector<float> size_three_value{1., 2., 3.};
  TensorHandlePtr size_two(
      VectorFloatTensorHandle(size_two_value, status.get()));
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());