EXPECT_EQ(dltensor_out->ndim, dltensor_in->ndim);
  EXPECT_EQ(dltensor_out->dtype.code, dltensor_in->dtype.code);
  EXPECT_EQ(dltensor_out->dtype.bits, dltensor_in->dtype.bits);
  EXPECT_EQ(dltensor_out->dtype.lanes, dltensor_in->dtype.lanes);
  for (int i = 0; i < dltensor_in->ndim; ++i) {
    EXPECT_EQ(dltensor_out->shape[i], dltensor_in->shape[i]);
    if (dltensor_out->strides) {

    TF_RETURN_IF_ERROR(forward_attrs_.Get("data_format", &data_format));

    // Grad for A
    grad_inputs[0] = upstream_grad;
    grad_inputs[0]->Ref();

    // Grad for bias
    std::string name = "bias_add_grad";
    TF_RETURN_IF_ERROR(BiasAddGrad(ctx, upstream_grad, &grad_inputs[1],
                                   data_format.c_str(), name.c_str()));

    return absl::OkStatus();
  }

  switch (dtype.code) {
    case DLDataTypeCode::kDLBool:
      if (dtype.bits != 8) {
        return tensorflow::errors::InvalidArgument(
            "Only DLPack bools of bitwidth 8 are supported, got: ", dtype.bits);
      }
      *tf_dtype = TF_DataType::TF_BOOL;
      return absl::OkStatus();

    case DLDataTypeCode::kDLUInt:
      switch (dtype.bits) {
        case 8:
          *tf_dtype = TF_DataType::TF_UINT8;

  char* dst = buffer;
  int64_t read = 0;

  while (n > 0) {
    // Some platforms, notably macs, throw `EINVAL` if `pread` is asked to read
    // more than fits in a 32-bit integer.
    size_t requested_read_length;
    if (n > INT32_MAX)
      requested_read_length = INT32_MAX;
    else
      requested_read_length = n;

    A.reset(A_raw);
  }
  // Bias
  float Bias_vals[] = {2.0f, 3.0f};
  int64_t Bias_dims[] = {2};
  AbstractTensorHandlePtr Bias;
  {
    AbstractTensorHandle* Bias_raw;
    status_ = TestTensorHandleWithDims<float, TF_FLOAT>(
        immediate_execution_ctx_.get(), Bias_vals, Bias_dims, 1, &Bias_raw);
    ASSERT_EQ(errors::OK, status_.code()) << status_.message();
    Bias.reset(Bias_raw);
  }