"//tensorflow/c/eager:tfe_tensorhandle_internal",
        "@com_google_absl//absl/strings",
        "@com_google_absl//absl/types:optional",
        "@com_google_absl//absl/types:variant",
    ],
)

cc_library(
    name = "parallel_device_lib",
    srcs = [":lib_sources"],
    hdrs = [":lib_headers"],
    visibility = ["//tensorflow:internal"],
    deps = [
        "//tensorflow/c:c_api",

};

using OpPtr = std::unique_ptr<TFE_Op, OpDeleter>;

using MaybeParallelTensorOwned =
    absl::variant<std::unique_ptr<ParallelTensor>, TensorHandlePtr>;

using MaybeParallelTensorUnowned =
    absl::variant<ParallelTensor*, TFE_TensorHandle*>;

// A ParallelDevice on its own is not registered with a TFE_Context, and so has

#include <memory>
#include <optional>
#include <string>
#include <utility>
#include <vector>

#include "absl/types/optional.h"
#include "absl/types/span.h"
#include "absl/types/variant.h"
#include "tensorflow/c/c_api.h"
#include "tensorflow/c/eager/c_api.h"
#include "tensorflow/c/eager/c_api_experimental.h"
#include "tensorflow/c/eager/tfe_op_internal.h"
#include "tensorflow/c/safe_ptr.h"

        "//tensorflow/core/platform:refcount",
    ] + if_not_mobile([
        "//tensorflow/core/kernels:tensor_list",
        "//tensorflow/core/kernels:tensor_list_util",
        "//tensorflow/core/kernels:variant_ops_util",
        "//tensorflow/core/kernels/data:optional_ops_util",
        "//tensorflow/core/platform:abi",
    ]),
    alwayslink = 1,
)

tf_cuda_library(
    name = "ops",
    srcs = [

// Adds operations to compute the partial derivatives of sum of `y`s w.r.t `x`s,
// i.e., d(y_1 + y_2 + ...)/dx_1, d(y_1 + y_2 + ...)/dx_2...
// This is a variant of TF_AddGradients that allows to caller to pass a custom
// name prefix to the operations added to a graph to compute the gradients.
//
// `dx` are used as initial gradients (which represent the symbolic partial

  // REQUIRES: "v2_reader_ != nullptr && v2_reader_.status().ok()".
  std::pair<std::unique_ptr<TensorSliceReader::VarToShapeMap>,
            std::unique_ptr<TensorSliceReader::VarToDataTypeMap> >
  BuildV2VarMaps();

  // Invariant: exactly one of "reader_" and "v2_reader_" is non-null.
  std::unique_ptr<TensorSliceReader> reader_;
  std::unique_ptr<BundleReader> v2_reader_;

  std::unique_ptr<TensorSliceReader::VarToShapeMap> var_to_shape_map_;

        "//tensorflow/c/experimental/filesystem:filesystem_interface",
        "@com_github_googlecloudplatform_google_cloud_cpp//:storage_client",
        "@com_google_absl//absl/strings",
        "@com_google_absl//absl/types:variant",
    ],
)

cc_library(
    name = "gcs_helper",
    srcs = ["gcs_helper.cc"],
    hdrs = ["gcs_helper.h"],
    linkstatic = 1,
    deps = [
        "//tensorflow/c:env",
    ],
)

    // Ownership of `aid` transferred to CallBackwardFunction below.
    Gradient* aid;
    if (output_tensor.GetDType() == tensorflow::DT_VARIANT) {
      // Note: Needs to be zeros rather than ones since there's currently no
      // ones_like for variants.
      aid = output_tensor.ZerosLike();
    } else {
      // TODO(allenl): Figure out why using zeros_like everywhere causes issues

  "$python" -c 'import sys; import tensorflow as tf; sys.exit(0 if "keras" in tf.keras.__name__ else 1)'
fi
# VERY basic check to ensure the [and-cuda] package variant is installable.
# Checks TFCI_BAZEL_COMMON_ARGS for "gpu" or "cuda", implying that the test is
# relevant. All of the GPU test machines have CUDA installed via other means,

    return tensorflow::ops::Identity(ctx, input, output, name);
  } else if (!DataTypeIsComplex(BaseType(dtype)) &&
             BaseType(dtype) != DT_VARIANT) {
    return errors::InvalidArgument(
        "Expected numeric or variant tensor, got dtype ", dtype);
  }
  return tensorflow::ops::Conj(ctx, input, output, name);
}

class AddGradientFunction : public GradientFunction {
 public: