# CuDNN: https://docs.nvidia.com/deeplearning/sdk/cudnn-install/index.html#ubuntu-network-installation
libcudnn8=8.9.6.50-1+cuda12.2
libcudnn8-dev=8.9.6.50-1+cuda12.2

# This can be removed once NVIDIA publishes a cuda-12.3.2 Docker image.
# For now it ensures that we install at least version 12.3.107 of PTXAS,
# since 12.3.103 has a bug.

# Libtensorflow archives) to GCS buckets. TFCI Mac VMs need to authenticate as
# a service account that has the right permissions to be able to do so.
set +x
if [[ -n "${GOOGLE_APPLICATION_CREDENTIALS:-}" ]]; then
  # Python 3.12 removed the module `imp` which is needed by gcloud CLI so we set
  # `CLOUDSDK_PYTHON` to Python 3.11 which is the system Python on TFCI Mac
  # VMs.
  export CLOUDSDK_PYTHON=$(which python3.11)

import distutils as _distutils
import importlib
import inspect as _inspect
import os as _os
import site as _site
import sys as _sys

# Do not remove this line; See https://github.com/tensorflow/tensorflow/issues/42596
from tensorflow.python import pywrap_tensorflow as _pywrap_tensorflow  # pylint: disable=unused-import
from tensorflow.python.tools import module_util as _module_util

     *
     *    dA = U
     *    dbias = reduceSum(U, dims = channel_dim)
     */

    AbstractTensorHandle* upstream_grad = grad_outputs[0];
    DCHECK(upstream_grad);

    // Recover data format from forward pass for gradient.
    std::string data_format;
    TF_RETURN_IF_ERROR(forward_attrs_.Get("data_format", &data_format));

    // Grad for A
    grad_inputs[0] = upstream_grad;

    std::string& summary) const {
  Status status;
  AbstractTensorPtr resolved(
      // TODO(allenl): Resolve should be const, and the caches that get updated
      // marked mutable.
      const_cast<ImmediateExecutionTensorHandle*>(this)->Resolve(&status));
  if (!status.ok()) {
    return status;
  }
  summary = resolved->SummarizeValue();
  return absl::OkStatus();
}

# Advanced: Use Remote Build Execution (RBE) (internal developers only)
#
#   RBE dramatically speeds up builds and testing. It also gives you a
#   public URL to share your build results with collaborators. However,
#   it is only available to a limited set of internal TensorFlow developers.
#
#   RBE is incompatible with local caching, so you must remove

        "Tensor expected to have scalar shape found rank: ", shape.dims());
  }
  return absl::OkStatus();
}

TEST_P(UnifiedAPI, TestTensorShapeScalar) {
  if (UseFunction() && UseMlir()) {
    // TODO(b/173074167): Remove this.
    GTEST_SKIP() << "MlirTensor::Shape is not implemented yet.";
  }
  AbstractContextPtr ctx;
  {
    AbstractContext* ctx_raw = nullptr;
    Status s =

  if [[ "$TFCI_DOCKER_REBUILD_UPLOAD_ENABLE" == 1 ]]; then
    docker push "$TFCI_DOCKER_IMAGE"
  fi
fi

# Keep the existing "tf" container if it's already present.
# The container is not cleaned up automatically! Remove it with:
# docker rm tf
if ! docker container inspect tf >/dev/null 2>&1 ; then
  # Pass all existing TFCI_ variables into the Docker container
  env_file=$(mktemp)
  env | grep ^TFCI_ > "$env_file"

    ],
)

tf_cuda_cc_test(
    name = "custom_gradient_test",
    size = "small",
    srcs = [
        "custom_gradient_test.cc",
    ],
    args = ["--heap_check="],  # TODO(b/174752220): Remove
    tags = tf_cuda_tests_tags(),
    deps = [
        "//tensorflow/c:tf_status_helper",
        "//tensorflow/c/eager:abstract_context",
        "//tensorflow/c/eager:c_api",