This policy protects our limited review budget, allowing us to invest our attention wisely.

A pull request is not a finished product—it is the start of a conversation, and an implied contract: we promise to provide thoughtful feedback, and we expect the contributor to engage with us in return.

## What we expect from contributors

    env:
      PIP_NO_CACHE_DIR: 1
    defaults:
      run:
        shell: bash
    container: ${{ (contains(inputs.runner, 'linux-x86') && 'us-docker.pkg.dev/ml-oss-artifacts-published/ml-public-container/ml-build:latest') ||
                   (contains(inputs.runner, 'linux-arm64') && 'us-docker.pkg.dev/ml-oss-artifacts-published/ml-public-container/ml-build-arm64:latest') ||
                   (contains(inputs.runner, 'windows-x86') && null) }}

# ==============================================================================
TFCI_DOCKER_ENABLE=1
if [[ "$TFCI_GITHUB_ACTIONS" == "true" ]]; then
  TFCI_DOCKER_ENABLE=0
fi
TFCI_DOCKER_PULL_ENABLE=1
TFCI_DOCKER_IMAGE="us-docker.pkg.dev/ml-oss-artifacts-published/ml-public-container/ml-build-windows@sha256:023fa901fda70be8890b05689d0a667287692c1f59d0abe4add7208ec5a12707"
TFCI_BAZEL_BAZELRC_ARGS="--output_user_root=C:/x"

The build cache has anonymous read access, so you don't need to authenticate in order to use it. You can use a different build cache node by specifying `-Ddevelocity.edge.discovery=false -DcacheNode=us` for a build cache node in the US or `-Ddevelocity.edge.discovery=false -DcacheNode=au` for a build cache node in Australia.

Backports are typically handled by maintainers; 
however, after agreement you may open a separate backport PR to the target maintenance branch.
-->

Following this checklist to help us incorporate your
contribution quickly and easily:

- [ ] Your pull request should address just one issue, without pulling in other changes.

    return accessibleObject.getDeclaredAnnotations();
  }

  // We ought to be able to implement GenericDeclaration instead its parent AnnotatedElement.
  // That would give us this method declaration. But for some reason, implementing
  // GenericDeclaration leads to weird errors in Android tests:
  // IncompatibleClassChangeError: interface not implemented

# ==============================================================================
TFCI_DOCKER_ENABLE=1
if [[ "$TFCI_GITHUB_ACTIONS" == "true" ]]; then
  TFCI_DOCKER_ENABLE=0
fi
TFCI_DOCKER_PULL_ENABLE=1
TFCI_DOCKER_IMAGE="us-docker.pkg.dev/ml-oss-artifacts-published/ml-public-container/ml-build-windows@sha256:023fa901fda70be8890b05689d0a667287692c1f59d0abe4add7208ec5a12707"
TFCI_BAZEL_BAZELRC_ARGS="--output_user_root=C:/x"

            if (context.payload.pull_request && context.payload.pull_request.head.repo.fork) {
                core.setOutput('sys-prop-args', '-DagreePublicBuildScanTermOfService=yes -Ddevelocity.edge.discovery=false -DcacheNode=us --scan')
            } else {
                core.setOutput('sys-prop-args', '')
            }
      - name: Setup Gradle
        uses: gradle/actions/setup-gradle@v6
        with:

    visibility = ["//visibility:public"],
)

# By default, XLA GPU is compiled into tensorflow when building with
# --config=cuda even when `with_xla_support` is false. The config setting
# here allows us to override the behavior if needed.
config_setting(
    name = "no_xla_deps_in_cuda",
    define_values = {"no_xla_deps_in_cuda": "true"},
    visibility = ["//visibility:public"],
)

             * corresponds to the least significant nonzero byte in lw ^ rw, since lw and rw are
             * little-endian. Long.numberOfTrailingZeros(diff) tells us the least significant
             * nonzero bit, and zeroing out the first three bits of L.nTZ gives us the shift to get
             * that least significant nonzero byte.
             */
            int n = Long.numberOfTrailingZeros(lw ^ rw) & ~0x7;