Tooling

The build system, code generators, and supporting scripts.

Bazel

Bazel is the only blessed build system. Important pieces:

• WORKSPACE and MODULE.bazel — repo-level dependencies. The transition from WORKSPACE to bzlmod is partial; both files are still active.
• .bazelrc (~60 KB) — the project's default flag set. Defines --config=cuda, --config=rocm, --config=xla_cpu, --config=tflite_with_xnnpack, --config=asan, --config=msan, --config=tsan, --config=opt, --config=mkl.
• .bazelignore and .bazelversion.
• tensorflow/tensorflow.bzl — the project's central macro library (~130 KB). It defines tf_cc_library, tf_kernel_library, tf_py_test, tf_proto_library, tf_cuda_cc_test, etc. Almost every BUILD file in the repo calls into it.
• tensorflow/tensorflow.default.bzl, tensorflow/tf_version.bzl, tensorflow/strict.default.bzl, tensorflow/pytype.default.bzl — auxiliary toggles.

For day-to-day work, bazelisk (any version) is fine — it reads .bazelversion and downloads the right Bazel automatically.

./configure

The top-level ./configure script is a thin wrapper around configure.py (~50 KB). It writes .tf_configure.bazelrc based on:

• Python interpreter path (PYTHON_BIN_PATH).
• CUDA support (TF_NEED_CUDA, TF_CUDA_PATHS, TF_CUDA_VERSION, TF_CUDNN_VERSION).
• ROCm support (TF_NEED_ROCM).
• TensorRT support (TF_NEED_TENSORRT).
• MPI support, MKL support, monolithic build, build optimization flags, Android build settings.

For unattended builds, set the env vars and run ./configure non-interactively. CI uses pre-built Docker images that bake in answers (ci/official/).

Code generation

• Op wrappers — tensorflow/python/ops/gen_*.py. Generated by Bazel rule tf_gen_op_wrapper_py (in tensorflow.bzl) from OpDef registrations. The generator binary lives in tensorflow/python/framework/python_op_gen_main.cc.
• C++ ops — tensorflow/cc/ops/*.h,cc produced by tf_gen_op_wrappers_cc for each op family (tensorflow/core/ops/). Resulting headers expose MatMul(...), Add(...), etc., as a builder API for C++.
• Python API stubs — tensorflow/api_template.__init__.py and tensorflow/api_template_v1.__init__.py are templates. Bazel rule gen_api_init_files_py (tensorflow/tools/api/) walks @tf_export(...) decorators and produces the full tensorflow/_api/v2/ tree at build time.
• Protobufs — tensorflow/core/protobuf/*.proto, tensorflow/core/framework/*.proto. Generated *_pb2.py files for Python; C++ generated headers in bazel-bin/.
• FlatBuffers — tensorflow/lite/schema/schema.fbs, tensorflow/lite/schema/conversion_metadata.fbs. Generated *.h for C++ and *.py for Python.

Don't edit any gen_*.py or *_pb2.py by hand.

Lint

• C++: clang-format (config .clang-format).
• Python: pylint (config tensorflow/tools/ci_build/pylintrc, root symlink .pylintrc).
• Bazel: buildifier.
• Markdown: light style only; some directories run mdformat in CI.

Doc generation

• API docs are generated by tensorflow/tools/docs/.
• The generator emits Markdown the tensorflow/docs repo (separate) renders.
• tensorflow/lite/g3doc/ and tensorflow/python/keras/ (legacy) keep small in-tree doc trees.

CI scripts

• ci/official/ — the modern CI configurations and docker images. Subdirectories per platform (linux_x86, linux_arm64, macos_arm64, windows, …).
• ci/devinfra/ — internal scripts (some only meaningful inside Google).
• .github/workflows/ — GitHub Actions for PR-side checks (formatting, basic builds, license/CLA).

The tensorflow/tools/ directory has older but still-used helpers:

Path	Purpose
tensorflow/tools/pip_package/	Pip wheel build rule (//tensorflow/tools/pip_package:wheel).
tensorflow/tools/api/	API export and golden-file checks.
tensorflow/tools/docs/	Docs generator + tag scrubber.
tensorflow/tools/test/	CI helpers and benchmarks.
tensorflow/tools/git/	Git utilities used by versioncontrol_utils.

Related

• development-workflow — how CI gates changes.
• reference/configuration — the env vars surfaced by ./configure.