Dependencies

External libraries this repo pulls in. Two flavours: Python (pip) and C++ (Bazel/third_party/).

Python dependencies

Pip dependencies are pinned per supported Python version:

requirements_lock_3_10.txt
requirements_lock_3_11.txt
requirements_lock_3_12.txt
requirements_lock_3_13.txt
requirements_lock_3_14.txt

Each is ~70 KB and contains every transitive package with sha256 hashes. The list is regenerated by the Bazel rule //ci/official:requirements.

Notable runtime deps (subset; full list in the lockfile):

• numpy — required pin per Python version.
• absl-py — Google's absl/Python.
• protobuf — runtime for the generated Python protos.
• flatbuffers — runtime for the TFLite schema.
• gast — used by AutoGraph for AST manipulation.
• astunparse — AST → source.
• wrapt — decorator helpers.
• typing-extensions — backports.
• ml_dtypes — bfloat16 and other ML-specific dtypes.
• keras — Keras 3 (the standalone) is also a runtime dependency; it co-exists with the bundled v2 Keras snapshot in tensorflow/python/keras/.
• tensorboard — only as an optional runtime dep; TensorBoard itself ships separately.
• grpcio, grpcio-tools — gRPC bindings.

C++ / Bazel dependencies

Vendored under third_party/ and pulled in by Bazel rules in WORKSPACE/MODULE.bazel. Each vendored project usually has:

• A WORKSPACE.bazel snippet that downloads the source tarball.
• One or more *.BUILD files that override the upstream's build rules.
• Optional patches to fit TF's build conventions.

Notable third-party libraries:

Project	Where in third_party/	What for
Eigen	third_party/eigen3/	Header-only linear-algebra; the CPU kernels' workhorse.
Abseil	third_party/absl/	C++ utilities (status, strings, hash, time).
Protobuf	third_party/protobuf/	Cross-process serialization.
gRPC	(via bazel_deps)	RPC.
MKL / oneDNN	third_party/mkl_dnn/	Optimised CPU primitives on Intel.
ICU	third_party/icu/	Unicode (TF strings).
NCCL	third_party/nccl/	Multi-GPU collectives.
CUDA / cuDNN	third_party/gpus/	GPU SDK glue.
TensorRT	third_party/tensorrt/	NVIDIA inference accelerator.
NumPy headers	third_party/py/	C-API wrappers.
FlatBuffers	third_party/flatbuffers/	TFLite schema runtime.
ml_dtypes	third_party/ml_dtypes/	bfloat16, float8, etc.
ruy	third_party/ruy/	Optimised matmul library used by TFLite.
XNNPACK	(referenced by TFLite xnnpack delegate)	CPU SIMD acceleration for TFLite.
xla	third_party/xla/	The XLA compiler (separated out to openxla/xla).
stablehlo	third_party/stablehlo/	StableHLO MLIR dialect.
LLVM/MLIR	third_party/llvm/	Pulled via bazel-deps; the compiler infra TF MLIR rides on.
Triton	third_party/triton/	OpenAI's Triton GPU compiler (used by some XLA paths).
gast/astunparse	(Python, via lockfile)	AutoGraph backbone.

Run ls third_party/ to see the full list (~50 entries).

Update flow

• Python: edit the requirements input and regenerate the lockfiles. Bot PRs from Dependabot land regularly (see by-the-numbers).
• C++: typically a "version bump" PR that updates the URL/SHA in workspace2.bzl (or MODULE.bazel) and re-runs the patch sets in third_party/<name>/.

Why third_party/xla/ exists when XLA is a separate repo

XLA was extracted into openxla/xla in 2023 for governance reasons (multi-vendor). This repo still depends on it, and the Bazel rules in third_party/xla/ express that dependency. The same pattern is used for StableHLO (which has its own repo) and a few other former-internal modules.

Related

• how-to-contribute/tooling — how Bazel resolves these.
• reference/configuration — runtime feature flags that gate optional deps (TensorRT, MKL, ROCm).