TensorFlow

TensorFlow is an end-to-end open source platform for machine learning. The repository at github.com/tensorflow/tensorflow contains the core runtime, kernel library, language frontends (Python, C++, Java, Go), the XLA/MLIR compiler stack, the TensorFlow Lite mobile/embedded runtime, and the bundled (legacy) Keras frontend.

What it does

TensorFlow lets users define computations as dataflow graphs of operations over multidimensional tensors, then executes those graphs on CPUs, GPUs, TPUs, and other accelerators. Users typically write Python (the dominant frontend), but the same graph can be saved and executed from C++, Java, Go, JavaScript-via-tfjs-converter, or run on mobile devices via TensorFlow Lite. The platform covers training (large-scale, distributed, possibly on TPU pods), inference (eager, graph, ahead-of-time-compiled, or quantized for mobile), and the surrounding tooling (TensorBoard logging via tf.summary, SavedModel serialization, tf.data input pipelines, tf.distribute strategies).

Repository scope

This repo contains:

• Core runtime (tensorflow/core/) — C++ implementation of graphs, executors, devices, eager runtime, distributed master/worker, protobufs, kernels.
• Compiler stack (tensorflow/compiler/) — XLA bridge (tf2xla), JIT/AOT compilation (jit, aot), MLIR-based passes, and TensorRT integration.
• Language frontends (tensorflow/python/, tensorflow/cc/, tensorflow/c/, tensorflow/java/, tensorflow/go/, tensorflow/js/).
• TensorFlow Lite (tensorflow/lite/) — interpreter, kernels, delegates, converter, micro runtime, iOS/Android/Swift/ObjC bindings.
• Tooling (tensorflow/tools/, ci/, configure.py) — pip packaging, docs generation, CI scripts, the project's ./configure step, build defs.
• Third-party vendor sources (third_party/) — Bazel rules and patches for Eigen, gRPC, protobuf, ICU, MKL, etc.

How users interact with it

graph LR
    User[User Python script]
    PyAPI[tf.* Python API]
    CppRuntime[Core C++ runtime]
    Kernels[Op kernels - CPU/GPU/TPU]
    XLA[XLA / MLIR compiler]
    SavedModel[SavedModel files]
    Lite[TF Lite interpreter]

    User --> PyAPI
    PyAPI -- pybind11 / SWIG --> CppRuntime
    CppRuntime --> Kernels
    CppRuntime -- optional --> XLA
    XLA --> Kernels
    PyAPI -- export --> SavedModel
    SavedModel --> CppRuntime
    SavedModel -- convert --> Lite

• Training a model: import tensorflow as tf, build a tf.keras.Model or tf.function-decorated function, call model.fit(...) or run training loop with tf.distribute.Strategy.
• Serving a model: export with tf.saved_model.save(...) and load from C++ (SavedModelBundle), TF Serving, or convert via TFLiteConverter for mobile.
• Embedded inference: convert to a .tflite flatbuffer and run on Interpreter (C++/Java/Swift/Obj-C/Python).

Where to start

• New contributor → how-to-contribute
• Want to build TF from source → getting-started
• Want the high-level component map → architecture
• Need term definitions (Op, Kernel, Device, Rendezvous, Resource) → glossary
• Want a feel for the codebase → by-the-numbers, lore, fun-facts

Quick links inside this wiki

• Lens pages: apps, systems, compilers, features
• Reference: configuration, data models, dependencies
• Ownership: maintainers

Licensing

Apache License 2.0 (LICENSE). The project is governed by Google with public contributions accepted via PR. The PR workflow is unusual: contributions are mirrored into Google's internal monorepo via a tool called Copybara, then re-exported. See development-workflow.