Glossary

Project-specific terms used throughout the codebase. Definitions are paired with the file you'll find the canonical implementation in.

Computation

• Tensor — a typed n-dimensional array with a DataType and TensorShape. Implementation in tensorflow/core/framework/tensor.h. The Python user-facing class is tf.Tensor (tensorflow/python/framework/ops.py); the eager-only subclass is EagerTensor.
• Op / Operation — a node in the dataflow graph. The schema (name, inputs, attrs) is an OpDef declared in tensorflow/core/ops/*.cc via REGISTER_OP(...). Python tf.Operation mirrors this in tensorflow/python/framework/ops.py.
• Kernel — a concrete implementation of an op for a particular (device_type, type_constraint). Kernels are subclasses of OpKernel (tensorflow/core/framework/op_kernel.h) and registered with REGISTER_KERNEL_BUILDER(...) in tensorflow/core/kernels/.
• Graph — tensorflow.Graph is a GraphDef plus runtime state. C++ types in tensorflow/core/graph/graph.h; GraphDef proto in tensorflow/core/framework/graph.proto.
• FuncGraph — a tf.Graph produced by tracing a @tf.function. Lives in tensorflow/python/framework/func_graph.py.
• ConcreteFunction — a FuncGraph specialized to specific input signatures. See tensorflow/python/eager/polymorphic_function/concrete_function.py.
• HLO — XLA's High-Level Operations IR. Produced by tf2xla (tensorflow/compiler/tf2xla/) and consumed by the XLA compiler in the openxla/xla repo.

Runtime

• Session — the legacy graph-mode execution context (tf.compat.v1.Session). Implementation in tensorflow/core/common_runtime/direct_session.cc and tensorflow/core/distributed_runtime/master_session.cc.
• Executor — runs a partitioned graph asynchronously. tensorflow/core/common_runtime/executor.cc.
• Device — abstraction over CPU, GPU, TPU. Base in tensorflow/core/framework/device.h; concrete devices in tensorflow/core/common_runtime/{threadpool_device.cc,gpu/} and tensorflow/core/tpu/.
• Allocator — Allocator/AllocatorAttributes (tensorflow/core/framework/allocator.h). The GPU uses BFC (tensorflow/core/common_runtime/bfc_allocator.h).
• Resource / ResourceMgr — long-lived state attached to a device (variables, hashtables, iterators). tensorflow/core/framework/resource_mgr.h.
• Rendezvous — channel for sending tensors between devices/hosts in a graph. tensorflow/core/framework/rendezvous.h.
• Eager — the default mode that runs ops immediately. Runtime under tensorflow/core/common_runtime/eager/; Python in tensorflow/python/eager/.
• Function library / FunctionDef — serialised function graphs callable as ops. tensorflow/core/framework/function.proto, function.h.

Graph optimization & compilation

• Grappler — graph-level optimizer (tensorflow/core/grappler/). Runs constant folding, layout, arithmetic, memory, etc.
• XLA — Accelerated Linear Algebra compiler. Bridge here is tensorflow/compiler/tf2xla/; the compiler proper is in openxla/xla (referenced via third_party/xla).
• JIT (auto-clustering) — tensorflow/compiler/jit/ decides which subgraphs to fuse into XLA clusters at runtime.
• AOT — tfcompile ahead-of-time compiler under tensorflow/compiler/aot/ that turns a graph into a static binary.
• MLIR — LLVM project's multi-level IR. TF defines several dialects under tensorflow/compiler/mlir/: tf, tfl (TFLite), quant, stablehlo, etc.
• StableHLO — versioned IR shared with other ML frameworks; conversion lives in tensorflow/compiler/mlir/stablehlo/ and tensorflow/lite/stablehlo/.

Distribution

• Strategy / tf.distribute — Python API for multi-device training. tensorflow/python/distribute/.
• Master / Worker — gRPC roles for distributed graphs (tensorflow/core/distributed_runtime/).
• DTensor — typed distributed tensor (tensorflow/dtensor/ and tensorflow/python/dtensor/).
• Coordinator — coordinates parameter servers / workers in a ParameterServerStrategy cluster.
• Collective — collective communication ops (all-reduce, all-gather) implemented via NCCL (tensorflow/core/nccl/) or rings.

Data and IO

• tf.data — input pipeline framework. Python in tensorflow/python/data/, C++ in tensorflow/core/data/ and dataset kernels in tensorflow/core/kernels/data/.
• TFRecord — record-oriented binary file format used by tf.data.TFRecordDataset (tensorflow/core/lib/io/record_*).
• Iterator — a stateful resource that yields elements from a Dataset.
• Example / SequenceExample — protobuf message families used to serialize training records (tensorflow/core/example/).

Persistence

• Checkpoint — variable-state-only saving via tf.train.Checkpoint (tensorflow/python/checkpoint/, tensorflow/python/training/saving/).
• SavedModel — graph + variables + signatures, the default export format. tensorflow/python/saved_model/ and tensorflow/cc/saved_model/.
• MetaGraphDef — protobuf that describes a SavedModel's signatures, asset list, etc.

TFLite-specific

• Interpreter — TFLite runtime that loads a flatbuffer model and dispatches ops (tensorflow/lite/interpreter.h).
• Delegate — pluggable backend that accepts subgraphs (GPU, NNAPI, XNNPACK, CoreML, Hexagon). tensorflow/lite/delegates/.
• Builtin op — TFLite op enumerated in tensorflow/lite/builtin_ops.h.
• Custom op — user-supplied op registered with MutableOpResolver.
• TFLite Micro — the embedded variant, lives in a separate repo but historical code is in tensorflow/lite/micro/.

Project terminology

• A. Unique TensorFlower / TensorFlower Gardener — the bot author handle for changes mirrored from Google's internal monorepo. Roughly 60% of recent commits in master are attributed to this name; it is not a person.
• Copybara — the tool that mirrors PRs into Google's internal codebase and re-exports internal changes to GitHub. See development-workflow.
• Kokoro — Google's internal CI system that runs the post-merge tests.
• TFRT — TensorFlow Runtime, a newer asynchronous runtime fallback (tensorflow/core/tfrt/).
• tf.function — Python decorator that traces a Python function into a FuncGraph for graph-mode execution (tensorflow/python/eager/polymorphic_function/).
• AutoGraph — Python source-to-source converter that turns Pythonic control flow (if, while, for) into TF control-flow ops (tensorflow/python/autograph/).
• DTensor — distributed-tensor extension (tensorflow/dtensor/).
• Trackable — base class for objects that participate in checkpointing (tensorflow/python/trackable/).