Systems

Internal building blocks of the TensorFlow runtime. Things that aren't user-facing surfaces (those are apps) and aren't compilers (those are compilers).

Page	What it covers	Code under
Core runtime	OpKernel, Tensor, Graph, Executor, Device, Allocator, function lib	tensorflow/core/framework/, tensorflow/core/common_runtime/, tensorflow/core/graph/
Kernels and ops	The ~10 000 op kernels and how they're registered	tensorflow/core/kernels/, tensorflow/core/ops/
Eager execution	The default mode that runs ops immediately	tensorflow/core/common_runtime/eager/, tensorflow/python/eager/
Distributed runtime	gRPC master/worker, rendezvous, collectives	tensorflow/core/distributed_runtime/, tensorflow/core/nccl/
Grappler optimizer	Graph-level optimizations (constant folding, layout, memory)	tensorflow/core/grappler/
Profiler	TraceMe instrumentation, profiler service, TensorBoard integration	tensorflow/core/profiler/, tensorflow/python/profiler/
Distribution strategy	tf.distribute.Strategy, multi-device training plumbing	tensorflow/python/distribute/, tensorflow/dtensor/
TFRT	Newer asynchronous runtime fallback	tensorflow/core/tfrt/, tensorflow/core/runtime_fallback/
TPU support	TPU device + ops + topology	tensorflow/core/tpu/, tensorflow/python/tpu/

The directory tensorflow/core/ itself is the common ancestor for most of these. Its top-level subdirectories map roughly as:

tensorflow/core/
├── framework/                # OpKernel, Tensor, OpDef, Resource — the type system
├── common_runtime/           # Local execution: Executor, Placer, eager runtime
│   ├── eager/                # ↳ Eager-specific runtime
│   └── gpu/                  # ↳ GPU device + BFC allocator
├── distributed_runtime/      # Multi-host: master, worker, rendezvous, collectives
├── kernels/                  # ~10 000 op implementations
├── ops/                      # OpDef registrations + shape inference
├── graph/                    # Graph type, partitioning
├── grappler/                 # Graph-level optimizer
├── data/                     # tf.data dataset/iterator runtime ops + kernels
├── debug/                    # Tensor-level debugging support
├── platform/                 # OS abstraction: env, filesystem, logging
├── profiler/                 # Profiler service + TraceMe
├── protobuf/                 # Cross-process protos
├── public/                   # Public-ish C++ headers
├── runtime_fallback/         # TFRT ↔ TF runtime bridge
├── tfrt/                     # New async runtime
├── transforms/               # MLIR-based core transforms
├── ir/                       # MLIR `tf` dialect / TFG
├── tpu/                      # TPU
├── nccl/                     # NCCL collectives
├── summary/                  # tf.summary
├── lib/                      # Misc shared lib code (io, gtl, hash, ...)
├── util/                     # Utilities
└── ...

The next pages dive into the most-touched of these.