Configuration

Configuration knobs you'll meet most often.

./configure (configure.py)

The script asks the questions and writes .tf_configure.bazelrc. The same answers can be set as environment variables for unattended builds:

Env var	Purpose
PYTHON_BIN_PATH	Path to the Python interpreter to build against.
TF_NEED_CUDA	1 to enable CUDA support.
TF_CUDA_PATHS	Comma-separated list of CUDA install roots.
TF_CUDA_VERSION	CUDA version (e.g. 12.5).
TF_CUDNN_VERSION	cuDNN version.
TF_CUDA_COMPUTE_CAPABILITIES	Compute capabilities to compile for, e.g. 8.0,9.0.
TF_NEED_ROCM	1 for ROCm/AMD GPU support.
TF_NEED_TENSORRT	1 to integrate TensorRT.
TF_NEED_MPI	1 for MPI distributed support.
TF_DOWNLOAD_CLANG	1 to use Clang shipped via Bazel.
CC_OPT_FLAGS	Optimization flags (default -Wno-sign-compare -mavx).
TF_SET_ANDROID_WORKSPACE	1 to set up Android NDK paths.

Source: configure.py at the repo root (~50 KB).

.bazelrc configs

The .bazelrc (~60 KB) defines named configurations you pass via --config=<name> to bazel build/test:

Config	Effect
--config=opt	Release optimization flags
--config=cuda	Enable CUDA build
--config=cuda_clang	Use Clang as CUDA compiler
--config=rocm	Enable ROCm
--config=mkl	Use Intel MKL/oneDNN paths
--config=mkl_aarch64	MKL on aarch64
--config=tflite_with_xnnpack	Build TFLite with XNNPACK delegate
--config=monolithic	Statically link runtime into single binary
--config=v1	Build with TF1 API surface
--config=asan / msan / tsan	Sanitizer builds
--config=dynamic_kernels	Make op kernels dynamically loadable
--config=android_arm / android_arm64 / android_x86_64	Android cross-compiles
--config=ios	iOS cross-compile

Plus per-CPU configs (--config=avx2_linux, etc.). The full list is in .bazelrc.

Runtime environment variables

Variables that influence behaviour at runtime (not build time):

Var	Effect
TF_CPP_MIN_LOG_LEVEL	Suppress logs at or below severity (0 info, 1 warn, 2 err, 3 none).
TF_CPP_VMODULE	Per-file VLOG verbosity (e.g. executor=2,direct_session=2).
CUDA_VISIBLE_DEVICES	NVIDIA-driver level GPU masking; respected by TF.
TF_FORCE_GPU_ALLOW_GROWTH	If true, BFC allocator grows on demand instead of pre-reserving.
TF_GPU_ALLOCATOR	cuda_malloc_async to use the async allocator.
TF_USE_TFRT	Switch some paths to TFRT.
TF_USE_MLRT	Use the MLRT runtime (a TFRT variant).
TF_NUM_INTEROP_THREADS	Inter-op pool size.
TF_NUM_INTRAOP_THREADS	Intra-op pool size.
TF_DETERMINISTIC_OPS	Force deterministic kernels where possible.
TF_DUMP_GRAPH_PREFIX	Where to dump pre/post-Grappler graphs.
XLA_FLAGS	Pass-through to XLA (e.g. --xla_dump_to=...).
TF_ENABLE_ONEDNN_OPTS	Toggle oneDNN code paths in CPU kernels.
TF_USE_CUDNN_DETERMINISTIC	Force deterministic cuDNN paths.
TF_DATA_EXPERIMENT_OPT_IN	Comma list of experimental tf.data optimizations to enable.

Use tf.config.experimental.* Python APIs to query and set most of these at runtime where applicable.

Python-side runtime config

The tf.config module:

• tf.config.list_physical_devices("GPU") — see what TF detects.
• tf.config.experimental.set_memory_growth(device, True) — equivalent to TF_FORCE_GPU_ALLOW_GROWTH=1.
• tf.config.threading.set_inter_op_parallelism_threads(n) / intra_op_parallelism_threads.
• tf.config.optimizer.set_jit(True) — enable XLA auto-clustering.
• tf.config.optimizer.set_experimental_options({...}) — toggle individual Grappler passes.
• tf.debugging.set_log_device_placement(True) — print every op's device.

Useful build targets

Target	Output
//tensorflow/tools/pip_package:wheel	Pip wheel.
//tensorflow:libtensorflow.so	C API shared lib.
//tensorflow:libtensorflow_cc.so	C++ API shared lib.
//tensorflow/lite:libtensorflowlite.so	TFLite shared lib.
//tensorflow/lite/java:tensorflow-lite	Android AAR.
//tensorflow/lite/tools/benchmark:benchmark_model	TFLite benchmark binary.

Related

• how-to-contribute/tooling — what ./configure and Bazel actually do.
• how-to-contribute/getting-started — full build walkthrough.