tensorflow/tensorflow
An Open Source Machine Learning Framework for Everyone
How tensorflow is put together
TensorFlow is an end-to-end open source machine learning platform, primarily written in C++ with a Python front-end API (imported as `tf`), plus C, Go, Java, and JavaScript bindings. The Python API is generated from templates (api_template.__init__.py) that wire up lazy-loaded Keras integration, plugin loading from site-packages/tensorflow-plugins, and the pywrap_tensorflow native binding. Builds are orchestrated by Bazel (WORKSPACE/MODULE.bazel/.bazelrc/configure.py), with official CI in ci/official producing wheels for Linux/macOS/Windows across CPU/GPU configurations using NVIDIA CUDA 12/13 pip packages. Data flows from user Python/C++/Go code through session/graph APIs into the core runtime kernels and device backends.
Languages
Frameworks
Infrastructure
Major components
Python API layer (tensorflow/api_template.__init__.py)
Generates the top-level `tf` module, enabling TF2 behavior, lazy-loading Keras v2/v3, and loading dynamic kernel/plugin libraries from site-packages.
C API (tensorflow/c)
Stable C ABI (c_api.cc, tf_tensor, tf_status, ops, kernels) that underpins all language bindings including Go, Java, and Python's pywrap_tensorflow.
Go bindings (tensorflow/go)
Go wrappers over the C API providing Graph, Session, Tensor, SavedModel loading, and PartialRun incremental execution.
Core runtime (tensorflow/core)
The central graph executor, kernels, protobuf definitions (ConfigProto), and local rendezvous implementation.
Compiler stack (tensorflow/compiler)
XLA/MLIR-based compilation of graphs to optimized device code.
TensorFlow Lite (tensorflow/lite)
Lightweight on-device inference interpreter with arena planner, builtin op registry, and iOS/Android tooling.
Build configuration (configure.py, WORKSPACE, MODULE.bazel)
Interactive script that probes CUDA/Python environments and writes .tf_configure.bazelrc driving Bazel builds.
Official CI (ci/official)
Shell scripts and containerized environments for building/testing/publishing wheels across platforms, including NVIDIA requirements management.
A busy week dominated by community bug fixes plus major XLA compiler milestones, including Bzlmod as the default build and a new default CPU compilation pipeline.
5 weeksWeek of 2026-08-24
Week of 2026-08-24latest
A busy week dominated by community bug fixes plus major XLA compiler milestones, including Bzlmod as the default build and a new default CPU compilation pipeline.
Week of 2026-08-17
A high-volume week of mostly internal cleanups and automated changes, punctuated by targeted fixes to GPU compilers, dynamic-shape handling, and Python free-threading safety.
Week of 2026-08-10
The busiest week in the window, blending a wave of user-reported correctness and security fixes with notable TPU/GPU compiler capabilities like native Pallas convolutions.
Week of 2026-08-03
An exceptionally quiet week with a single small change to NumPy compatibility.
Week of 2026-07-20
Another single-commit week, focused on preventing a crash when encoding empty images to PNG.
Over the past few weeks the TensorFlow/XLA project has been in a heavy stabilization phase: dozens of community-submitted bug fixes landed across Python APIs, TFLite conversion, and numerical edge cases (like overflows, crashes on empty inputs, and wrong results). Alongside that, the GPU and CPU compiler teams shipped significant infrastructure work — enabling a new CPU compilation pipeline by default, moving XLA's build system to Bzlmod, and adding new GPU features like Triton scan-fusion rewriting and CuTe DSL support. Quiet weeks with only one or two commits were interleaved with very busy ones of 86–169 commits, showing bursts of merged work rather than steady output.
Week by week
2026-08-24A busy week dominated by community bug fixes plus major XLA compiler milestones, including Bzlmod as the default build and a new default CPU compilation pipeline.latest6 changes
Feature
Triton block-level rewriting for GPU scan fusions
The GPU compiler can now use Triton to optimize 'scan'-style operations at the block level, improving performance for those patterns.
Chore
XLA switches its default build to Bzlmod
XLA now uses Google's newer dependency-management system by default, modernizing how the project builds.
Feature
New xtile compilation pipeline enabled by default on CPU
A redesigned CPU code-generation pipeline is now the standard path, replacing the older flow.
Fix
Digamma returns -inf for input 0
A math operation that previously misbehaved on zero input now returns the mathematically correct result.
Fix
More stable variance in batch normalization
Batch norm now uses a two-pass calculation so it no longer produces impossible negative variances with large input values.
Feature
CuTe DSL handlers wired into Jet Engine
New NVIDIA CuTe-based operations were registered and linked into XLA's Jet Engine, expanding supported GPU kernels.
2026-08-17A high-volume week of mostly internal cleanups and automated changes, punctuated by targeted fixes to GPU compilers, dynamic-shape handling, and Python free-threading safety.6 changes
Fix
Fix crash and wrong results in dynamic reduce-window on CPU
Scan-like reductions over dynamically sized data no longer crash or produce incorrect output.
Fix
Fix byte-vs-bit shift bug in GPU PadToStatic
Corrected a unit-mixup in how dynamic sizes are stored on GPU that could corrupt metadata.
Fix
Debug event writers made safe for free-threaded Python
TensorFlow's logging writers are now thread-safe for users running the new free-threaded builds of Python.
Fix
Handle 0-D inputs in numpy cross
Cross-product operations on scalar-shaped inputs now behave correctly instead of erroring out.
Feature
New CollectiveReduce HLO operation for NVIDIA GPUs
A new collective reduction operation was added to the compiler's vocabulary for NVIDIA hardware.
Fix
Fix DynamicUpdateSlice inside loop fusions on CPU
Updating a slice of an array within a fused loop now emits correct code on CPU.
2026-08-10The busiest week in the window, blending a wave of user-reported correctness and security fixes with notable TPU/GPU compiler capabilities like native Pallas convolutions.6 changes
Fix
Integer overflow fix in ExtractImagePatches
Patching operations on very large images could overflow and crash; sizes are now computed safely and bad inputs return a clear error instead.
Feature
Native convolutions in JAX Pallas for TPUs
Developers writing custom TPU kernels can now call convolution primitives directly from Pallas.
Fix
tf.tuple() works with variables under jit_compile
Grouping TensorFlow variables into tuples no longer throws an error when just-in-time compilation is on.
Fix
Infinite loop fixed in the memory allocator
The GPU memory allocator could get stuck growing allocations for small requests; that backpedal loop is gone.
Fix
Type confusion and memory leak fixed in RPC ops
Two security-adjacent issues in remote-operation handling were patched.
Refactor
Thread-safe latency-hiding scheduler
The GPU scheduling simulator can now run analyses in parallel, speeding up compile-time optimization.
2026-08-03An exceptionally quiet week with a single small change to NumPy compatibility.1 change
Refactor
Use static last dimension in numpy cross when available
The cross-product operation now takes advantage of known dimension sizes for slightly more efficient execution.
2026-07-20Another single-commit week, focused on preventing a crash when encoding empty images to PNG.1 change
Fix
EncodePng no longer fails on empty images
Saving a zero-size image as PNG previously triggered an internal check failure; it is now handled gracefully.
Dependencies and code review
Dependency advisories
Security Watch
No known advisories across 0 scanned dependencies.
No known advisories in the scanned dependencies.
Code review
No concrete code-level issues confirmed.
Get this report every week for your repos.
GitZoid learns each repo, reports what changed, and flags what needs attention. One flat price for the whole team.
$19 a month, flat · First 10 outputs free · No card required