Weekly GitHub Report for Tensorflow

Thank you for subscribing to our weekly newsletter! Each week, we deliver a comprehensive summary of your GitHub project's latest activity right to your inbox, including an overview of your project's issues, pull requests, contributors, and commit activity.

Table of Contents

• I. News
  • 1.1. Recent Version Releases
  • 1.2. Other Noteworthy Updates
• II. Issues
  • 2.1. Top 5 Active Issues
  • 2.2. Top 5 Stale Issues
  • 2.3. Open Issues
  • 2.4. Closed Issues
  • 2.5. Issue Discussion Insights

• III. Pull Requests
  • 3.1. Open Pull Requests
  • 3.2. Closed Pull Requests
  • 3.3. Pull Request Discussion Insights

• IV. Contributors
  • 4.1. Contributors

I. News

1.1 Recent Version Releases:

The current version of this repository is v2.19.0

1.2 Version Information:

Released on March 5, 2025, TensorFlow version 2.19.0 introduces breaking changes to the tf.lite API, including the deprecation of tf.lite.Interpreter in favor of ai_edge_litert.interpreter and changes to certain C++ constants for improved API flexibility. Key updates also include runtime support for the bfloat16 data type in the tfl.Cast operation, alongside the discontinuation of separate libtensorflow package publishing, while still allowing unpacking from PyPI.

II. Issues

2.1 Top 5 Active Issues:

We consider active issues to be issues that that have been commented on most frequently within the last week. Bot comments are omitted.

1. [STAT:CONTRIBUTION WELCOME] [TYPE:BUG] [COMP:AUTOGRAPH] [TF 2.9] AutoGraph cannot handle python 3.10's structural pattern matching: This issue reports that TensorFlow's AutoGraph cannot properly transform Python 3.10's structural pattern matching (match/case statements), resulting in warnings and fallback to running the function as-is rather than converting it into a TensorFlow graph. The user highlights that while the code executes, AutoGraph fails to handle the new syntax, causing limitations in tracing and graph conversion, and requests support for this Python feature to be added.

   • The comments reveal that the issue is reproducible and acknowledged as a lack of support for structural pattern matching in AutoGraph; users discuss the nature of the warning, possible workarounds, and the importance of supporting this syntax. A contributor reports working on a fix with regression tests and opens a draft pull request to address the problem, while others note the slow pace of TensorFlow updates and encourage community contributions.
   • Number of comments this week: 3

2. [STAT:CONTRIBUTION WELCOME] [TYPE:BUG] [COMP:OPS] [TF 2.18] A heap oob write in TensorArray.write: This issue reports a heap out-of-bounds write occurring in the TensorArray.write operation when used inside a lambda layer of a TensorFlow model, causing the model service to crash during inference. The problem arises because a large index value, when cast to a signed integer, leads to incorrect resizing of an internal vector and subsequent invalid memory writes, highlighting a potential security vulnerability.

   • The comments clarify that the crash is due to improper handling of large or negative indices in the dynamic resizing logic, with a suggestion to add validation checks; a contributor volunteers to fix the bug by adding an explicit non-negative index check and a regression test, and later confirms that the issue is already fixed in the current master branch, which rejects invalid indices with an error instead of crashing.
   • Number of comments this week: 2

3. [TYPE:BUG] [COMP:OPS] [COMP:MKL] [TF 2.19] DepthwiseConv2dNative and related APIs core dump exclusively when TF_ENABLE_ONEDNN_OPTS='1': This issue reports a bug in TensorFlow where the DepthwiseConv2dNative API and related depthwise convolution APIs crash when given a 1D tensor input instead of the expected 4D tensor, but only when oneDNN optimizations are enabled via TF_ENABLE_ONEDNN_OPTS='1'. The problem arises because the oneDNN kernel does not validate the input tensor's rank, leading to a runtime crash instead of a proper shape-mismatch error, and this behavior has been confirmed and fixed in the latest TensorFlow nightly builds.

   • The comments confirm the issue was reproduced in TensorFlow 2.21 and nightly versions, a fix was proposed and merged, and the problem is now resolved with the API raising an appropriate InvalidArgumentError instead of crashing, as verified by users.
   • Number of comments this week: 2

4. [TYPE:BUG] [COMP:DATA] [2.21.0] [AWAITING PR MERGE] tf.data.FixedLengthRecordDataset aborts with std::bad_alloc on oversized buffer_size instead of raising an error: This issue reports that the tf.data.FixedLengthRecordDataset in TensorFlow 2.21.0 aborts the entire process with a std::bad_alloc error when given an excessively large buffer_size instead of raising a recoverable Python or TensorFlow exception. The problem arises because the C++ kernel only checks that buffer_size is non-negative but does not enforce an upper limit, causing an attempt to allocate an unreasonably large buffer and crashing the interpreter; a proposed fix adds a Python-level sanity check to reject oversized values with a clear error before reaching the kernel.

   • The comments confirm the root cause as missing upper-bound validation on buffer_size in the C++ kernel and describe a Python-only fix that adds a sanity check to raise a ValueError for excessive buffer sizes; testing verified the fix prevents the crash and maintains normal behavior, and a regression test was added with a forthcoming pull request.
   • Number of comments this week: 2

5. [TYPE:BUG] [COMP:OPS] [2.21.0] SavedModels containing DebugIdentity/DebugNanCount/DebugNumericSummary raw ops segfault at inference: This issue reports a segmentation fault occurring during inference when loading SavedModels that include TensorFlow debug raw operations such as DebugIdentity, DebugIdentityV3, DebugNanCount, and DebugNumericSummary. The problem is traced to an infinite recursion in the RecursiveCreateDir function within the debug IO utilities, caused by handling empty directory paths improperly, which leads to a stack overflow and crash.

   • The comments confirm reproduction of the crash on multiple debug ops with TensorFlow nightly builds, identify the root cause as infinite recursion due to missing base case for empty paths in RecursiveCreateDir, and propose adding a base case and regression test to fix the issue.
   • Number of comments this week: 2

2.2 Top 5 Stale Issues:

We consider stale issues to be issues that has had no activity within the last 30 days. The team should work together to get these issues resolved and closed as soon as possible.

As of our latest update, there are no stale issues for the project this week.

2.3 Open Issues

This section lists, groups, and then summarizes issues that were created within the last week in the repository.

Issues Opened This Week: 9

Summarized Issues:

• TFLiteConverter Fusion and Conversion Failures: The TFLiteConverter has issues with its fusion pass for chained MatMul operations producing a 'tfl.mul' operation with incompatible result types, causing MLIR verifier rejections and conversion failures without clear diagnostics. Additionally, conversion fails with an opaque error when Keras models contain multiple @tf.function-decorated methods, even if unused, leading to uninformative failure messages.
• [issues/124103, issues/124220]

• Incorrect Behavior of TensorFlow Math Functions on CPU: The tf.math.floor function incorrectly returns -0.0 instead of -1.0 for negative float32 subnormal inputs on CPU due to kernel flushing behavior, differing from NumPy's correct output. This bug affects numerical correctness for specific subnormal inputs on CPU platforms.
• [issues/124151]

• Segmentation Faults Due to Symbol Clashes with LLVM Libraries: TensorFlow wheels version 2.20+ cause segmentation faults when loaded with RTLD_GLOBAL alongside Triton 3.7+ due to symbol clashes involving statically linked LLVM libraries, resulting in crashes when both are used together. This regression impacts interoperability with Triton in newer TensorFlow versions.
• [issues/124205]

• Axis Index Handling Inconsistencies in tf.transpose with XLA: While tf.transpose correctly normalizes negative axis indices in eager execution, it fails with an InvalidArgumentError under XLA compilation because negative axes are not normalized during tf2xla conversion. This inconsistency causes errors when using JIT compilation with negative axis indices.
• [issues/124303]

• FlatBuffers Version Override Failures in TensorFlow Lite Build: The documented method to override the FlatBuffers version in TensorFlow Lite's CMake build is incorrect due to a wrong variable name, and even when fixed, the build enforces an exact FlatBuffers version via a static assertion, preventing successful overrides. This issue blocks building TensorFlow Lite with alternative FlatBuffers versions.
• [issues/124399]

• Control Dependency Redundancy and Over-Serialization in TensorFlow: The CreateControlDependencies function emits redundant, address-dependent control edges due to an incomplete reachability map, causing pruning to fail and resulting in over-serialization of concurrent collective operations. This deterministic defect leads to unnecessary serialization and potential performance degradation.
• [issues/124405]

• Inaccurate Executor Cost Estimation on aarch64 Linux: The executor's cost estimation on aarch64 Linux uses unscaled CNTVCT_EL0 timer cycles, which differ significantly from x86 TSC frequency, causing inaccurate classification of operations as expensive and leading to scheduling inefficiencies. This discrepancy affects performance optimization on ARM platforms.
• [issues/124406]

• Integer Overflow in TensorFlow Lite Reference Backend: Performing element-wise division of INT32_MIN by -1 in the TensorFlow Lite reference backend causes a signed integer overflow, resulting in undefined behavior and crashes due to exceeding the 32-bit integer limit. This bug leads to instability when handling specific edge-case integer operations.
• [issues/124409]

2.4 Closed Issues

This section lists, groups, and then summarizes issues that were closed within the last week in the repository. This section also links the associated pull requests if applicable.

Issues Closed This Week: 38

Summarized Issues:

• Crashes and Fatal Errors in TensorFlow Operations: Several issues report crashes or fatal CHECK failures in TensorFlow operations due to invalid tensor shapes, data types, or device configurations. These include segmentation faults, core dumps, and aborts triggered by shape mismatches, unsupported data types, or invalid tensor dimension indices, often without graceful error handling.
• issues/78994, issues/80331, issues/93391, issues/99894, issues/104796, issues/118340, issues/118345

• XLA Compilation and Execution Bugs: Multiple issues describe failures or incorrect behavior when using XLA compilation, including dimension size errors, cross-device resource access problems, inconsistent bitcast results, unsupported complex types, and axis normalization errors. These bugs cause crashes, compilation failures, or incorrect outputs that do not occur in eager execution mode.
• issues/105639, issues/105646, issues/118727, issues/122051, issues/122054

• GPU and Multi-GPU Related Issues: There are issues related to GPU usage including kernel launch count overflow, memory leaks during training, device query crashes when GPU drivers are missing or outdated, and performance overheads in multi-GPU training with MirroredStrategy. These problems affect stability, resource usage, and training efficiency on GPU hardware.
• issues/88288, issues/99894, issues/100085, issues/103564

• TensorFlow API and Function Bugs: Several bugs involve TensorFlow API functions producing incorrect results, raising errors due to type or axis mismatches, or lacking proper validation. Examples include SparseFillEmptyRows causing integer underflow, numpy.take_along_axis raising InvalidArgumentError, tf.experimental.numpy.isclose returning wrong results for integers, and tf.raw_ops.ResourceApplyAdagradV2 crashing on mixed data types.
• issues/63066, issues/99163, issues/108657, issues/104796

• Documentation and Link Issues: One issue reports a broken link in the conv1d_transpose documentation where the padding argument incorrectly redirects to the top of the tf.nn page instead of the specific notes section, suggesting an update to improve clarity and navigation.
• issues/91613

• Platform and Import Errors: Issues include import failures on Windows due to missing native runtime DLLs and crashes on machines with specific CPU flags (AVX-512) causing std::system_error exceptions, indicating platform-specific compatibility problems.
• issues/100335, issues/105562

• TensorFlow Model and Layer Bugs: Bugs affecting model layers include Attention layer broadcast errors on macOS M1/M4 with Python 3.12 and TFLiteConverter failures due to invalid internal representations or conversion errors in fused FullyConnected and reduction-style tf.einsum operations.
• issues/100332, issues/117453, issues/124013, issues/124161

• TFLite Delegate and Conversion Bugs: Issues in TFLite include incorrect cosine values from the XNNPACK delegate's COS kernel due to missing range reduction, division by zero returning FLT_MAX instead of infinity, and conversion failures caused by invalid output dimensions in lowered operations.
• issues/124080, issues/124150, issues/124161

• Numerical and Mathematical Function Bugs: Bugs include TensorFlow's logdet function returning NaN instead of -inf for singular matrices and TFLite's COS kernel producing out-of-range cosine values, indicating numerical inaccuracies in mathematical computations.
• issues/115768, issues/124080

• Spam and Phishing Reports: Multiple issues were identified as spam or phishing links related to Web3 crypto wallets and were closed and locked to prevent further activity, indicating ongoing moderation challenges.
• issues/124202, issues/124252, issues/124291, issues/124385, issues/124388, issues/124389, issues/124515

2.5 Issue Discussion Insights

This section will analyze the tone and sentiment of discussions within this project's open and closed issues that occurred within the past week. It aims to identify potentially heated exchanges and to maintain a constructive project environment.

Based on our analysis, there are no instances of toxic discussions in the project's open or closed issues from the past week.

III. Pull Requests

3.1 Open Pull Requests

This section provides a summary of pull requests that were opened in the repository over the past week. The top three pull requests with the highest number of commits are highlighted as 'key' pull requests. Other pull requests are grouped based on similar characteristics for easier analysis. Up to 25 pull requests are displayed in this section, while any remaining pull requests beyond this limit are omitted for brevity.

Pull Requests Opened This Week: 34

Key Open Pull Requests

1. Add support for building Tensorflow on Windows ARM64 CPUs: This pull request adds support for building TensorFlow on Windows ARM64 CPUs by modifying Bazel build files to include Windows ARM64 toolchain support and platform detection, patching rules_python to enable Windows ARM64 compatibility with the legacy WORKSPACE + pip_parse mechanism, adding platform-specific guards in TensorFlow source files to exclude unsupported functions, and addressing CPU naming issues in LLVM toolchain for Snapdragon Windows ARM64 systems.

• URL: pull/124329

• Associated Commits: 0d4b6, d5646, 2abe3, 6badb, 2b470

2. Fix thread count validation to prevent segfault (issue #105297): This pull request fixes issue #105297 by adding upper bound validation to the thread count settings in context.py to prevent segmentation faults caused by excessively large values, updates documentation to reflect these limits, and includes comprehensive unit tests to ensure proper validation and error handling.

• URL: pull/124253

• Associated Commits: f1e93, 3c514, a5ac1, 732d7

3. Fix/issue 124017 Fix GPU Pack kernel failure when autopacking 0-D scalar tensors: This pull request fixes a GPU kernel failure in TensorFlow's autopacking mechanism when handling 0-D scalar tensors by modifying the _autopacking_helper function to reshape scalars to 1-D tensors and use concat_v2 instead of the unsupported Pack operation on GPU, ensuring consistent behavior across CPU and GPU devices and adding tests to verify correct dynamic rank handling.

• URL: pull/124235

• Associated Commits: 7414e, 20064, 0c3bc

Other Open Pull Requests

• Buffer size validation in datasets: Multiple pull requests add Python-level validation to prevent excessive buffer sizes in TensorFlow datasets, raising ValueError when parameters exceed defined upper bounds. These changes prevent native crashes caused by large memory allocations in underlying C++ operations, improving robustness in FixedLengthRecordDataset and tf.data.Dataset.shuffle.
  • pull/124245, pull/124248

• Security and crash fixes: Several pull requests address critical security vulnerabilities and crash bugs by adding strict validation, overflow guards, and safer handling of inputs. These include fixes for code injection, integer overflow in TFLite operations, negative index validation, and path traversal prevention, all supported by regression and integration tests.
  • pull/124374, pull/124567, pull/124300

• Control edge and graph optimization fixes: A pull request fixes the transitive reduction of collective control edges by correcting reachability closure logic, eliminating over-serialization and iteration-order dependence. This ensures minimal and correct control edges between concurrent collective operations, validated by enhanced unit tests.
  • pull/124410

• Numerical stability and math function fixes: Fixes include resolving overflow issues in XLA implementations of tf.math.sin and tf.math.cos for complex inputs and stabilizing gradient computations for Euclidean norms at magnitude extremes. These changes replace unstable computations with stable methods and normalize components to prevent underflow/overflow, accompanied by regression tests.
  • pull/124480, pull/124518

• AutoGraph and control flow enhancements: AutoGraph is enhanced to support Python structural pattern matching (match statements) in its control flow graph builder and activity analyses. The implementation correctly handles match subjects, patterns, guards, and case bodies while preserving statement-block behavior, verified by comprehensive regression tests.
  • pull/124521

• TFLite conversion and optimization fixes: Fixes address issues in TFLite conversion including fallback handling for Keras models with multiple @tf.function methods and preventing invalid rewrites in the FuseMulAndFullyConnected optimization pass by adding shape guards. These ensure successful conversion and prevent failures during optimization.
  • pull/124545, pull/124109

• Random number generator and CPU math fixes: The ThreeFry stateless RNG is fixed by padding keys/counters and correcting counter shape checks outside XLA, with added documentation and regression tests. Additionally, the CPU implementation of tf.math.floor is specialized to correctly handle negative subnormal float32 values, matching NumPy behavior.
  • pull/124227, pull/124232, pull/124227

• Symbol and build system improvements: A pull request prevents leaking LLVM/MLIR symbols into the global process symbol table by omitting a helper in shared pip wheels, avoiding conflicts and segmentation faults on Linux. Another fixes a documentation example in OverridableFetchContent_Declare() to use the correct override variable name, preventing silent failures.
  • pull/124230, pull/124407

• Documentation and code clarity improvements: Several pull requests fix typos and improve clarity in documentation and docstrings, including adding Python code examples for report_tensor_allocations_upon_oom and related RunOptions fields. These changes enhance user understanding without altering functionality.
  • pull/124105, pull/124240, pull/124450

• Tensor handling and buffer safety fixes: Fixes include capping the number of int32 words swapped in ByteSwapBuffer for string tensors to prevent buffer overruns and improving GFile handling by initializing lazy writable handles on close to ensure empty files are created and truncation is applied correctly. These changes improve safety and correctness in file and tensor operations.
  • pull/124241, pull/124485

• Executor and performance improvements: A fix improves cost estimation in TensorFlow's executor on ARM64 Linux by applying a scale factor based on CPU frequency and virtual timer counter ratios. This enhances scheduling decisions and overall inference performance on affected platforms.
  • pull/124308

• Third-party library integration: An opt-in integration of Huawei's KDNN library is introduced with build-flag support, operator/kernel scaffolding, and Grappler remapper integration. This skeleton contribution is not yet ready to merge due to outstanding issues like license confirmation and missing benchmarks.
  • pull/124543

3.2 Closed Pull Requests

This section provides a summary of pull requests that were closed in the repository over the past week. The top three pull requests with the highest number of commits are highlighted as 'key' pull requests. Other pull requests are grouped based on similar characteristics for easier analysis. Up to 25 pull requests are displayed in this section, while any remaining pull requests beyond this limit are omitted for brevity.

Pull Requests Closed This Week: 48

Key Closed Pull Requests

1. Convert tensor/variable axes in moments and weighted_moments to preve…: This pull request addresses and fixes TypeErrors caused by improper handling of tensor and variable axes in the moments and weighted_moments functions by converting these axes appropriately, while also resolving related AttributeErrors and test issues to ensure compatibility under both Eager and Graph execution modes.

• URL: pull/121030

• Associated Commits: f3d6e, 35427, 96a8d, d2094, c312a, a87de, eab2c, ec325, e0558, ea896

2. [Bug] tf.bitcast after tf.cast to uint64 differs between eager and XLA: This pull request fixes a bug causing inconsistent behavior between eager TensorFlow and XLA-compiled execution when casting negative floating-point values to unsigned integers by updating the TensorFlow-to-XLA cast lowering to preserve TensorFlow's wrapping behavior through an intermediate signed integer conversion, and includes corresponding MLIR rewrite patterns and regression tests.

• URL: pull/118744

• Associated Commits: 09411, 76006, 6912f, 218fd, 596b5, a6a29, 1676b, 2d797

3. Fix Inconsistent Error Handling in tf.raw_ops.SparseSegmentSqrtNGradV2 and tf.raw_ops.SparseSegmentSumGradV2 Between CPU and GPU Implementations: This pull request aims to fix inconsistent error handling between CPU and GPU implementations in the TensorFlow operations tf.raw_ops.SparseSegmentSqrtNGradV2 and tf.raw_ops.SparseSegmentSumGradV2 by adding error messages on GPU for negative indices, addressing issues #94151 and #94376.

• URL: pull/108041

• Associated Commits: 49bd8, c8429, f5658, f019c, 3b495, a4b72, c0445

Other Closed Pull Requests

3.3 Pull Request Discussion Insights

This section will analyze the tone and sentiment of discussions within this project's open and closed pull requests that occurred within the past week. It aims to identify potentially heated exchanges and to maintain a constructive project environment.

Based on our analysis, there are no instances of toxic discussions in the project's open or closed pull requests from the past week.

IV. Contributors

4.1 Contributors

Active Contributors:

We consider an active contributor in this project to be any contributor who has made at least 1 commit, opened at least 1 issue, created at least 1 pull request, or made more than 2 comments in the last month.

If there are more than 10 active contributors, the list is truncated to the top 10 based on contribution metrics for better clarity.

Access Last Week's Newsletter:

• Link