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 a breaking change in the LiteRT (tf.lite) API, including the deprecation of tf.lite.Interpreter in favor of ai_edge_litert.interpreter and changes to certain C++ constants for better API compatibility. Additionally, the tfl.Cast operation now supports bfloat16 in the runtime kernel, and the libtensorflow packages are no longer published separately but remain accessible via the PyPI package.

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:AWAITING TENSORFLOWER] [TYPE:FEATURE] [TYPE:BUILD/INSTALL] [SUBTYPE:WINDOWS] [TF 2.18] Tensorflow not supported on Windows + ARM CPUs: This issue reports that TensorFlow cannot be imported on Windows 11 running on ARM CPUs, despite successful installation, due to missing native runtime support and DLL load failures. The discussion clarifies that TensorFlow currently does not provide official Windows wheels for ARM architectures, making the CPU unsupported out of the box, and suggests possible workarounds like using WSL or compiling from source, with some users eventually reporting success compiling TensorFlow for Windows ARM64 after code modifications.

   • The comments reveal initial confusion about the issue being a duplicate of an older Intel CPU compatibility problem, but it was re-opened as a distinct ARM-related issue; users shared CPU specs and discussed the lack of official ARM Windows wheels, explored alternatives like WSL and Colab, and in later comments, some contributors mentioned attempts and eventual success in compiling TensorFlow for Windows ARM64, indicating ongoing development toward official support.
   • Number of comments this week: 2

2. [STAT:CONTRIBUTION WELCOME] [TYPE:BUG] [TF 2.19] TensorShape CHECK failure in nightly & stable: This issue reports a bug in TensorFlow versions 2.20.0 and nightly where running the tf.raw_ops.Unbatch() operation with simple 1-D inputs causes a native crash due to a C++ CHECK failure related to tensor shape handling, resulting in the Python interpreter being terminated without a Python-level exception. The problem appears to stem from improper validation of tensor dimensions within the UnbatchResource::Compute function, specifically accessing a dimension size without confirming the tensor's rank, which leads to an abort signal during execution.

   • The comments confirm the reproducibility of the issue on the specified TensorFlow versions and provide a reference notebook for replication; maintainers acknowledge the bug but have not prioritized a fix, encouraging community contributions instead. One contributor shared a preliminary analysis identifying the root cause related to dimension checks and subsequently withdrew from pursuing a fix, leaving the issue open for others to address.
   • Number of comments this week: 2

3. [STAT:AWAITING RESPONSE] [TYPE:BUG] [COMP:OPS] [2.17] Aborted (core dumped) in tf.raw_ops.MatrixSolve: This issue reports a crash occurring in the TensorFlow operation tf.raw_ops.MatrixSolve when the matrix argument has an empty shape and a GPU is available, causing an abort with a core dump. The problem is reproducible on TensorFlow nightly builds and involves a failure triggered by invalid tensor shapes during GPU execution.

   • The comments discuss testing the operation with valid tensor inputs instead of empty datasets, confirming that the crash does not occur in those cases, and suggest that the API should handle such edge cases more gracefully. The TensorFlow team acknowledges the issue but marks it for community contributions, noting that recent versions raise a ValueError instead of crashing and recommending users update to the latest TensorFlow release or open a new issue if problems persist.
   • Number of comments this week: 1

4. [TYPE:DOCS-BUG] [TYPE:BUG] [AWAITING PR MERGE] [TF 2.16] Notes on padding missing in documentation: This issue addresses a documentation bug where the link for the padding argument in the conv1d_transpose function incorrectly redirects to the top of the TensorFlow neural network documentation page instead of the specific section explaining padding. The reporter suggests that the link should point to the relevant "Notes on padding" section in the source code to provide proper context and clarity.

   • The comments reveal interest from multiple contributors to fix the broken anchor links affecting many functions in the same module, with one user planning a comprehensive PR. However, a maintainer clarifies that the fix belongs in the tensorflow/docs repository rather than tensorflow/tensorflow, advising the issue be refiled there for proper handling. Another commenter requests assignment to work on updating the docstrings accordingly.
   • Number of comments this week: 1

5. [STAT:CONTRIBUTION WELCOME] [TYPE:BUG] [COMP:GPU] [TF 2.19] [GPU] CUDA_ERROR_ILLEGAL_ADDRESS / “No algorithm worked” in Conv2D after full spatial self-attention at 224×224 (N≈50k → N²≈2.5B). Looks like index/size limit; should fail earlier with a clearer error: This issue describes a bug where running a Keras model with full spatial self-attention on 224×224 feature maps followed by a Conv2D layer causes a GPU crash with a CUDA illegal memory access error and a failure to find a suitable cuDNN convolution algorithm. The problem appears related to index or size limits in the attention mechanism, as the error surfaces at the Conv2D call after the attention operation, and users report the issue occurs on multiple GPU setups with TensorFlow 2.20.0 and later versions but not on 2.19.0.

   • Commenters confirmed reproducibility on various GPUs and TensorFlow versions, shared a workaround involving specific NVIDIA driver and CUDA toolkit versions, noted the issue does not occur on TensorFlow 2.19.0, and maintainers acknowledged the bug but marked it for community contributions without immediate prioritization.
   • Number of comments this week: 1

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: 13

Summarized Issues:

• Memory leaks in TensorFlow components: Multiple issues report memory leaks occurring in different parts of TensorFlow, including host memory leaks during GPU execution with tf.data.Dataset.from_generator and TensorFlow Lite FFT kernels where buffers are not released properly. These leaks cause resource exhaustion and can degrade system performance over time.
• issues/123269, issues/123387

• Crashes due to improper error handling in ops: Several bugs cause TensorFlow to crash with fatal signals instead of raising recoverable exceptions when given invalid inputs, such as empty tensors in CropAndResizeGradImage, zero FFT length in RFFT, and zero-dimension features in SequenceExample parsing. This lack of proper validation leads to denial-of-service conditions and process aborts.
• issues/123397, issues/123399, issues/123476

• Memory safety and crash bugs in low-level ops: There are memory-safety bugs and crashes in low-level TensorFlow operations, including a bad-free error in the CPU implementation of tf.raw_ops.CTCLoss and integer conversion errors causing infinite loops and out-of-memory crashes in Dataset cache functions. These issues cause process aborts and hard crashes without graceful error handling.
• issues/123396, issues/123484, issues/123485

• Numerical inaccuracies and inconsistent results in math operations: Some TensorFlow math operations produce incorrect or inconsistent results, such as tf.reduce_mean returning inaccurate values for large float16 tensors and tf.math.rsqrt yielding different outputs depending on input length due to CPU bfloat16 implementation differences. These inaccuracies affect the reliability of numerical computations.
• issues/123299, issues/123551

• Incorrect algebraic simplification causing NaN to 1.0 conversion: A bug in XLA's algebraic simplification rewrites exp(a) * exp(b) to exp(a + b) incorrectly, changing NaN results to 1.0 in cases involving floating-point overflow and underflow. This causes discrepancies between XLA-compiled and eager or graph execution in TensorFlow.
• issues/123169

• Import errors due to DLL initialization failure: TensorFlow 2.19 encounters an ImportError when importing _pywrap_tensorflow_internal because the DLL initialization routine fails, preventing the native runtime from loading. This blocks TensorFlow from starting properly.
• issues/123395

• NaN gradients in special function computations: The gradient of tf.math.igamma with respect to x returns NaN at specific inputs (a=1, x=0) instead of the mathematically expected value 1.0, due to an indeterminate form in the gradient computation. This affects gradient-based optimization involving this function.
• issues/123578

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: 50

Summarized Issues:

• Model Saving and Loading Issues: Several issues report problems with saving and loading models or optimizers in TensorFlow, including incorrect preservation of optimizer states in SavedModel format and backward compatibility failures when loading Keras 2 models in TensorFlow 2.17 using Keras 3. These issues cause discrepancies in training continuation, prediction results, and errors related to unsupported file formats or missing attributes.
• issues/44670, issues/77356

• Build and Compilation Failures: Multiple reports describe build failures across different platforms and configurations, including unresolved repository errors, invalid static_cast errors in Eigen for AVX512 FP16, linking errors on Fedora 40, and Bazel build issues on Windows and AWS Graviton3. These failures prevent successful compilation and wheel generation, often requiring dependency updates or specific build flags.
• issues/69367, issues/69674, issues/72081, issues/72166, issues/77072, issues/90448

• Memory Usage and Performance Problems: Several issues highlight excessive memory consumption, memory leaks, or performance regressions, such as increased memory usage in model.predict in newer TensorFlow versions, CPU memory leaks with tf.image.crop_and_resize, slow data reading with dynamic sharding, and performance degradation in oneDNN convolution operations. These problems impact training stability and inference efficiency.
• issues/56904, issues/58676, issues/70541, issues/90784, issues/72014

• TensorFlow API and Operation Bugs: Various bugs affect TensorFlow operations and APIs, including segmentation faults in RaggedTensorToTensor and sparse_bincount, incorrect results from tf.math.reciprocal on complex128 inputs, errors in tf.nn.weighted_moments with keepdims=False, and incorrect gradients in tf.math.xlogy and tf.math.xlog1py. These bugs cause crashes, incorrect computations, or unexpected behavior during model execution.
• issues/85240, issues/74918, issues/78298, issues/101580, issues/119476

• TensorFlow Profiler and Debugging Limitations: Users report that the TensorFlow profiler fails to show profiling data in TensorBoard, and that AutoGraph application to custom loss functions during eager execution hinders debugging despite run_eagerly=True. These issues reduce the ability to effectively profile and debug TensorFlow models.
• issues/61212, issues/41189

• TensorFlow Build Configuration and Dependency Issues: Problems include cyclic dependencies with tensorflow-io-gcs-filesystem, ignored compute capability and TensorRT settings when using --config=cuda_clang, and missing Windows builds in nightly libtensorflow packages. These issues complicate building and configuring TensorFlow for specific environments or hardware.
• issues/56636, issues/74370, issues/46538

• GPU and CUDA Integration Problems: Issues include persistent warnings about missing TensorRT despite correct installation, inability to extend GPU operator compute functions, and build failures due to missing CUDA build flags. These problems affect GPU acceleration and custom GPU operation development.
• issues/68335, issues/72422, issues/86405

• Data Pipeline and Dataset API Limitations: Requests and bugs include the need for a public API to clear CacheDataset caches, inability to inspect the actual runtime value of tf.data.AUTOTUNE, and slowdowns caused by dynamic sharding policies. These issues limit control and performance tuning of data input pipelines.
• issues/54157, issues/72369, issues/70541

• Security Vulnerabilities: A critical security issue is reported where custom layers in TensorFlow/Keras can execute arbitrary malicious code during model loading or execution, posing a risk of remote code execution and system compromise.
• issues/82214

• TensorFlow Version and Platform Support Questions: Users seek clarification on dropped GPU support on native Windows in TensorFlow versions above 2.10, plans to upgrade manylinux support from manylinux 2014 to newer versions, and challenges building TensorFlow on Power architecture without boringssl. These questions reflect concerns about platform compatibility and future support.
• issues/69750, issues/70801, issues/70352

• TensorFlow Model API Behavior Changes: A regression is reported where tf.keras.Sequential models assigned as attributes in custom tf.Modules no longer behave as tf.Modules themselves, causing variables to be excluded from the parent module's variables list. This change breaks previous assumptions about model composition.
• issues/74297

• TensorFlow Distributed and DTensor Usage Challenges: Users discuss difficulties optimizing large matrix computations with distributed strategies and DTensor, seeking more efficient sharding and multiplication methods beyond current approaches involving jacobian gathering and mesh copying.
• issues/71930

• TensorFlow Image Processing Bugs: Issues include crashes when encoding illegal images with tf.io.encode_png, and incorrect results from tf.image.adjust_gamma on integer inputs due to normalization and type conversion. These bugs affect image preprocessing and augmentation pipelines.
• issues/76726, issues/123290

• TensorFlow Memory Optimization Proposals: A feature request suggests replacing the static list of "cheap" operations for recomputation in TensorFlow's memory optimizer with a dynamic cost model that adapts to operation characteristics and runtime factors, aiming to improve memory efficiency and performance.
• issues/94653

• TensorFlow Conversion and Interoperability Performance Issues: A performance bottleneck is reported when converting TensorFlow tensors to CuPy arrays using tf.experimental.dlpack.to_dlpack(), where the conversion is significantly slower than the reverse, causing delays in inference workflows.
• issues/88731

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: 23

Key Open Pull Requests

1. Validate vocab_size against dataset cardinality in table_from_dataset: This pull request adds validation to ensure that the vocab_size parameter in table_from_dataset and its wrapper does not exceed the dataset's known static cardinality or a maximum allowable value, raising a Python exception before reaching the C++ layer to prevent process aborts caused by oversized vocabulary sizes.

• URL: pull/123366

• Associated Commits: c325c, 47520, 5e638, 08b89, a7406, b02a2

2. Fix tf.sigmoid on bfloat16 CPU non-monotonicity and rounding issues: This pull request fixes the non-monotonicity and incorrect rounding issues of the tf.sigmoid function on bfloat16 CPU inputs by specializing the Eigen scalar_logistic_op to perform computations in float32 precision before casting back to bfloat16, and disabling SIMD packet vectorization to ensure correct, monotonic, and accurately rounded sigmoid outputs on CPU.

• URL: pull/123323

• Associated Commits: ee080, 2248c, 1cdcf

3. Fix zero-day division by zero NaN in Keras categorical_crossentropy: This pull request fixes a critical issue in Keras's categorical_crossentropy function where division by zero could produce NaN values by ensuring proper handling of zero outputs and preventing the loss gradients from being corrupted.

• URL: pull/123494

• Associated Commits: 58769, f35ec, 51112

Other Open Pull Requests

• oneDNN integration updates: This topic includes updates to the TensorFlow/XLA oneDNN integration for AArch64 by upgrading the oneDNN source archive from version 3.7.0 to 3.11.3, updating the Bazel build, adding compatibility patches, and revising threadpool integration to match the newer oneDNN interface. These changes ensure improved compatibility and performance on AArch64 platforms.
  • pull/123217

• Reciprocal operation and involution fixes: Multiple pull requests address the handling of the Reciprocal operation in TensorFlow's Grappler and MLIR dialect by removing incorrect involution traits and excluding reciprocal from involution simplifications. These fixes prevent incorrect folding optimizations and ensure that double or triple reciprocals produce correctly rounded results consistent with eager execution.
  • pull/123195, pull/123282

• Security and memory safety fixes: Several pull requests fix vulnerabilities and memory issues by rejecting invalid inputs such as empty row_splits in RaggedTensorVariant decoding and zero-element non-memcpyable components in UncompressElement. Additionally, a host memory leak in tf.data.Dataset.from_generator is fixed by properly managing closure references, preventing out-of-bounds reads and memory leaks.
  • pull/123446, pull/123530, pull/123572

• Bug fixes in tensor operations and kernels: This group of pull requests fixes various bugs including a critical issue in the oneDNN quantized depthwise-convolution kernel by validating scalar inputs, a buffer overrun in ByteSwapBuffer for complex types, a regression in tf.raw_ops.Round for integer tensors, and a crash in ScatterNd ops due to rank mismatches. These fixes improve stability and correctness of tensor operations.
  • pull/123149, pull/123293, pull/123319, pull/123379

• CoreML delegate and build fixes: Fixes include moving the GraphBuilder constructor and destructor out-of-line to resolve compilation errors in CoreML delegate builders and correcting a typo in convolution_op_builder.cc. These changes enable successful linking of the CoreMLDelegate on Chrome for iOS.
  • pull/123414

• CTC loss and iterator cache fixes: Fixes address a heap out-of-bounds write in the native CTC loss CPU implementation by guarding matrix accesses and resolve issues in IteratorRandomAccessCache and DatasetRandomAccessCache by fixing unguarded vector resize and signed-to-unsigned conversion bugs. These ensure memory safety and correctness in dataset caching and loss computations.
  • pull/123451, pull/123489, pull/123491

• RFFT validation and error handling improvements: A fix changes the validation in tf.raw_ops.RFFT to reject zero FFT lengths cleanly by raising an InvalidArgumentError instead of crashing, with a regression test added to verify this behavior. This improves robustness of the RFFT operation.
  • pull/123516

• Numerical precision and gradient fixes: Fixes include addressing a length-dependent double rounding issue in tf.math.rsqrt for bfloat16 by specializing scalar computation, and correcting the gradient of tf.math.igamma at specific inputs to avoid NaN values by replacing a problematic term. These changes improve numerical accuracy and gradient correctness.
  • pull/123575, pull/123580

• Argmax support in MaxPool gradient: This pull request adds support for int32 argmax in CPU MaxPool gradient computations by templating the gradient implementation and extending tests to validate both int32 and int64 argmax tensors, fixing an automatic differentiation failure with int32 indices.
  • pull/123156

• Documentation and migration guidance: This pull request provides detailed guidance and documentation updates to assist users and agents migrating from TensorFlow Lite to LiteRT, including README.md updates and a new SKILL.md file.
  • pull/123632

• Miscellaneous updates: This pull request updates some files in the TensorFlow project without further specification.
  • pull/123391

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: 25

Key Closed Pull Requests

1. fix(math): correct xlogy/xlog1py gradient w.r.t. x at x=0: This pull request fixes the incorrect gradient computation for the TensorFlow functions tf.math.xlogy and tf.math.xlog1py with respect to x at x = 0 by replacing the erroneous zero-mask logic in the backward pass so that the gradient correctly returns log(y) or log1p(y) instead of zero, thereby resolving dead zones in optimization affecting models using KL divergence, cross-entropy, and mixture training.

• URL: pull/119869

• Associated Commits: 03a41, 82f81, 27183, c6c29, 7f3a7, 9f573, 88c28, a6ab8, 88bf1, 4e662, d1cee, 54d6b, cbc82, 1fd39, 94cca, 7532e, 44e74, ec497, bce11, 5ed9c, 61395, a9115

• Associated Commits: 03a41, 82f81, 27183, c6c29, 7f3a7, 9f573, 88c28, a6ab8, 88bf1, 4e662, d1cee, 54d6b, cbc82, 1fd39, 94cca, 7532e, 44e74, ec497, bce11, 5ed9c, 61395, a9115

2. Validate tf.name_scope names with spaces in eager and graph modes: This pull request improves TensorFlow by adding validation for tf.name_scope names containing spaces in both eager and graph modes, ensuring invalid names raise a ValueError immediately rather than causing confusing errors later, and includes tests to cover these validation cases.

• URL: pull/113161

• Associated Commits: d14b3, 62ba5, 738ac, 55c69, 7d988, cd45b, abdba, 9e4b9

• Associated Commits: d14b3, 62ba5, 738ac, 55c69, 7d988, cd45b, abdba, 9e4b9

3. Fix tf.signal.idct docstring: n parameter is supported: This pull request corrects the docstring of the tf.signal.idct function to accurately reflect that the n parameter is supported and functional—removing incorrect statements that it must be None and that its use raises a ValueError—while also updating related documentation and tests to validate proper behavior of n in truncation and padding scenarios.

• URL: pull/118322

• Associated Commits: f9a6f, 379c4, 757bd, 7e23b, 5a181, 05af2, be45f, 5b735

• Associated Commits: f9a6f, 379c4, 757bd, 7e23b, 5a181, 05af2, be45f, 5b735

Other Closed Pull Requests

• Precision fix in tf.math.sign for complex64 inputs: This pull request fixes an issue where very small nonzero complex64 inputs to tf.math.sign incorrectly returned zero due to underflow in magnitude computation. It promotes intermediate calculations to float64 and complex128 precision to avoid subnormal values and kernel signature mismatches, ensuring accurate sign results.
• pull/117023

• Robust error handling in parsing and feature extraction: Multiple pull requests address critical bugs and security issues in parsing functions such as ParseBytesFeature and TFLite's ParseExample. They add checks for unchecked return values, prevent heap out-of-bounds writes, handle empty serialized batches safely, and fix integer overflow vulnerabilities to improve robustness and security.
• pull/120699, pull/121168

• Default oneDNN asynchronous execution engine: This pull request makes the oneDNN asynchronous execution engine the default for x86_64 Linux and Windows builds in TensorFlow, improving performance on these platforms.
• pull/120403

• Dynamic cost model for Grappler memory optimizer: The static hardcoded list of "cheap to recompute" operations in the Grappler memory optimizer is replaced with a dynamic cost model using the OpLevelCostEstimator. This allows more accurate identification of inexpensive operations based on tensor shape information, with fallback to the static list when shape inference is unreliable.
• pull/109358

• Enhanced debug traceability in TensorFlow Lite operations: New location logic in flatbuffer_import.cc enables TensorFlow Lite operations to carry subgraph and operation identifiers along with fused-activation suffixes. This improves debug traceability and is supported by added and updated MLIR test files validating the new location format.
• pull/115950

• Improved exception specificity in speech_commands example: Generic Exception usage is replaced with more specific built-in exceptions like ValueError, FileNotFoundError, and ImportError in the speech_commands example code. Corresponding docstrings are updated to reflect these changes, enhancing error handling precision.
• pull/117858

• Improved error handling and deprecation in to_numpy() for remote DTensor meshes: The silent return of np.array([None]) is replaced with a clear NotImplementedError in to_numpy() for remote DTensor meshes. Unit tests are added and a transitional environment variable is introduced to softly deprecate the old behavior, ensuring safer usage.
• pull/119410

• Fix for XLA compiler failure on tf.squeeze with bounded dynamic dimensions: The SqueezeOp::Compile method is updated to correctly preserve dynamic dimensions, fixing a compilation failure in the XLA compiler for tf.squeeze when input tensors have bounded dynamic dimensions. A new unit test verifies this fix.
• pull/123015

• Integer overflow fix in Conv2D GEMM dimension calculations: This pull request fixes a critical integer overflow by promoting intermediate multiplication operands from 32-bit to 64-bit integers in Conv2D GEMM dimension calculations. This prevents process crashes caused by SIGABRT when input dimension products exceed 32-bit integer limits.
• pull/112776

• Fix for cumulative_logsumexp NaN issue with multiple +inf inputs: The tf.math.cumulative_logsumexp function is fixed to handle inputs with multiple +inf values by short-circuiting to return +inf directly. This prevents invalid inf - inf operations and ensures correct cumulative results.
• pull/115123

• Validation added to QuantizedMaxPool MKL kernel ksize parameter: Validation ensures the ksize parameter does not exceed input tensor dimensions in the QuantizedMaxPool MKL kernel. This provides a clear InvalidArgumentError instead of an unclear low-level oneDNN backend error, with test coverage and updates based on review feedback.
• pull/122516

• Validation added to Huber loss delta parameter: The Huber loss function now validates that the delta parameter is greater than zero, raising a ValueError at construction time to prevent nonsensical results like division by zero. This validation is guarded to accommodate symbolic TensorFlow types during graph construction.
• pull/122611

• Pinning GitHub Actions to immutable commit SHAs for security: Several pull requests enhance CI workflow security by pinning unpinned GitHub Actions references to immutable full-length commit SHAs. This prevents supply-chain attacks from mutable tags while retaining version tags as inline comments for reference.
• pull/122209, pull/122210, pull/122211

• Fix TypeError in tf.nn.weighted_moments with axes as tf.Tensor: This pull request fixes a TypeError in tf.nn.weighted_moments when the axes parameter is passed as a tf.Tensor or numpy array with keepdims=False. It normalizes axes to a Python list before calling squeeze and adds regression tests to verify the fix.
• pull/122402

• Runtime ResourceExhausted check in SnappyInputBuffer::Inflate: The debug-only buffer size check is replaced with a runtime ResourceExhausted check to prevent heap overflow vulnerabilities when decompressing untrusted snappy streams. A corresponding test ensures safe handling of small output buffers.
• pull/122667

• Validation in Keras Conv*DTranspose layers for unsupported configs: Validation is added to prevent simultaneous use of strides > 1 and dilation_rate > 1 in Conv*DTranspose layers, which causes crashes in oneDNN/MKL kernels. A clear ValueError is raised to improve user feedback and prevent runtime errors.
• pull/121423

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