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 standalone 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. [TYPE:DOCS-BUG] [TYPE:BUG] Language dropdown becomes inaccessible near bottom of viewport at tensorflow.org website.: This issue reports a problem on the tensorflow.org website where the language selector dropdown in the footer becomes inaccessible when near the bottom of the viewport because it opens downward outside the visible area. The expected behavior is for the dropdown to reposition itself, such as opening upward, to keep all language options visible and accessible.

   • The comments confirm the issue is reproducible and clarify that the dropdown at the top of the page works correctly. Contributors express interest in fixing the problem but discover that the relevant frontend code is part of Google's internal DevSite platform and not publicly available, meaning external contributions cannot directly resolve the issue; it must be escalated internally to the DevSite team.
   • Number of comments this week: 8

2. [TYPE:BUG] [COMP:XLA] [2.21.0] XLA fails to compile tf.norm with ord=2 over multiple axes due to norm/ListDiff compile-time constant error: This issue describes a bug where the XLA compiler fails to compile the TensorFlow tf.norm function with ord=2 over multiple axes due to a requirement that the input to the ListDiff operation must be a compile-time constant, which is not met when using dynamic rank information. The problem occurs specifically when using jit_compile=True and multi-axis norms, causing a compilation error during the TF-to-XLA conversion, despite the function working correctly in eager mode and non-XLA graph mode.

   • The comments reveal that a contributor has identified the root cause as the use of dynamic rank in the ListDiff operation, which prevents XLA from determining the output shape at compile time. A fix involving a static-rank fast path was proposed and implemented in a pull request, which is being tracked and reviewed by the community.
   • Number of comments this week: 3

3. [TYPE:BUG] [COMP:CORE] [AWAITING PR MERGE] [2.21.0] tf.data.Dataset.concatenate docstring shows outdated error type/message: This issue reports that the docstring for the tf.data.Dataset.concatenate method in TensorFlow version 2.21.0 shows an outdated error type and message that no longer match the actual behavior observed when concatenating datasets with incompatible element types. The user provides a reproducible example demonstrating the discrepancy and suggests that the documentation should be updated to reflect the current error message format introduced when the concatenate function was moved into a different module.

   • The comments include an offer from the user to submit a pull request to fix the docstring, a question about PyTorch version and device context which appears unrelated, and confirmation from a maintainer acknowledging the issue and noting that a pull request has already been opened to update the documentation, crediting the user for the discovery and offering to defer to the user's PR if available.
   • Number of comments this week: 3

4. [TYPE:BUG] [COMP:GPU] [2.20.0] Non-deterministic TensorFlow errors when executing tf.boolean_mask inside Grain MapTransform with num_threads> 1 on GPU: This issue describes a non-deterministic error occurring when executing the TensorFlow operation tf.boolean_mask inside a Grain MapTransform with multiple threads on a GPU, which does not happen with a single thread or when running outside Grain. The user provides reproducible code and logs showing various TensorFlow errors and notes that replacing tf.boolean_mask with tf.where plus tf.gather_nd avoids the problem, raising questions about whether the issue lies in Grain's multithreading, TensorFlow's runtime, or the specific operation.

   • The comments include an expression of interest to work on the issue, a discussion about using alternative TensorFlow operations to avoid the problem, and references to related TensorFlow issues, indicating ongoing investigation and community engagement.
   • Number of comments this week: 3

5. [TYPE:BUG] [COMP:GPU] [TF 2.19] TensorFlow was not built with CUDA kernel binaries compatible with compute capability 12.0 CUDA_ERROR_INVALID_HANDLE: This issue describes a problem where TensorFlow is not built with CUDA kernel binaries compatible with the new compute capability 12.0 GPUs, such as the NVIDIA RTX 5080, causing CUDA kernels to be just-in-time compiled from PTX and resulting in errors like CUDA_ERROR_INVALID_HANDLE during execution. Users report difficulties building TensorFlow from source for these GPUs, and while some have found workarounds using nightly builds or NVIDIA containers, native support for these newer GPUs remains incomplete, leading to ongoing requests for official support and guidance.

   • The comments discuss attempts to reproduce and resolve the issue, with some users confirming that nightly TensorFlow builds with CUDA 12.8 support the RTX 50 series GPUs, while others share workarounds involving specific Docker images or conda environments; there is also a call for official native support and updates on progress, with some users reporting partial success but continued challenges on certain setups.
   • 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: 14

Summarized Issues:

• UI Accessibility Issues: The language selector dropdown in the footer of tensorflow.org becomes inaccessible near the bottom of the viewport because it opens outside the visible area, preventing users from selecting a language. This issue affects user experience by limiting language selection functionality on the website.
• issues/123738

• Documentation Accuracy: The docstring for tf.data.Dataset.concatenate contains outdated error type and message information that no longer matches the actual TypeError message in TensorFlow 2.21.0. This discrepancy indicates a need to update the documentation to reflect the current behavior accurately.
• issues/123757

• Build and Runtime Failures: TensorFlow 2.8.0 fails to build on SUSE Linux Enterprise Server 15 SP6 with CUDA 11.2 and Bazel 4.2.1 due to an error loading the CUDA enable option resulting in a null value. Additionally, loading the native TensorFlow runtime can fail due to missing Microsoft Visual C++ Redistributable packages or unsupported CPU instruction sets like AVX/AVX2, preventing proper module import.
• issues/123789, issues/123933

• Non-Deterministic GPU Multi-threading Errors: Executing TensorFlow operations such as tf.boolean_mask, tf.transpose, and tf.concat inside a Grain MapTransform with multiple GPU threads causes non-deterministic errors including invalid argument errors and incorrect vector sizes. These issues do not occur with single-threaded execution, on CPU, or when using alternative implementations, indicating threading-related concurrency problems.
• issues/123980, issues/124017

• TFLiteConverter Bugs and Conversion Failures: The TFLiteConverter has multiple bugs including silently producing models with zero operations and outputs instead of raising errors, generating incorrect values and indices for models with multiple tf.nn.top_k calls, truncating output when slicing tensors near INT64_MAX, and producing internally inconsistent fused FullyConnected operations that fail MLIR verification. These bugs cause invalid or non-functional converted models without clear diagnostics.
• issues/123995, issues/123996, issues/123997, issues/124013

• Third-Party Library Integration Requests: There is a request to support integration of the KDNN library into TensorFlow to improve execution performance by enabling calls to KDNN functions. Initial implementation for the sigmoid operator is complete, with plans to extend support to additional operators.
• issues/124076

• TFLite XNNPACK Delegate Kernel Accuracy: The COS kernel in the TFLite XNNPACK delegate produces incorrect output values outside the expected range [-1, 1] for large float32 inputs, unlike eager TensorFlow and non-delegated TFLite implementations which correctly bound cosine results. This discrepancy affects the accuracy of cosine computations in delegated models.
• issues/124080

• Pull Request Stagnation: A pull request for the TFLite converter has been open for over five months, passed all checks, and received approval but remains unmerged. The author is requesting assistance to move the PR forward and complete the merge process.
• issues/124085

• Unclear Issue Reference: One issue references a URL linking to a GitHub Actions run in the ton-blockchain/verifier repository but lacks further descriptive details, making the problem unclear.
• issues/123979

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: 45

Summarized Issues:

• Gradient and Numerical Inconsistencies: Multiple issues report incorrect or inconsistent gradient computations and numerical results across different TensorFlow operations and devices. These include incorrect gradients for tf.dynamic_stitch, NaN gradients in tf.math.igamma, and inconsistent casting or arithmetic results between CPU and GPU, highlighting challenges in maintaining numerical correctness and device consistency.
• issues/7397, issues/123578, issues/93648, issues/106175, issues/97125, issues/94442, issues/94655, issues/98410, issues/111925

• Memory Leaks and Resource Management: Several issues describe memory leaks occurring in different TensorFlow components, including dataset saving loops, custom gradients, and GPU execution with dataset generators. These leaks cause increased memory usage and eventual crashes, indicating problems with resource cleanup and garbage collection in TensorFlow's runtime.
• issues/56177, issues/97697, issues/123269

• Device and Backend Inconsistencies: Numerous bugs highlight inconsistent behavior or crashes between CPU and GPU executions, such as differing outputs for linear algebra operations, error handling discrepancies, and crashes on CUDA for high-dimensional tensors. These issues reveal challenges in ensuring consistent and stable multi-device support.
• issues/79157, issues/79162, issues/97102, issues/97105, issues/97204, issues/93923, issues/97780, issues/98349, issues/97125, issues/93648

• Crashes and Fatal Errors in TensorFlow Operations: Multiple issues report crashes, core dumps, or fatal errors triggered by specific TensorFlow operations or configurations, including LU decomposition on multi-GPU systems, convolution backpropagation on Windows, and invalid parameter settings in session configurations. These failures often lack graceful error handling and cause abrupt termination of processes.
• issues/104852, issues/104881, issues/105331, issues/112404, issues/113144, issues/95762

• Build and Environment Compatibility Issues: Some issues describe build failures or runtime incompatibilities related to environment variables, Linux kernel versions, protobuf versions, or CUDA toolchain versions. These problems prevent TensorFlow from building or running correctly in certain setups, complicating deployment and usage.
• issues/68966, issues/73978, issues/77948, issues/94030

• API and Functionality Inconsistencies: Several issues report inconsistent or unexpected behavior in TensorFlow APIs, such as tf.reshape behaving differently in eager vs. graph mode, tf.math.bincount error handling differing between eager and compiled modes, and tf.nn.local_response_normalization producing different outputs or errors on CPU vs. GPU. These inconsistencies affect usability and reliability of the APIs.
• issues/51241, issues/117960, issues/97105

• TensorFlow Lite and XLA Bugs: Issues include integer overflow and memory safety vulnerabilities in TensorFlow Lite, incorrect model conversion dropping broadcast dimensions, and severe numerical inconsistencies with XLA compilation producing non-deterministic outputs. These problems impact model deployment and optimization workflows.
• issues/121166, issues/121631, issues/124048, issues/122952

• Performance and API Usability Concerns: One issue highlights poor performance and high memory usage in the tf.data.Dataset API for windowed and timeseries data, calling for improved windowing and shuffling methods to enhance training efficiency and user experience.
• issues/44675

• Binary Size Regression: A significant increase in the Windows TensorFlow binary size between versions 2.16.2 and 2.17.0/2.18.0 is reported, with no clear cause identified, raising concerns about build optimization and distribution size.
• issues/84962

• Session and Configuration Parameter Bugs: Issues report crashes or aborts when setting negative values for parallelism thread parameters in ConfigProto or tf.compat.v1.ConfigProto, indicating improper validation and error handling in session configuration.
• issues/95762, issues/112404

• Graph Optimization Bugs: Bugs in TensorFlow's Grappler optimizer cause incorrect transformations and floating-point rounding errors, leading to wrong results or discrepancies between eager and compiled execution modes.
• issues/118675, issues/119429

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: 37

Key Open Pull Requests

1. Fix crash in FromProtoField on DecodeUnaryVariant failure: This pull request fixes a crash caused by partially-decoded Variant objects in FromProtoField<Variant> when DecodeUnaryVariant fails, by ensuring these objects are properly cleared before buffer unreferencing to prevent destructor-induced Python thread state corruption and adds comprehensive regression tests to verify that malformed DT_VARIANT TensorProtos raise clean exceptions instead of causing segmentation faults.

• URL: pull/123829

• Associated Commits: a883a, 9758f, c5faa, 976c4, 18714

2. Fix RFFT crash on zero fft_length by changing validation from >= 0 to > 0: This pull request fixes a crash in the TensorFlow RFFT operation caused by allowing zero fft_length values by changing the validation checks in the FFTBase and FFTNBase Compute functions from >= 0 to > 0, ensuring that zero lengths raise a clean InvalidArgumentError instead of triggering backend assertion failures, and includes updated error messages and new tests to verify this behavior.

• URL: pull/123902

• Associated Commits: ed14b, ace08, 0a776, 6a21f

3. Return empty tensors instead of uninitialized scalars for unused linalg outputs: This pull request fixes an issue where certain linear algebra operations in TensorFlow, such as SelfAdjointEigV2, Eig, and CPU Svd, returned uninitialized scalar outputs for unused eigenvector or singular-vector results by changing the allocation to empty tensors of shape [0], thereby aligning runtime output shapes with shape functions, preventing garbage data returns, updating GPU kernel behavior, clarifying API documentation, and adding regression tests to ensure correctness.

• URL: pull/123861

• Associated Commits: 82dfc, af8ba, bbd60

Other Open Pull Requests

• Fixes for input validation and error handling in TensorFlow ops: Multiple pull requests improve input validation by adding rank and dimension checks to prevent crashes and aborts in various TensorFlow operations such as LSTMBlockCell, crop-and-resize gradient kernels, WriteScalarSummary, EncodePng, Unbatch, and ResourceApply ops. These changes ensure that invalid inputs produce catchable errors instead of fatal crashes, improving robustness and adding regression tests to verify correct error handling.
  [pull/124034, pull/123864, pull/123672, pull/123767, pull/123865, pull/124045, pull/124019, pull/123672, pull/123865]

• Fixes and improvements to numerical operations and math functions: Several pull requests address correctness and edge cases in TensorFlow math functions, including fixes to tf.math.floordiv for negative infinite divisors, tf.experimental.numpy.isclose for integer inputs, tf.math.round for complex types, and tf.math.igamma gradient NaN issues. These fixes ensure consistent behavior with NumPy, add proper error messages, and correct numerical gradients with added test coverage.
  [pull/123862, pull/123867, pull/124019, pull/123803]

• Security and resource management improvements: Pull requests fix security vulnerabilities by capping decompression sizes in DecodeCompressed and address unpinned-uses linter warnings in GitHub Actions workflows. Additionally, a memory leak in TFLiteSavedModelConverterV2.convert() is fixed by changing how debug info is loaded, improving efficiency and preventing leaks.
  [pull/123846, pull/123838, pull/123818]

• Fixes to TensorFlow Lite and Keras fallback and metrics: Fixes include preserving and rethrowing Keras tracing errors during TFLite fallback conversion to prevent empty models, and correcting the AUC.result() method in Keras to return NaN when true labels contain only a single class, addressing incorrect ROC-AUC calculations.
  [pull/124016, pull/123668]

• Updates to documentation and examples: Documentation fixes include adding missing imports, correcting broken links, updating docstring examples for tf.data.Dataset.concatenate, and clarifying supported data types and error messages in math functions. These changes improve clarity and accuracy without affecting functionality.
  [pull/124018, pull/123855, pull/124019]

• Graph shape inference and sparse matrix validation improvements: Updates to shape inference logic allow zero-sized outputs for Conv2D and pooling ops, aligning static inference with eager and XLA execution. Validation is added for CSR sparse matrix components to prevent out-of-bounds memory access from malformed data.
  [pull/123894, pull/123845]

• Algebraic simplifier and optimization gating: The AlgebraicSimplifier rewrite rule converting exp(LHS)*exp(RHS) to exp(LHS+RHS) is gated behind an enable_fast_math option to avoid incorrect numerical results at overflow/underflow boundaries, with tests added to verify correct behavior.
  [pull/123991]

• Build and tooling updates: A pull request updates the rules_ml_toolchain reference, fixing invalid calls and missing repository initializations, and another adds new HTML and CSS files to the project.
  [pull/123893, pull/123758]

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: 50

Key Closed Pull Requests

1. Reject rank-0 shape tensors in tf.reshape for consistency: This pull request introduces Python-level validation in TensorFlow's tf.reshape function to consistently reject rank-0 (scalar) shape tensors across both eager execution and graph modes, providing a clear error message that suggests using a one-element shape list instead, thereby resolving inconsistencies between execution modes.

• URL: pull/109356

• Associated Commits: eb68d, 6390e, 77f03, fe7c5

• Associated Commits: eb68d, 6390e, 77f03, fe7c5

2. Add size limit to scalar tensor handle cache to prevent memory leak: This pull request introduces a maximum size limit of 1024 entries to the scalar tensor handle cache in TensorFlow's eager execution mode to prevent unbounded memory growth caused by caching numerous distinct scalar values, ensuring that when the cache reaches capacity, all existing entries are properly released before new ones are inserted, thereby fixing a memory leak while maintaining caching benefits for frequently used scalar values.

• URL: pull/109357

• Associated Commits: 167db, eeabd, ff179, d2e15

• Associated Commits: 167db, eeabd, ff179, d2e15

3. Fix MaxPool3D MKL/oneDNN CHECK failure on non-5D input: This pull request fixes a fatal CHECK failure in the MKL/oneDNN MaxPool3D kernel by adding input rank validation that returns an InvalidArgument error instead of aborting when the input tensor does not have the expected 5D shape, thereby improving robustness for incorrect input ranks.

• URL: pull/112652

• Associated Commits: b22e6, ee64d, 6a1d5, fc96d

• Associated Commits: b22e6, ee64d, 6a1d5, fc96d

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