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 and the discontinuation of publishing libtensorflow packages, 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:BUG] [COMP:CORE] [TF 2.19] Keras model.fit() crashes with ValueError: as_list() is not defined on an unknown TensorShape when dataset elements come from tf.numpy_function without shapes: This issue describes a bug where using tf.numpy_function within a tf.data.Dataset without explicitly setting the shapes of the returned tensors causes tf.keras.Model.fit() to crash with a ValueError related to calling as_list() on an unknown TensorShape. The expected behavior is that Keras should instead raise a clear validation error indicating that dataset element shapes are unknown and must be specified, rather than crashing internally.

   • The comments confirm the issue is reproducible and explain that tf.numpy_function outputs tensors with unknown rank unless shapes are manually set using set_shape(), which resolves the error. Users shared environment details and a workaround, and requested improved user-facing error messages to prevent confusion from the internal TensorShape.as_list() crash.
   • Number of comments this week: 2

2. [TYPE:BUG] [COMP:LITE] [TF 2.13] AGP 9 build fails when two libs share the same namespace "org.tensorflow.lite.": This issue describes a build failure occurring when migrating an Android project to AGP 9.0 due to duplicate namespaces in multiple libraries sharing the same namespace "org.tensorflow.lite.support." The problem arises during manifest processing because AGP 9 now fails if two libraries declare the same namespace, specifically affecting TensorFlow Lite support libraries.

   • The comment suggests a workaround by defining unique namespaces for each library, including transitive dependencies, to resolve the namespace conflict causing the build failure.
   • Number of comments this week: 1

3. [TYPE:BUG] [TF 2.19] [CUDA] illegal memory write on Unique: This issue reports an illegal memory write detected by Compute-Sanitizer in the CUDA kernel RangeInitKernel when using the Unique operation on a very large tensor size near the INT32 limit. The problem arises because the GPU implementation relies on 32-bit indexing, which overflows at this size, causing out-of-bounds memory access.

   • The comment explains that the bug is due to index arithmetic overflow when the input tensor size reaches INT32_MAX, and questions whether the Unique operation should support such large tensors on GPU or instead fail fast with a validation error; the commenter also offers to help implement a guard or test for this case.
   • Number of comments this week: 1

4. [TYPE:BUG] [TF 2.19] [CUDA] illegal memory read on TopK: This issue reports an illegal memory read error occurring in the CUDA TopK kernel when processing an extremely large input tensor close to the maximum 32-bit integer size, which likely causes an index overflow. The user provides a minimal reproducible example and detailed environment information, highlighting a potential limitation or bug in handling large tensors on the GPU.

   • The comment acknowledges the minimal repro and suggests that the problem arises from the TopK kernel's reliance on 32-bit indexing, which cannot handle inputs at the INT32_MAX size, proposing either support for such large tensors or a fail-fast validation to prevent the error.
   • Number of comments this week: 1

Since there were fewer than 5 open issues, all of the open issues have been listed above.

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: 16

Summarized Issues:

• CUDA Kernel Memory Errors: Multiple CUDA kernels in TensorFlow 2.20 exhibit illegal out-of-bounds memory reads or writes during various GPU operations, as detected by Compute-Sanitizer on Ubuntu 22.04 with CUDA 12.5.1 and cuDNN 9.2.1. These errors affect kernels such as BiasGradNHWC_Naive, CheckNumericsKernel, GatherSliceOpKernel, LegacyResizeBilinearKernel, RangeInitKernel, RollKernel, SwapDimension1And2InTensor3UsingTiles, and TopK, causing crashes or invalid memory access.
• issues/109515, issues/109516, issues/109517, issues/109518, issues/109519, issues/109520, issues/109522, issues/109523

• Build and Compatibility Issues: There are concerns about compatibility when upgrading TensorFlow to version 2.20 and building it with LLVM while other components like Protobuf, gRPC, brpc, and self-developed code are built with GCC. Additionally, importing TensorFlow on an aarch64 server with Ubuntu 24.04.3 LTS and Python 3.10 results in an ImportError related to pybind11_protobuf, despite multiple build attempts.
• issues/109505, issues/109869

• Android Build and Namespace Conflicts: Building Android projects with AGP 9 causes failures due to multiple libraries sharing the same namespace "org.tensorflow.lite.support," leading to duplicate namespace conflicts during manifest processing. There is also a request for cross compilation support to build TensorFlow wheel (.whl) files for Android, as current tools and formats are limited to .aar files.
• issues/109508, issues/109605

• Model and API Bugs: TensorFlow versions above 2.15.0 have a bug in Bidirectional LSTM layers with asymmetric units causing shape mismatch errors during model fitting. Another bug causes the pretrained MobileNetV3Large model to incorrectly load the MobileNetV3Small structure and weights, resulting in parameter mismatches.
• issues/109637, issues/109949

• XLA GPU JIT Compilation Failure: Using tf.map_fn on a zero-length tensor in TensorFlow 2.21.0-dev triggers an XLA GPU JIT compilation failure with an InvalidArgumentError due to slicing issues during conversion to XLA operations.
• issues/109648

• Segmentation Fault with Python and OpenGL: TensorFlow 2.20.0 combined with Python 3.13 and PyOpenGL 3.1.10 on Debian GNU/Linux 12 causes a segmentation fault, suspected to be due to version mismatches between libllvm, OpenGL, and TensorFlow. The fault disappears when TensorFlow is not imported.
• issues/109933

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: 2

Summarized Issues:

• Type Mismatch in tf.maximum with XLA Compilation: The tf.maximum function triggers a type mismatch error when called with mixed integer and float arguments during XLA compilation mode, although the same code executes without error in eager mode. This indicates a discrepancy in type handling between the two execution modes that affects compilation.
• issues/108305

• AttributeError Due to Protobuf Compatibility: Importing TensorFlow 2.18.0 with Protobuf 6.33.5 causes an AttributeError because the 'MessageFactory' object lacks the 'GetPrototype' attribute, leading to multiple warnings. This reflects compatibility conflicts between TensorFlow and the Protobuf version used.
• issues/109469

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: 18

Key Open Pull Requests

1. fix: Raise ValueError when max_output_size < 0 in tf.image.non_max_suppression: This pull request fixes issue #109150 by adding input validation to the max_output_size parameter in tf.image.non_max_suppression, ensuring that a ValueError is raised when a negative value is passed instead of causing a SIGABRT crash, thereby improving error handling and debugging clarity.

• URL: pull/109433

• Associated Commits: 3b8e5, ba263, d54e8, 5ca3d, ec13b

2. Fix int32 overflow in BiasAdd/BiasAddGrad GPU kernels: This pull request fixes an int32 overflow issue in the BiasAdd and BiasAddGrad GPU kernels by widening kernel iteration variables and thread parameters from int32_t to int64_t, adding overflow protection and proper bounds checking, and removing unnecessary CPU restrictions to handle tensors with more than 2^31 elements safely.

• URL: pull/109619

• Associated Commits: 7f9b7, 360f0, 68c43, 797a7

3. Fix MatrixSolve GPU handling for singular matrices: This pull request fixes the handling of singular matrices in the MatrixSolve operation on GPU by ensuring it consistently raises the standard "Input matrix is not invertible." error, aligning GPU behavior with CPU, and adds regression tests to cover this GPU error case.

• URL: pull/109950

• Associated Commits: 4da54, 0fc1e

Other Open Pull Requests

• Bug fix in gradient computation for tf.dynamic_stitch: This pull request fixes a decade-old correctness bug where gradients were incorrectly propagated to all writers for duplicate indices instead of only the last writer. It assigns unique IDs to input elements, uses dynamic_stitch to identify winning entries, and masks out gradients for overwritten entries to ensure accurate gradient flow in differentiable programs.
  • pull/109354

• Validation improvements for TensorFlow operations: Multiple pull requests add validation checks to prevent crashes and ensure consistent behavior. These include scalar checks for QuantizeAndDequantizeV3 when axis=-1 and Python-level validation in tf.reshape to reject rank-0 shape tensors across eager and tf.function modes.
  • pull/109355, pull/109356

• Memory management enhancements in eager execution: A maximum size limit of 1024 is introduced to the scalar tensor handle cache to prevent unbounded memory growth caused by caching many distinct scalar values. When the cache reaches capacity, all existing entries are cleaned up before inserting new ones, fixing a memory leak while maintaining caching benefits.
  • pull/109357

• Dynamic cost model for Grappler memory optimizer: The static hardcoded list of "cheap to recompute" operations is replaced with a dynamic cost model using OpLevelCostEstimator based on tensor shapes. This allows more accurate identification of inexpensive operations while falling back to the static list if shape inference fails or estimates are unreliable.
  • pull/109358

• CMake configuration and build system improvements: Several pull requests improve CMake configurations by requiring and locating protobuf headers and libraries, adding EXCLUDE_FROM_ALL to prevent unnecessary third-party compilation, fixing label image dependency linking, and adding install rules for the C interface shared library. These changes enhance build reliability and reduce build time and resource usage.
  • pull/109375, pull/109378, pull/109381, pull/109382

• OpenCL and delegate linking improvements: Dynamic OpenCL library loading is introduced using pkg-config to detect installations and set library names based on OpenCL versions, resolving loading issues on various Linux distributions. Additionally, linker options are added to ensure delegate create and delete functions for XNNPACK, external, and GPU delegates are resolved from shared libraries with platform-specific support.
  • pull/109377, pull/109379

• C API versioning and shared library support: Compile-time definitions for major, minor, and patch versions are set up along with shared library properties for version and soversion to ensure proper C API versioning. This also includes installing the C interface shared library and managing static versus shared library installations.
  • pull/109380, pull/109382

• Bug fixes and behavior improvements in core components: Fixes include adding a missing return statement in ScopeGuard::operator=(ScopeGuard&&) to prevent undefined behavior and ensuring that both branches of tf.cond in eager mode are validated regardless of execution path, allowing earlier bug detection before XLA compilation.
  • pull/109406, pull/109413

• Code quality and logging enhancements: User-facing print statements are replaced with structured logging across build and tool scripts, over 200 automated ruff fixes and safe manual style improvements are applied in the tensorflow/lite codebase, and a high-priority TODO issues list is generated to improve code clarity, style compliance, and logging best practices.
  • pull/109769

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: 6

Key Closed Pull Requests

1. Update requirements.in and setup.py.tpl and generate lock files: This pull request updates the requirements.in and setup.py.tpl files, generates updated lock files, upgrades Keras to version 3.12.0, and restores accidentally removed CUDA dependencies.

• URL: pull/109808

• Associated Commits: 89e39, 4085e, f21e6

• Associated Commits: 89e39, 4085e, f21e6

2. Update release notes for TensorFlow 2.21.0: This pull request updates the release notes for TensorFlow version 2.21.0 by initially inserting placeholder text and subsequently modifying the RELEASE.md file, with the intention that a Release Owner will complete the internal release notes before final submission.

• URL: pull/109558

• Associated Commits: 4ff72, c47c0

• Associated Commits: 4ff72, c47c0

3. Update version numbers for TensorFlow 2.21.0-rc0: This pull request updates the version numbers to 2.21.0-rc0 in key TensorFlow files including core/public/version.h, tools/pip_package/setup.py, and tensorflow/tensorflow.bzl, while also warning about and advising manual checks for lingering old version strings in other source files.

• URL: pull/109643

• Associated Commits: 4aa34, 6fd2f

• Associated Commits: 4aa34, 6fd2f

Other Closed Pull Requests

• Test reliability improvements: This pull request enables the --flaky_test_attempts option in the release branch of TensorFlow to improve the reliability of tests. It aims to reduce test flakiness by allowing multiple attempts for flaky tests, enhancing overall test stability.
• pull/109416

• Security vulnerability fix: This pull request addresses a medium-severity security vulnerability by replacing the insecure MD5 hash function used for seed generation in data_flow_ops.py with a more secure alternative. This change mitigates risks related to predictable randomness that could compromise machine learning computations in security-sensitive contexts.
• pull/109842

• Build configuration updates: This pull request updates the BUILD files to include necessary dependencies, ensuring proper build configuration for the TensorFlow project. These changes help maintain the integrity and correctness of the build process.
• pull/109896

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