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 TensorFlow Lite (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 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:BUG] tf.math.erf returns nan for inf and -inf, inconsistent with PyTorch/SciPy: This issue reports that the TensorFlow function tf.math.erf returns NaN when given positive or negative infinity as input, which is inconsistent with PyTorch, SciPy, and the Python math library that correctly return 1.0 and -1.0 respectively. The expected behavior is for TensorFlow to handle infinite inputs explicitly and return ±1.0 to align with mathematical correctness and other standard libraries.

   • The comments confirm the inconsistency and suggest the root cause is missing explicit handling of infinite inputs in TensorFlow’s implementation. Multiple users express interest in fixing the issue by adding special case handling for ±inf to return ±1.0, with one user indicating they will submit a pull request soon.
   • Number of comments this week: 3

2. [TYPE:BUG] [COMP:OPS] [2.17] Crash in tf.compat.v1.train.shuffle_batch when enqueue_many=True with zero-sized dimension: This issue reports a crash occurring in TensorFlow when using the tf.compat.v1.train.shuffle_batch function with enqueue_many=True on input tensors that have a zero-sized dimension, causing a segmentation fault instead of a Python-level error. The user provides a reproducible code snippet and confirms the problem persists across TensorFlow versions 2.17.0 and 2.20.0, seeking clarification on whether an error should be raised instead of crashing.

   • The comments confirm the crash can be reproduced using the provided code snippet, with one user offering to help fix the issue if it is agreed that a Python error should be raised instead of a crash. Another user confirms the issue occurs on multiple TensorFlow versions and shares a reference gist for further investigation.
   • Number of comments this week: 3

3. [TYPE:BUG] [COMP:OPS] [TF 2.19] tf.eye with TF_ENABLE_ONEDNN_OPTS=1 causes crash (Aborted/SegFault) on large num_columns instead of OOM: This issue describes a bug where using tf.eye with an extremely large num_columns value causes TensorFlow to crash with a segmentation fault or abort signal when oneDNN optimizations are enabled, instead of raising a proper out-of-memory exception. The problem arises because the oneDNN kernel implementation does not properly check for integer overflow or memory exhaustion before attempting allocation, leading to a crash rather than a catchable Python exception.

   • The comments clarify that tf.eye only supports int32-sized dimensions and values larger than 2^31 - 1 are unsupported; while this limitation is known, the user expects a graceful exception rather than a crash, highlighting the need for improved error handling in this edge case.
   • Number of comments this week: 2

4. [TYPE:BUG] [COMP:LITE] arrch64 tflite 2.19/2.18 build failed when xnnpack is enabled.: This issue reports a build failure for TensorFlow Lite versions 2.19 and 2.18 on the arrch64 architecture when the xnnpack feature is enabled, caused by a SHA256 hash mismatch for the kleidiai package during the build process. The problem results in the build system repeatedly removing the downloaded kleidiai archive due to the hash not matching the expected value, preventing successful compilation.

   • The single comment confirms that this issue also disrupts the litert_1.2.0 recipe in the NXP Yocto builds, indicating the problem affects other related build environments and is currently blocking those workflows.
   • Number of comments this week: 1

5. [TYPE:BUG] tf.experimental.numpy.isclose returns incorrect results for integer inputs: This issue reports a bug in TensorFlow's tf.experimental.numpy.isclose function where it returns incorrect results when given integer inputs, unlike the expected behavior consistent with NumPy's implementation. The problem arises because the internal handling of integer arithmetic in this function is flawed, causing all comparisons to return False, whereas casting inputs to floating-point types yields correct results.

   • The single comment notes that a pull request has been made to fix the integer input behavior in tf.experimental.numpy.isclose, ensuring it matches NumPy's output.
   • 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: 15

Summarized Issues:

• Mathematical function inconsistencies: Several TensorFlow math functions return incorrect or inconsistent results compared to other libraries. For example, tf.math.erf returns NaN for infinite inputs instead of ±1.0, and tf.math.igamma returns 0 instead of NaN for certain negative inputs, indicating edge-case handling issues.
• [issues/108483, issues/108641]

• XLA JIT compilation bugs: Using XLA JIT compilation causes various functions to fail or raise errors due to unsupported operations or type mismatches. This includes errors with Keras model names containing '@', tf.nn.dropout seed parameter type issues, and failure of tf.image.resize with bicubic interpolation under JIT.
• [issues/108501, issues/108502, issues/108660]

• Segmentation faults and crashes due to input validation failures: TensorFlow crashes or segfaults occur when invalid or extreme inputs are passed without proper validation. Examples include crashes with zero-sized dimensions in shuffle_batch, negative parameters in AudioSpectrogram, extremely large dimensions in ImmutableConst, and large num_columns in tf.eye with oneDNN enabled.
• [issues/108489, issues/108621, issues/108664, issues/108665]

• Integer overflow and assertion failures with large tensor dimensions: Operations involving very large tensor shapes cause integer overflows leading to fatal assertion failures or process aborts instead of proper Python errors. This affects SparseConcat, ParallelConcat, and summary writer usage with multi-element step tensors.
• [issues/108659, issues/108661, issues/108663]

• Build and environment compatibility issues: TensorFlow Lite fails to build on aarch64 with XNNPACK enabled due to a SHA256 mismatch, and TensorFlow DLL loading fails in Java applications on Windows 11 with versions 2.19 and 2.20, indicating environment-specific compatibility problems.
• [issues/108497, issues/108594]

• Incorrect behavior in numpy compatibility functions: The TensorFlow implementation of tf.experimental.numpy.isclose returns incorrect results for integer inputs, failing to match NumPy's expected behavior.
• [issues/108657]

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: 13

Summarized Issues:

• XLA Compilation Errors: Multiple issues describe errors occurring during XLA compilation, including a NotImplementedError caused by calling numpy() on symbolic tensors inside tf.function with jit_compile=True, and a type error when using tf.image.extract_patches with tensor-valued sizes instead of integers. These problems highlight limitations and strict type requirements in XLA compilation that do not appear in eager execution mode.
• issues/105728, issues/108143

• Parameter Validation Bugs: Several issues report inconsistent or missing validation of function parameters, such as tf.concat not validating the axis parameter properly when concatenating tensors, tf.compat.forward_compatible allowing invalid date values without error, and tf.compat.as_str_any silently accepting invalid encoding parameters. These validation gaps lead to confusing errors or silent failures that contradict documented behavior.
• issues/108521, issues/108548, issues/108608, issues/108628

• TensorFlow Function Return Type Issues: There are reports of functions returning unexpected types, such as tf.compat.dimension_value returning a tf.Tensor instead of an int or None as documented, causing discrepancies between actual and expected return types. This inconsistency can lead to bugs when users rely on the documented function signatures.
• issues/108605

• TensorFlow Crashes and Inconsistent Behavior: Some issues describe crashes or inconsistent behavior, including TensorFlow crashing when tf.experimental.numpy.transpose is called with invalid axes, and tf.experimental.numpy.asarray producing inconsistent int8 results when casting from bfloat16 on CPU versus GPU. These bugs indicate incomplete error handling and platform-dependent discrepancies.
• issues/108488, issues/108629

• Installation and Environment Issues: One issue reports a failure to install TensorFlow on Windows 11 using pip due to no matching distribution found, suggesting compatibility or availability problems with the environment.
• issues/108503

• Documentation and Metadata Errors: A typo in the BertInputTensorsMd.md file where "tesnor" was used instead of "tensor" could cause bugs related to metadata properties, indicating the need for careful documentation review.
• issues/108471

• Spam Content Removal: An issue addresses the removal of spam content posted by a user identified as a shameless spammer in the TensorFlow repository, reflecting maintenance and moderation efforts.
• issues/108426

• Insufficient Issue Details: One issue reports a bug with TensorFlow Nightly and custom code but lacks detailed environment and error information, limiting the ability to diagnose or reproduce the problem.
• issues/108640

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.

1. tf.compat.dimension_value returns a Tensor when given a Tensor, despite documentation stating it returns only int or None.
   • Toxicity Score: 0.65 (Rapid escalation, sarcastic tone, dismissive language)
   • This GitHub conversation begins with a user reporting unexpected behavior in a TensorFlow function, providing detailed observations and reproducible code. Another participant responds with a sarcastic tone, dismissing the issue as insignificant and urging to stop filing such bugs, which introduces tension. The exchange reflects a mix of technical concern and frustration, with the latter comment potentially escalating conflict.

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: 10

Key Open Pull Requests

1. Fix capitalization in ApiDefMap docstring: This pull request fixes the capitalization of Tf_ApiDefMap to TF_ApiDefMap in the docstring to align with the actual C API naming convention used elsewhere in the codebase, addressing a documentation inconsistency without affecting functionality.

• URL: pull/108597

• Associated Commits: 04815, b902f

2. Improve error message for tf.image.extract_patches with XLA: This pull request improves the error message for the tf.image.extract_patches operation when used with XLA JIT compilation by introducing a helper function that detects non-constant windowing parameters and raises a clear, descriptive TypeError explaining the requirement for compile-time constant values, thereby replacing a previously cryptic error and enhancing user understanding.

• URL: pull/108476

• Associated Commits: 827a1

3. Fixed rfft backprop shape mismatch when fft_length < input shape: This pull request fixes the gradient computation for the tf.signal.rfft operation in TensorFlow when the fft_length parameter is smaller than the input shape, resolving a shape mismatch error during backpropagation by correctly handling truncated inputs and adding test coverage for multiple fft_length scenarios.

• URL: pull/108479

• Associated Commits: 185ff

Other Open Pull Requests

• GPU work-group size and buffer alignment fixes: This pull request adjusts the GPU work-group size calculation to respect maximum X-dimension limits and removes Adreno-specific pitch alignment logic to simplify buffer size calculations. These changes improve compatibility across GPUs and prevent oversized work-groups and incorrect buffer alignment on Adreno devices.
• pull/108500

• Typographical and unreachable code cleanup: These pull requests fix a typographical error and remove an unreachable return statement that was redundant because it was always executed beforehand. Both changes contribute to code clarity and maintainability by eliminating minor errors and dead code.
• pull/108532, pull/108587

• Fixes to mathematical function behavior: This pull request corrects the tf.math.erf function to return ±1 for ±infinity inputs instead of NaN by introducing a custom functor, aligning its behavior with PyTorch, SciPy, and Python's math library. It also includes comprehensive tests for float32 and float64 data types to verify this fix.
• pull/108592

• Integer overflow prevention in CUDA kernels: This pull request addresses integer overflow vulnerabilities in six CUDA kernel files by replacing unsafe int32 multiplications with a safer MultiplyWithoutOverflow() function that uses int64 space. This change enables proper error handling to prevent memory corruption and crashes when processing large tensors.
• pull/108695

• Fixes to numpy compatibility and segmentation fault prevention: These pull requests fix the tf.experimental.numpy.isclose function to correctly handle integer inputs matching NumPy's behavior while preserving float input behavior, and prevent segmentation faults in tf.raw_ops.ImmutableConst by avoiding size calculation overflow for extremely large shape dimensions. Both include new tests or error handling to ensure robustness.
• pull/108702, pull/108705

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: 31

Key Closed Pull Requests

1. Reshape output according to batch size: This pull request proposes reshaping the output dynamically according to the batch size instead of using padded output sizes, aiming to improve the handling of output dimensions before returning results.

• URL: pull/108590

• Associated Commits: 92148, 66aca, 5a423, bf0e7, a7902, f2d6e, 204a6, 4ab00, 69e6c, f1e7b, 8ba71, 893c4, f0235, de2b8, eca0e, d4abb, 439a1, ccdb4, cd1b1, ad321, 3bb90, 20855, 039b2, 12630, 33a78, 4ef54, a768c, 044df, 4b10c

• Associated Commits: 92148, 66aca, 5a423, bf0e7, a7902, f2d6e, 204a6, 4ab00, 69e6c, f1e7b, 8ba71, 893c4, f0235, de2b8, eca0e, d4abb, 439a1, ccdb4, cd1b1, ad321, 3bb90, 20855, 039b2, 12630, 33a78, 4ef54, a768c, 044df, 4b10c

2. Fix tf.case indexing inconsistency in XLA/Graph mode (Issue #108304): This pull request fixes an inconsistency in TensorFlow's XLA and Graph modes by modifying the tensor indexing logic to support advanced indexing (such as list or tuple indices) for Symbolic Tensors via a fallback to _numpy_style_getitem, thereby resolving TypeError exceptions when using list indices inside tf.function and tf.case and improving parity with Eager execution.

• URL: pull/108481

• Associated Commits: 37965, b641d, 77e66, 6e023, 0e2be, 5a372, f86bc

• Associated Commits: 37965, b641d, 77e66, 6e023, 0e2be, 5a372, f86bc

3. ci: disable legacy Java artifact repacking: This pull request disables the repacking of legacy Java artifacts in the repack_libtensorflow.sh script as part of deprecating the legacy Java client, including commenting out related copy commands and updating the Windows zip command to prevent build failures, while also cleaning up legacy JNI archives and build rules.

• URL: pull/108284

• Associated Commits: fee7c, 65ab7, 0d911, 2e214, 933cc, 9ec04

• Associated Commits: fee7c, 65ab7, 0d911, 2e214, 933cc, 9ec04

Other Closed Pull Requests

• Validation and Error Handling Improvements: Multiple pull requests enhance validation and error handling across TensorFlow operations. These include adding rank validation to BlockLSTMGrad to prevent segmentation faults, fixing tf.TensorArray.stack to propagate element shapes correctly with XLA, validating axis in tf.concat as a scalar int32 tensor, and adding encoding validation in tf.compat.as_str_any to raise appropriate errors. Additionally, crash issues in tf.experimental.numpy.transpose are addressed by validating perm/axes inputs and normalizing negative axes to prevent process aborts.
• [pull/108361, pull/108490, pull/108486, pull/108485, pull/108493, pull/108498]

• Build System and Annotation Cleanup: Several pull requests focus on cleaning up the TensorFlow build system and standardizing annotations. This includes removing legacy pywrap_rules logic, replacing deprecated @local_xla and @local_tsl annotations with @xla and @tsl, fixing local_xla references in build files, and updating TOSA build rules to use @xlaTested. A temporary disabling of a failing gRPC patch was also done to fix a macOS ARM64 build failure.
• [pull/108492, pull/108708, pull/108699, pull/108751, pull/108752, pull/108474, pull/108475]

• Documentation and Docstring Enhancements: Multiple pull requests improve documentation clarity and correctness. These include fixing typos in the BertInputTensorsMd.md file, correcting a docstring typo in ScopedTFImportGraphDefResults, adding comprehensive docstrings to include_frame and _process_traceback_frames in traceback_utils.py, and adding detailed docstrings to _infer_output_signature in base_layer.py. Documentation was also added to clarify that static configuration values in Keras layers should be Python primitives to avoid XLA errors.
• [pull/108472, pull/108473, pull/108491, pull/108494, pull/108519, pull/108598, pull/108602]

• Bug Fixes and Memory Leak Resolution: Some pull requests address bugs and memory leaks in TensorFlow components. This includes fixing a missing return statement in the mlir::TFTPU error check, resolving a memory leak in the @custom_gradient decorator by modifying gradient function closures to avoid capturing Tensor objects unnecessarily, and adding regression tests to prevent Tensor accumulation during repeated use.
• [pull/108201, pull/107094, pull/107197]

• TOSA Conversion and Validation Enhancements: One pull request enhances the tosa-convert-tfl-uint8 pass by supporting unsigned to signed integer conversion for any bitwidth, adding validation for bitwidths and zero-point restrictions, and replacing deprecated quantized type construction methods to comply with TOSA specifications.
• [pull/107197]

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.