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 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. [STAT:AWAITING TENSORFLOWER] [TYPE:FEATURE] python 3,14 support: This issue reports a bug where TensorFlow packages cannot be installed on Windows 11 using Python 3.14, resulting in errors indicating no matching distribution found. The user requests support for Python 3.14, highlighting the importance of upgrading due to security and performance improvements, and seeks updates on when TensorFlow will provide compatible builds for this Python version.

   • The comments clarify that the installation errors stem from incompatibility between TensorFlow and Python 3.14 on Windows, noting that nightly builds currently support Python 3.14 only on Linux and macOS; users express frustration over the lack of Windows support and reference ongoing discussions and merged pull requests aimed at resolving this issue.
   • Number of comments this week: 2

2. [TYPE:FEATURE] Support for Python 3.14: This issue requests the addition of support for Python 3.14 in TensorFlow, as the current versions are incompatible and users want to avoid downgrading their Python environments. The discussion confirms that Python 3.14 support has been integrated into the TensorFlow master branch and is available in the 2.22.0 release candidate and nightly builds, though it is not yet part of a stable release.

   • Comments include a suggestion to target Python 3.15 instead, confirmation that Python 3.14 support is present in pre-release versions, a note about the issue being a duplicate of another, and inquiries about the timeline for a stable release supporting Python 3.14.
   • Number of comments this week: 2

3. [STAT:AWAITING TENSORFLOWER] [TYPE:BUG] [COMP:XLA] [2.21.0] RealDiv accepts integer dtypes in its OpDef, but eager/graph have no integer kernel and reject, while XLA (jit_compile=True) silently does truncating integer division — a dtype-support / semantic-consistency gap: This issue highlights an inconsistency in TensorFlow's tf.raw_ops.RealDiv operation where integer data types are accepted in the operation definition but lack kernel support in eager and graph execution modes, causing errors, while XLA silently performs truncating integer division without error. The discrepancy leads to a semantic gap where the same integer inputs are rejected in some modes but processed with unintended truncation in XLA, and the issue requests a resolution for consistent behavior or explicit documentation of integer division semantics.

   • The comments confirm the problem and provide a detailed investigation tracing the cause to the OpDef and kernel registrations; a proposed fix involves constraining the XLA registration and MLIR patterns to floating-point and complex types, with an offer to submit a pull request including the fix and regression tests.
   • Number of comments this week: 2

4. [TYPE:BUG] [COMP:OPS] [AWAITING PR MERGE] [2.21.0] ResourceAccumulatorApplyGradient aborts on accumulator/gradient dtype mismatch: This issue reports that calling tf.raw_ops.ResourceAccumulatorApplyGradient with a gradient of type float16 and a ResourceConditionalAccumulator of type float32 causes the Python process to abort due to a dtype mismatch, instead of raising a recoverable TensorFlow exception. The user provides a minimal reproducible example and requests confirmation on whether an existing pull request addressing a similar problem also covers this ConditionalAccumulator path.

   • The comments acknowledge the problem, confirm that the issue causes a fatal process abort due to continued execution after a signature mismatch error, and inform that a fix has been proposed in a pull request which, once merged, will resolve the issue.
   • Number of comments this week: 2

5. [TYPE:FEATURE] [COMP:OPS] [AWAITING PR MERGE] [TF 2.11] how to assign value to a EagerTensor slice? ----'tensorflow.python.framework.ops.EagerTensor' object does not support item assignment: This issue addresses the inability to assign values directly to slices of a TensorFlow EagerTensor, which raises an error because EagerTensors are immutable and do not support item assignment like NumPy arrays or PyTorch tensors. The discussion explores various workarounds such as converting tensors to variables, using tensor_scatter_nd_update, or employing TensorArray, while highlighting the demand for native support of in-place tensor slice assignment to improve usability and performance in TensorFlow.

   • The comments include requests for reproducible code, confirmation of the issue across TensorFlow versions, suggestions of workarounds like converting to variables or using tensor_scatter_nd_update, expressions of frustration over the lack of native support, comparisons to PyTorch, ongoing discussions about potential fixes and RFCs, and updates indicating a proposed fix is under review but not yet merged.
   • 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: 10

Summarized Issues:

• Data type handling inconsistencies: Several issues highlight problems with TensorFlow operations mishandling data types, leading to crashes or incorrect behavior. These include inconsistent dtype support in RealDiv across execution modes, fatal errors from dtype mismatches in ResourceAccumulatorApplyGradient, and integer overflow causing crashes in tf.gather_nd with int16 indices.
• issues/129335, issues/129358, issues/129561

• Operation crashes and aborts: Multiple TensorFlow operations cause the Python process to abort unexpectedly due to fatal CHECK failures or segmentation faults. Examples include tf.raw_ops.ImportEvent failing on multi-element string tensors, SparseMatrixAdd crashing with complex128 on RTX 4090 GPUs, and tf.raw_ops.For loops aborting on GPU with CUDA errors.
• issues/129384, issues/129392, issues/129396

• Static shape and tensor inference issues: TensorFlow's static shape inference and tensor folding mechanisms exhibit bugs that cause incorrect or missing static values. Specifically, tf.get_static_value() fails to fold StridedSlice operations in graph mode, returning None instead of a static value, which complicates shape inference.
• issues/129438

• Signal processing function bugs: The tf.signal.overlap_and_add function incorrectly handles inputs with zero frames and certain frame lengths, producing non-empty zero-filled tensors instead of empty tensors. This leads to downstream errors such as fake audio samples when reconstructing from empty spectrograms using tf.signal.inverse_stft.
• issues/129484

• Numerical computation errors on CPU: The tf.matmul operation on rank-3 float32 tensors produces incorrect and nondeterministic results on CPU under specific conditions involving large contracted dimensions and batch sizes. This issue is traced to the oneDNN batched-matmul path and does not occur when oneDNN optimizations are disabled.
• issues/129498

• Build process failures due to checksum mismatch: Building TensorFlow 2.21.0 from source fails because of a checksum mismatch when downloading the LLVM archive. The expected SHA256 hash does not match the actual archive, causing Bazel to halt the build process due to verification failure.
• issues/129543

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: 51

Summarized Issues:

• Gradient computation errors and missing gradient registrations: Several issues report failures or incorrect results in gradient computations due to missing gradient registrations or improper handling of edge cases. These include LookupErrors for missing gradient entries in UpSampling2D, incorrect gradients for sparse segment operations with empty inputs, and inability to compute gradients for functions like tf.linalg.expm under JIT compilation, all of which disrupt model training workflows.
• [issues/43648, issues/72785, issues/77693]

• Crashes and core dumps due to input validation failures: Multiple bugs cause TensorFlow to crash or abort with core dumps when invalid inputs are provided or when internal checks fail. Examples include crashes from negative thread counts in set_intra_op_parallelism_threads, segmentation faults from invalid seq_len_max in BlockLSTMGradV2, and aborts caused by tf.raw_ops.LoopCond in debug builds. These highlight missing or insufficient input validation in critical TensorFlow components.
• [issues/62982, issues/73518, issues/106407]

• Inconsistent or incorrect function outputs across devices or execution modes: Several issues describe discrepancies in function outputs depending on hardware (CPU vs GPU) or execution mode (eager vs graph or XLA). For instance, bucketize returns wrong results on GPU, tf.nn.conv1d behaves differently in eager and graph modes, and tf.searchsorted produces inconsistent indices under XLA JIT. Such inconsistencies complicate debugging and reduce reliability.
• [issues/70343, issues/112371, issues/118381]

• CUDA and GPU kernel bugs causing illegal memory accesses and performance issues: There are reports of CUDA kernel bugs causing illegal out-of-bounds memory reads or writes during operations like BiasAddGrad and DepthToSpace on GPU, as well as long compilation times due to missing CUDA kernel binaries for new GPU architectures. These issues lead to runtime errors and degraded performance on GPU platforms.
• [issues/109515, issues/128999, issues/97387]

• Numerical inaccuracies and incorrect special function implementations: Several TensorFlow math functions return incorrect or inconsistent values compared to mathematical definitions or other libraries. Examples include tf.math.igamma returning 0 instead of NaN, tf.raw_ops.Digamma returning NaN instead of negative infinity, and tf.keras.ops.numpy.logaddexp producing wrong gradients. These inaccuracies affect scientific computing reliability.
• [issues/108641, issues/111917, issues/127232]

• XLA compiler bugs causing crashes, heap corruption, and inconsistent behavior: Multiple issues describe bugs in TensorFlow's XLA compiler where invalid inputs or dynamic shapes cause heap corruption, crashes, or silent incorrect results. Examples include tf.slice with out-of-range indices causing heap corruption, silent broadcasting of incompatible MatMul dimensions, and missing broadcast checks in element-wise binary ops. These undermine the safety and correctness of XLA-compiled models.
• [issues/110789, issues/128403, issues/128405, issues/129312]

• Shape inference and tracing errors in graph mode: Some bugs arise from TensorFlow's inability to correctly infer shapes or trace functions when input shapes are unknown or dynamic. For example, tf.linalg.set_diag fails to trace with partially known shapes, and tf.linalg.eigh_tridiagonal does not support symbolic tensors with unknown shapes, causing errors during graph construction.
• [issues/128429, issues/129061]

• Incorrect or missing error messages and warnings: Several issues request improved error handling or clearer messages, such as replacing confusing reshape errors in ForwardAccumulator with explicit ValueErrors, or addressing protobuf version mismatch warnings that do not affect functionality but may cause future incompatibilities. These improvements would aid developers in diagnosing problems.
• [issues/98980, issues/99160]

• TensorFlow package and build environment limitations: There are requests and reports about missing ARM64 wheels for CPU packages affecting CI/CD workflows, and build issues related to CUDA kernel compatibility with new GPU architectures, highlighting gaps in platform support and build configurations.
• [issues/106021, issues/97387]

• Bugs in TensorFlow core operations causing crashes or incorrect behavior: Various core ops exhibit bugs such as segmentation faults during attribute deletion, overflow in slice index calculations, and incorrect handling of signed zeros in scalar caches, leading to crashes or subtle computational errors.
• [issues/121923, issues/126125, issues/129231]

• Experimental numpy API inconsistencies and bugs: TensorFlow's experimental numpy functions show multiple issues including incorrect rounding, failure to promote integer inputs to float, incorrect handling of boolean inputs, and shape mismatches in window functions, causing deviations from NumPy's behavior and unexpected errors.
• [issues/129232, issues/129233, issues/129234, issues/129235, issues/129236, issues/129237, issues/129238, issues/129239]

• TensorList and tensor scattering bugs causing type mismatches and crashes: A bug in TensorListScatterIntoExistingList creates intermediate slots with incorrect dtypes when scattering into empty lists at non-contiguous indices, causing type errors and process termination during subsequent TensorList operations.
• [issues/129348]

• Debugger and tracing infrastructure bugs causing failures on reinitialization: Re-enabling Debugger V2 with the same dump root fails due to uninitialized graph execution traces writer, resulting in FailedPreconditionError, though using a fresh dump root works correctly.
• [issues/129409]

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: 35

Key Open Pull Requests

1. [ROCm] Enable hermetic build for TF under rocm configuration: This pull request enables a fully hermetic build environment for TensorFlow under the ROCm configuration by activating hermetic LLVM and ROCm dependencies, allowing TensorFlow to be built on any machine without requiring pre-installed ROCm components, thereby facilitating long-term support and integration with the ROCm fork of TensorFlow.

• URL: pull/129444

• Associated Commits: 13b3f, 37035, 25856, 7c715

2. Fix tf.experimental.numpy.cross axis-bound validation with NumPy-style AxisError messages: This pull request fixes an axis-bound validation gap in tf.experimental.numpy.cross by adding upfront checks that raise clear, NumPy-style AxisError messages for out-of-bounds axis parameters (axisa, axisb, axisc, or axis), ensuring errors are caught before any computation and improving error clarity and consistency without affecting valid-axis behavior or performance during tracing and eager execution.

• URL: pull/129378

• Associated Commits: a9878, a196b, 89a56

3. Validate single-element inputs in summary kernels: This pull request improves TensorFlow summary kernels by adding validation to ensure that certain inputs expected to contain exactly one element—such as those in ImportEvent, WriteSummary, WriteScalarSummary, WriteHistogramSummary, WriteImageSummary, WriteGraphSummary, and CreateSummaryDbWriter—are checked with a helper function that raises an InvalidArgumentError for empty or multi-element tensors instead of causing a process abort, thereby enhancing robustness and consistency in input handling.

• URL: pull/129416

• Associated Commits: e3a06, 1b547, 324b9

Other Open Pull Requests

• Gradient and Numerical Stability Fixes: Multiple pull requests address numerical stability and gradient correctness issues in TensorFlow operations. These include fixing the sigmoid gradient for float64 to avoid zero gradients in the positive tail, improving float16 and bfloat16 norm and normalization computations by performing calculations in float32, and correcting the gradient calculation for tf.pow when both base and exponent are zero to prevent NaNs.
• [pull/129437, pull/129512, pull/129513, pull/129514]

• TensorBundle and Checkpoint Validation Enhancements: Several pull requests improve the robustness of TensorBundle and checkpoint handling by validating tensor element counts, slice extents, and data type consistency to prevent memory corruption and out-of-bounds writes. These changes add checks against corrupted or untrusted V2 checkpoints and include comprehensive unit tests to ensure reliability.
• [pull/129496, pull/129488, pull/129495]

• RaggedTensor and Sparse Tensor Kernel Hardening: Pull requests enhance the safety and correctness of RaggedTensor and sparse tensor operations by adding overflow checks, dimension validations, and datatype handling improvements. These fixes prevent integer overflows, invalid memory accesses, and type mismatch crashes during decoding and kernel execution.
• [pull/129508, pull/129516]

• Input Validation and Error Handling Improvements: Multiple pull requests add or improve validation logic to prevent silent failures or process aborts. This includes validating contracted axis sizes in tf.tensordot, checking compression options in TFRecordWriter, rejecting invalid tensor content buffers, and adding axis validation tests for tf.experimental.numpy.ptp.
• [pull/129511, pull/129373, pull/129517, pull/129379]

• Bug Fixes in Core TensorFlow Operations: Several fixes address bugs in core TensorFlow functionality such as accumulator apply-gradient kernels aborting after invalid inputs, tf.signal.overlap_and_add returning incorrect signal lengths, and a wild pointer read in the strong_hash feature causing segmentation faults.
• [pull/129359, pull/129493, pull/129391]

• Image and JPEG Decoding Fixes: Pull requests fix issues related to image processing including correcting pixel value conversions for half-precision images in convert_image_dtype and fixing signed integer overflow in JPEG crop window validation to prevent crashes during decoding.
• [pull/129510, pull/129527]

• Static Shape and Constant Folding Enhancements: Improvements include fixing static shape inference for MaxPool operations with small inputs and adding support for folding StridedSlice operations in constant value inference to enable static evaluation of indexed constants in graph mode.
• [pull/129497, pull/129440]

• Memory Safety and Security Fixes: A security vulnerability is fixed by capping implicit resize operations in TensorListSetItem to prevent excessive memory allocation and potential crashes, with regression tests added to ensure enforcement.
• [pull/129488]

• Rounding and Conversion Accuracy Tests: Regression tests are added to verify correct midpoint rounding behavior in double-to-bfloat16 conversions, ensuring accuracy after recent Eigen updates.
• [pull/129529]

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: 66

Key Closed Pull Requests

1. Introduce tf.distribute.AutoStrategy: This pull request introduces the tf.distribute.AutoStrategy API, a zero-configuration distributed training optimizer that automatically detects the hardware environment—including multi-worker clusters, TPUs, multiple GPUs, single GPUs, or CPUs—and dynamically selects and initializes the optimal TensorFlow distribution strategy to simplify and streamline distributed machine learning setup.

• URL: pull/129224

• Associated Commits: caf37, f0274, ba05b, 5d112, f4d74, e2c7c, 47baf, 41a0b, a41b4, 28f85, 51f01, 8e24e, a42bf, f6337, 6a850, 0846a, 1a0ac, 34ab1, 30814, 911e4, 7bcb0, 1255e, 50534, 49460

2. Validate tensor rank in WriteAudioSummary: This pull request addresses the issue where tf.raw_ops.WriteAudioSummary aborts the process when given tensors that are not 2-D or 3-D by adding explicit validation in WriteAudioSummaryOp::Compute to return an InvalidArgumentError for tensors with invalid ranks, thereby preventing crashes and ensuring only tensors with shapes [batch_size, num_samples] or [batch_size, num_samples, num_channels] are accepted, along with adding comprehensive tests to verify correct rejection and acceptance of various tensor ranks.

• URL: pull/128257

• Associated Commits: 71927, 13f57, 74058, b00f9, 57fc6, b0225, c5eb3, f7ba6, 30bce, 80886, 4b0c2, 6c53d, 23395, 95adf, 95596, 1c5b1, cd19d, e4f4a, 554fa, d6d4d, 1a4dd, e1a3e, 85d52

3. Fix CWE-770: Add bounds checks to DecodeGifV2 and DecodeJpegV2 (consistent with BMP/PNG/WebP decoders): This pull request enhances the TensorFlow GIF and JPEG image decoders by adding pre-allocation bounds checks consistent with other decoders to prevent excessive memory allocation from pathological image headers, implements early rejection of oversized logical screens in GIFs, maintains a post-decode output size ceiling, and introduces deterministic regression and OSS-Fuzz tests to improve security and robustness against out-of-memory vulnerabilities.

• URL: pull/115498

• Associated Commits: 974d4, 9f5aa, 8e338, fe2be, 41a20, 4a2ef, 43dae, 43b1e, 66bd8, 9a659, 41e62, 68630, 22005, b125a

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.