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 Python with a migration path to ai_edge_litert.interpreter, and changes to certain C++ API constants for better compatibility. Key improvements include runtime support for the bfloat16 data type in the tfl.Cast operation, alongside the discontinuation of separate 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:CONTRIBUTION WELCOME] [TYPE:FEATURE] [COMP:DATA] Unspecific error message "None values not supported" when no labels are provided in dataset: This issue reports that when a dataset is missing labels in TensorFlow Keras code, the resulting error message "None values not supported" is vague and does not clearly indicate that the problem is due to missing labels or null y_true values. The user suggests improving the error message to explicitly reference the absence of labels to make debugging easier, and notes that they resolved the issue with the help of an AI assistant.

   • The comments discuss whether the error message falls under documentation, suggest providing labels to avoid the error, and clarify that the issue is not currently prioritized but open for community contributions. One commenter volunteers to work on it, and later updates that the root cause is in the separate Keras repository, so the fix will be submitted there instead.
   • Number of comments this week: 2

2. [STAT:CONTRIBUTION WELCOME] [TYPE:BUG] [COMP:OPS] [COMP:GPU] [2.20.0] Inconsistency on CPU and GPU float16 std: This issue reports a bug involving inconsistent results when computing the standard deviation of float16 tensors on CPU versus GPU using TensorFlow, where the CPU returns NaN and the GPU returns infinity. The user has reproduced the problem with TensorFlow Nightly and provided code demonstrating the discrepancy in behavior between devices.

   • The comments confirm the issue was reproduced on Colab with newer TensorFlow versions, provide a reference gist, and indicate that the issue is not currently prioritized but open for community contributions; one user expressed interest in working on a fix.
   • Number of comments this week: 2

3. [STAT:CONTRIBUTION WELCOME] [TYPE:BUG] [COMP:GPU] [2.20.0] tf.math.unsorted_segment_max returns empty-segment sentinel for NaN-containing non-empty segment on GPU: This issue reports a bug in TensorFlow where the function tf.math.unsorted_segment_max returns an empty-segment sentinel value instead of the expected result when processing segments containing NaN values on a GPU, while the CPU returns the correct output. The user has provided reproducible code and confirmed the bug occurs in TensorFlow version 2.20 and nightly builds.

   • The comments confirm the issue was reproduced on Colab with TensorFlow 2.21.0 and nightly builds, and the maintainers have marked the issue as a community contribution opportunity, encouraging users to submit pull requests; one user expressed willingness to work on a fix if the original reporter does not.
   • Number of comments this week: 2

4. [STAT:CONTRIBUTION WELCOME] [TYPE:BUG] [COMP:KERAS] [TF 1.15] serialize/deserialize round-trip gives different results in eager mode vs inside tf.function: This issue reports a bug where serializing and deserializing a Keras layer preserves the configuration correctly in eager execution mode but fails to do so when the same operation is performed inside a tf.function compiled graph, resulting in different outputs. The expected behavior is that the serialization round-trip should yield identical results regardless of whether it runs eagerly or within a compiled function, or at least raise a clear error if unsupported.

   • The comments show a contributor expressing interest in investigating the issue by reproducing it on current TensorFlow versions and adding regression tests, followed by a maintainer welcoming the contribution and advising to follow contribution guidelines and include tests with any fix.
   • Number of comments this week: 2

5. [STAT:CONTRIBUTION WELCOME] [TYPE:BUG] [COMP:KERAS] [TF 1.15] [COMP:TF.FUNCTION] tf.keras.layers.serialize returns a different config for the same layer in eager mode vs inside tf.function: This issue reports a bug where the tf.keras.layers.serialize function returns different configuration dictionaries for the same layer when called eagerly versus inside a tf.function, which is unexpected behavior. The user provides a minimal reproducible example demonstrating that serialization outputs differ depending on the execution mode, and expects consistent serialization results regardless of context.

   • The comments show a user expressing interest in working on the issue and being encouraged to proceed after reviewing related discussions and contribution guidelines, with a reminder to include appropriate tests in any proposed fix.
   • 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: 24

Summarized Issues:

• Serialization inconsistencies in tf.keras layers: Serialization and deserialization of Keras layers produce inconsistent configurations when performed eagerly versus inside a tf.function, leading to different outputs depending on the execution mode. This inconsistency affects both the preservation of layer configurations and the output of the tf.keras.layers.serialize function.
• issues/127003, issues/127005

• Debug mode causing dataset iteration errors: Enabling TensorFlow's debug mode triggers failures in tf.data.Dataset.from_generator due to callback registry errors after garbage collection, causing iteration errors that do not occur without debug mode enabled. This bug disrupts normal dataset iteration workflows under debug conditions.
• issues/127022

• Integer overflow in ScatterNd operation: The ScatterNd operation suffers from an integer overflow bug where int32 accumulation of output indices leads to out-of-bounds memory writes on CPU and GPU for large tensor shapes. This results in crashes without Python-level exceptions, posing a serious stability risk.
• issues/127164

• Incorrect higher-order derivatives in various functions: Multiple TensorFlow functions including tf.linalg.expm, tf.keras.ops.numpy.sinc, tf.experimental.numpy.vander, tf.keras.ops.hypot, tf.keras.ops.numpy.logaddexp, tf.math.cumulative_logsumexp, tf.keras.ops.numpy.ldexp, and tf.signal.kaiser_window produce incorrect second derivatives or gradients under nested or reverse-mode automatic differentiation. These errors manifest as NaNs, incorrect values, or missing derivative terms, affecting the accuracy of gradient-based computations.
• issues/127224, issues/127226, issues/127227, issues/127230, issues/127232, issues/127243, issues/127244, issues/127245

• Gradient calculation errors in automatic differentiation: Several TensorFlow functions such as tf.math.segment_mean, tf.nn.isotonic_regression, tf.raw_ops.MatrixDiagV2, tf.math.xdivy, tf.math.xlogy, and tf.keras.ops.image.affine_transform return incorrect gradients or Jacobian-vector products during forward or reverse-mode automatic differentiation. These issues cause inaccurate derivative values despite correct forward computations, impacting model training and optimization.
• issues/127233, issues/127234, issues/127235, issues/127238, issues/127240, issues/127241

• Incorrect outputs in image processing ops: The tf.keras.ops.image.gaussian_blur and tf.keras.ops.numpy.rot90 functions produce incorrect outputs due to parameter misinterpretation and rotation logic errors, respectively. These bugs cause swapped sigma parameters in Gaussian blur and wrong rotated tensor values for rectangular inputs, leading to unexpected image processing results.
• issues/127247, issues/127248

• Incorrect loss and metric calculations: The tf.keras.metrics.categorical_crossentropy function miscalculates class counts when label smoothing is applied with the class axis set to zero, resulting in inaccurate loss values. This affects model evaluation and training metrics.
• issues/127249

• Request for fused linear cross-entropy API: There is a feature request to add a fused linear cross-entropy API in tf.nn to optimize large language model training by combining linear projection and cross-entropy loss into a single GPU pass. This would reduce memory usage and prevent out-of-memory errors caused by large intermediate tensors.
• issues/127290

• Numerical discrepancies between CPU and GPU matmul: A significant numerical difference exists between CPU and GPU outputs for float32 matrix multiplication using tf.linalg.matmul in TensorFlow 2.20.0, with discrepancies exceeding typical tolerances despite identical inputs. This inconsistency challenges reproducibility across hardware.
• issues/127346

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: 33

Summarized Issues:

• TensorFlow Operation Bugs and Incorrect Behavior: Multiple TensorFlow operations exhibit incorrect behavior or bugs, such as Round not supporting complex inputs despite documentation claims, tf.raw_ops.Round outputting zeros for integer inputs, and tf.linalg.diag aborting on invalid band specifications instead of raising proper errors. These issues cause unexpected crashes, silent failures, or incorrect outputs, highlighting inconsistencies in error handling and operation implementations.
• [issues/65317, issues/74789, issues/110796]

• Segmentation Faults and Crashes in TensorFlow Functions: Several TensorFlow functions cause segmentation faults or crashes under specific conditions, including tf.compat.v1.train.shuffle_batch crashing with zero-sized dimensions, server creation crashing with excessive thread counts, and diagonal matrix creation aborting due to shape checks. These faults often stem from improper input validation or unchecked internal assumptions, leading to process aborts rather than manageable exceptions.
• [issues/105297, issues/108489, issues/69471]

• Gradient and Automatic Differentiation Errors: TensorFlow's gradient computations have multiple precision and correctness issues, such as zero gradients returned for tf.math.tanh and tf.math.sqrt at finite inputs, and incorrect second derivatives for functions like tf.nn.silu and tf.nn.softplus. These bugs cause loss of gradient signals or incorrect gradient sums, affecting model training and backpropagation accuracy.
• [issues/89110, issues/124836, issues/126524, issues/126525, issues/126630, issues/126632, issues/126635, issues/126637]

• XLA/JIT Compilation and GPU Compatibility Issues: TensorFlow's XLA/JIT compilation fails or aborts in various scenarios, including missing XLA OpKernel support for complex data types in tf.math.cumsum and tf.math.cumprod, fatal aborts after failed sparse operation compilations, and incompatibility with NVIDIA Blackwell GPUs due to PTX version mismatches. These issues cause runtime errors and prevent successful GPU execution or JIT compilation.
• [issues/111958, issues/117546, issues/120244]

• TensorFlow Lite and Model Conversion Problems: Converting models to TensorFlow Lite or running TFLite tutorials encounter issues such as non-runnable mnist_tflite.py due to deprecated APIs, indefinite hangs in TFLite Micro for stacked LSTM models with long sequences, and discrepancies in GPU memory reporting during benchmarking. These problems hinder deployment and accurate resource monitoring on edge devices.
• [issues/119439, issues/126061, issues/127163]

• TensorFlow Sparse and Tensor Handling Inconsistencies: TensorFlow exhibits inconsistent behavior in sparse tensor handling and tensor stacking, where tf.SparseTensor silently casts int32 indices to int64 under graph tracing but errors in eager mode, and tf.stack behaves differently under eager execution versus jit_compile due to input type expectations. These inconsistencies complicate debugging and code portability.
• [issues/125374, issues/125537]

• TensorFlow Slice and Indexing Bugs: The tf.slice operation behaves inconsistently with out-of-bounds size arguments, silently succeeding under some tracing modes but raising errors in eager and jit_compile modes, indicating issues with constant folding rather than XLA compilation. This leads to unpredictable runtime behavior depending on execution context.
• [issues/125544]

• TensorFlow Kernel and API Support Regression: Starting with TensorFlow 2.20.0, the removal of experimental kernel C API symbols breaks training support for PluggableDevices across all platforms, preventing access to resource variable operations. This regression impacts extensibility and device compatibility for custom hardware accelerators.
• [issues/126374]

• TensorFlow Internal Validation and Overflow Bugs: Internal validation bugs such as unchecked output feature dimension validation in quantized convolution parameters and integer overflow in TensorFlow Lite Detection PostProcess cause potential crashes or heap buffer overflows. These low-level bugs pose security and stability risks.
• [issues/111638, issues/127080]

• TensorFlow Padding and Memory Corruption: The tf.pad function with mode='reflect' silently corrupts output for large inputs due to 32-bit indexing overflow, causing incorrect results without error notification. This silent corruption undermines data integrity in large-scale tensor operations.
• [issues/91027]

• CUDA and GPU Execution Errors: CUDA-related errors such as cudaErrorInvalidConfiguration occur during specific model executions on H100 GPUs with certain layer configurations and memory growth settings, causing runtime failures on Ubuntu with Python 3.12. These errors indicate configuration or compatibility issues in GPU execution environments.
• [issues/87455]

• TensorFlow State and Constraint Bugs: Bugs in state updates and constraints include unexpected non-zero values after sliding-window state updates with tf.gather_nd and discrepancies in tf.keras.constraints.NonNeg() outputs compared to manual computations, leading to subtle correctness issues in model state and parameter constraints.
• [issues/127002, issues/127004]

• TensorFlow Build and Integration Support: Documentation and scripts for building TensorFlow Lite C library XCFrameworks for iOS and Mac Catalyst architectures facilitate integration into Xcode projects, supporting multiple architectures and CocoaPods integration for easier deployment.
• [issues/126507]

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: 19

Key Open Pull Requests

1. Preserve sigmoid gradients for large positive inputs: This pull request addresses the issue of sigmoid gradient computation for large positive inputs by replacing the existing y * (1 - y) formula with y * sigmoid(-x) to preserve nonzero gradients in floating-point precision, updates the gradient tape to retain inputs for accurate backpropagation, corrects the exclusion table size after rebasing onto master, and includes extensive regression tests and validation to ensure correctness across multiple data types and scenarios.

• URL: pull/127010

• Associated Commits: 93fa7, a0b11, c4e2e

2. Fix L2 normalization fusion for multi-axis reductions: This pull request fixes the L2 normalization fusion logic by preventing the fusion of reductions over multiple axes, ensuring that only single-axis reductions along the trailing dimension are fused, thereby preserving correct normalization behavior and including a regression test to verify the preservation of multi-axis normalization patterns.

• URL: pull/127026

• Associated Commits: 77ca4, b9558

3. Propagate onednn async build in Tensorflow and compatibility fixes for newer XLA: This pull request propagates the integration of the oneDNN asynchronous build for AArch64 in TensorFlow and includes compatibility updates to support newer XLA APIs, contingent on updating TensorFlow to a more recent third-party XLA version.

• URL: pull/127109

• Associated Commits: 87372, 0fb99

Other Open Pull Requests

• GPU kernel overflow and out-of-bounds fixes: Multiple pull requests address overflow and out-of-bounds memory issues in various GPU kernels by changing index calculations from 32-bit to 64-bit integers and adding proper bounds checks. These fixes prevent invalid memory reads and writes in ResourceScatterUpdate, ResizeBilinear, MatrixDiagPart, TopK, and ScatterNd operations, improving stability and correctness on large tensors.
• pull/127113, pull/127116, pull/126974, pull/127118, pull/127165

• Input validation and error handling improvements in numpy and matrix operations: Several pull requests add or improve input validation for diagonal offsets, axis bounds, and input sizes in numpy-related functions and matrix operations. These changes prevent fatal backend errors and improve error messages by raising appropriate exceptions for invalid inputs in tf.experimental.numpy.diag, compress, cumsum, cumprod, and oneDNN conv backprop-input kernels.
• pull/127030, pull/127293, pull/127296, pull/127297

• Memory safety and buffer handling fixes: Fixes include addressing multiple memory safety vulnerabilities in the TFLite GPU delegate model builder by correcting buffer size calculations and validating tensor pointers, as well as modifying the slim model's weights file handling to use unbuffered streams and improve error reporting. These changes prevent out-of-bounds accesses and improve robustness in file operations.
• pull/127253, pull/127292

• Gradient computation corrections: Pull requests fix gradient computations by correcting masks and alignment parameters to ensure mathematically correct derivatives and prevent incorrect gradient flows. These fixes apply to the xdivy function and the _MatrixDiagV2Grad function, aligning implementations with expected behavior and adding regression tests.
• pull/127294, pull/127316

• Code quality and logging improvements: One pull request fixes invalid Python 3 syntax in docstring examples and improves logging efficiency by switching to lazy logging arguments without changing behavior. Another replaces internal asserts with proper exceptions to improve error handling consistency.
• pull/127295, pull/127296

• XLA Pad kernel integer handling fixes: A pull request fixes integer narrowing and overflow issues in the TensorFlow XLA Pad kernel by properly handling int64 paddings and adding checks to prevent padded shapes from exceeding int64 limits. This prevents crashes and ensures errors are raised correctly when padding sizes overflow.
• pull/127324

• YNNPACK broadcast reduction fix: One pull request fixes a partial-output cropping issue in YNNPACK broadcast reductions by ensuring all reduction bounds are processed, preventing incorrect and nondeterministic sums under XLA CPU. The fix includes regression tests, dependency patching, and validation with native YNNPACK kernels.
• pull/127196

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: 59

Key Closed Pull Requests

1. Fix sqrt gradient for float64 subnormals: This pull request fixes the gradient computation of the square root function for positive float64 subnormal inputs on CPUs that flush these subnormals to zero by detecting them via their IEEE-754 bit representation and applying a stable calculation method that reconstructs the subnormal division from the mantissa, ensuring accurate higher-order and mixed gradients across dense, RaggedTensor, SparseTensor, and internal calls while preserving existing behavior for other data types.

• URL: pull/125178

• Associated Commits: ebf27, 84b31, 3c254, e78f7, ce6e7, a46a1, fe40a, 8e398, e60fa, 31c25, 0d483, 51cae, a49fc, b8d83, 033d9, 60a89

2. Fix tf.linalg.expm float16 support by computing in float32: This pull request addresses the runtime failure of tf.linalg.expm with float16 inputs by modifying the computation to occur in float32 before casting the result back to float16, thereby preserving the documented API while avoiding unsupported CPU/GPU kernels, and includes added regression tests to cover this float16 path.

• URL: pull/122575

• Associated Commits: 89059, 03950, 72287, a8f30, 4334b, 23a50, 67664, 1dd86

3. Validate values sequence input in tf.stack (fix #125374): This pull request adds explicit validation to the tf.stack function to ensure that its values parameter is a sequence of tensors, thereby raising consistent TypeError exceptions across eager, graph, and XLA execution modes when a single tensor is incorrectly passed instead of a sequence.

• URL: pull/125386

• Associated Commits: f2bfc, 04aa9, 5d248, b8a07, a7572, 59608, f04d3, 9c62a

Other Closed Pull Requests

• Validation and Error Handling Improvements: Multiple pull requests enhance validation and error handling across TensorFlow operations to prevent crashes and incorrect results. These include adding upper bound validation for thread counts, weight sum validation in np.average, rank validation in fake quant ops, axis validation in tf.experimental.numpy functions, and negative bin checks in Bincount, all accompanied by comprehensive tests to ensure robustness.
• pull/124253, pull/126325, pull/118136, pull/126204, pull/126603

• Crash and Segmentation Fault Fixes: Several pull requests address crashes and segmentation faults by adding early returns, input guards, and out-of-bounds checks. Fixes include handling empty tensors in scatter_nd, guarding profiler functions against null pointers, preventing out-of-bounds permutation indices in SparseMatrixSparseCholesky, and fixing floating point exceptions in oneDNN 3D maxpooling.
• pull/126770, pull/122959, pull/124996, pull/99921, pull/124227

• XLA and Backend Compilation Fixes: Pull requests improve XLA and backend compilation by enforcing minimum PTX versions for Blackwell GPUs, restoring missing kernel registrations, and fixing JIT compilation crashes. These changes ensure compatibility with newer CUDA versions and restore critical OpKernel registrations for successful graph compilation.
• pull/126241, pull/112039, pull/126377, pull/126935, pull/126936

• Indexing and Overflow Bug Fixes: Fixes address integer overflow and indexing bugs in tensor operations to prevent silent corruption and incorrect results. Notably, MirrorPad and MirrorPadGrad kernels were updated to use 64-bit indexing for large tensors, and range kernel arithmetic was hardened to prevent overflow and invalid output sizes.
• pull/125321, pull/125144, [pull/122548](https://github.com/pull/122548]

• Refactoring and Code Quality Enhancements: Some pull requests focus on refactoring and improving code clarity without changing behavior. These include modularizing the configure.py script, adding type hints and comments to the DLPack test suite, and improving documentation and error messages in random number generator implementations.
• pull/125792, pull/125795, pull/124227

• Build and Linking Fixes: Updates to build configurations and linking ensure proper dependency resolution and symbol exports. This includes fixing protobuf detection in CMake for benchmarks, restoring experimental kernel API symbols, and correcting Windows linker exports for TFLite functions.
• pull/109375, pull/126377, pull/127012

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