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 better API compatibility. Key updates also include runtime support for the bfloat16 data type in the tfl.Cast operation, alongside the discontinuation of standalone libtensorflow package publishing.

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:BUILD/INSTALL] [SUBTYPE: UBUNTU/LINUX] [TF 2.18] [TF-2.18] Protoc-related Segmentation Fault on GH200 when Building from Source : This issue describes a segmentation fault occurring during the Bazel build of TensorFlow 2.18 from source on a Grace Hopper GH200 system, specifically related to the protoc compiler. The user suspects the fault arises from incorrect instruction sets used for building protoc on the aarch64 architecture, as there is no prebuilt pip package for GH200 and the build fails despite following recommended LLVM and Bazel versions.

   • The comments include a request for clarification on whether the build was attempted on CPU or GPU, the user confirming CUDA support was enabled and the build is on ARM CPU, and a suggestion from the TensorFlow team to try the latest TensorFlow version and hardware configuration since the issue may be outdated; the discussion ends with acknowledgments but no resolution.
   • Number of comments this week: 4

2. Update the pinned ruy dependency — current pin (Sep 2022) predates upstream fixes for ruy::CpuInfo::Initialize() SIGSEGV on Android: This issue requests updating the pinned ruy dependency in TensorFlow Lite because the current pin from September 2022 predates important upstream fixes that prevent a null-pointer dereference crash in ruy::CpuInfo::Initialize() on certain Android devices. The problem affects many users in production, and the suggested fix is to bump the pinned commit to a more recent version that includes these critical null-guard improvements, which have already been merged upstream and verified to resolve the crash without additional code changes.

   • The comments include a polite request for feedback on whether the proposed change correctly addresses the issue, a suggestion to create a minimal reproducible example and bisect the problem, and a response acknowledging the existing stack trace and upstream fixes while planning to prepare a minimal test to validate the fix.
   • Number of comments this week: 3

3. [STAT:CONTRIBUTION WELCOME] [TYPE:BUILD/INSTALL] [SUBTYPE:WINDOWS] [2.18.RC] Failed to build tensorflow_cc in Windows when linking: This issue describes a failure to build the tensorflow_cc C++ library on Windows when linking, where many essential symbols such as Session and SavedModelBundleInterface are missing despite successful compilation. The problem is attributed to incomplete dependency declarations in Bazel build files and missing symbol exports, with ongoing efforts including manual dependency additions, Windows-specific build configuration fixes, and a forked repository providing patched Windows-compatible builds.

   • The comments reveal users confirming the issue and sharing detailed analysis of missing dependencies and build configuration problems; maintainers acknowledge the complexity and limited official support for Windows builds of tensorflow_cc, suggesting workarounds like using Linux VMs or WSL; a contributor has created a fork with patched builds that partially resolve linking errors, and users provide feedback on these releases while awaiting further improvements.
   • Number of comments this week: 2

4. [STAT:AWAITING TENSORFLOWER] [TYPE:BUILD/INSTALL] [SUBTYPE: UBUNTU/LINUX] [COMP:CORE] ../tensorflow/third_party/xla/third_party/tsl/tsl/platform/ml_dtypes.h:19:10: error: 'ml_dtypes/include/float8.h' file not found [clang-diagnostic-error]: This issue reports a build error caused by a missing header file 'ml_dtypes/include/float8.h' when compiling TensorFlow code on an Arm64 platform using clang++, specifically within the third-party XLA and TSL components. The user is attempting to build TensorFlow without Bazel on a smartphone environment (Termux on AOSP) and faces challenges due to missing dependencies, submodules, and the complexity of TensorFlow’s build system, seeking a minimal setup to verify code validity without full Bazel-based compilation.

   • The comments discuss attempts to reproduce the error on different platforms, clarify that the missing file is part of a submodule not included by default, explore partial workarounds like cloning the ml_dtypes repository manually, and highlight the difficulty of building TensorFlow without Bazel, especially on resource-constrained devices; suggestions include using submodules properly and installing prebuilt packages, but no definitive solution is reached.
   • Number of comments this week: 2

5. [STAT:AWAITING RESPONSE] [TYPE:BUG] [COMP:OPS] [TF 2.18] Crash in tensoflow protobuf.: This issue reports a crash occurring in TensorFlow when running on CPUs with AVX512 flags, caused by an invalid argument exception thrown within the protobuf library during the MklLayoutRewritePass optimization. The user experiences this problem on TensorFlow version 2.18.1 with a specific hardware setup and seeks a workaround or fix to prevent the crash related to oneDNN optimizations.

   • The comments confirm the crash is linked to oneDNN optimizations interacting poorly with protobuf on AVX512 CPUs, suggesting disabling oneDNN via an environment variable or build flag as a workaround. The discussion also advises building TensorFlow with --define=build_with_mkl=false to disable oneDNN for production and recommends testing on the latest TensorFlow release, as newer oneDNN versions may have resolved the issue.
   • 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: 10

Summarized Issues:

• XLA Compilation Failures and Errors: Multiple issues report failures and errors when compiling TensorFlow functions with XLA under jit_compile=True. These include problems with tf.norm over multiple axes, tf.squeeze(tf.where(...)) with bounded dynamic dimensions, repeated tf.while_loop calls causing missing body functions, and incorrect loop folding in the simplify-while-loops optimization pass, all leading to compilation errors or incorrect results.
• issues/122865, issues/122885, issues/122886, issues/123067

• Numerical and Determinism Issues with XLA: There are severe numerical inconsistencies and non-deterministic outputs when using XLA compilation with certain models, such as a HyperNetwork producing zero-dimension outputs. Additionally, tf.sigmoid on bfloat16 inputs on CPU shows non-monotonic and incorrectly rounded results in eager and graph modes, while XLA-compiled versions produce correct rounding, causing output inconsistencies.
• issues/122952, issues/123104

• Memory and Crash Bugs in TensorFlow Operations: Some TensorFlow operations cause critical memory issues and crashes, including a heap corruption bug in eager tf.matmul on CPU with specific tensor shapes causing out-of-bounds writes, and SavedModels containing debug raw operations that crash the Python process with segmentation faults during inference.
• issues/123109, issues/123114

• Dataset Caching and Pipeline Ordering Warning: A warning occurs during training when using tf.data.Dataset.cache() with a file path, indicating that the dataset iterator did not fully read the cached dataset mid-epoch. This partial reading may cause cache discarding and suggests reordering the pipeline to use take() before cache() to avoid unexpected truncation.
• issues/122752

• RangeOp Memory Allocation Bug: Calling tf.range with extremely large floating-point arguments causes the RangeOp to infer an excessively large output size, leading to an unbounded memory allocation attempt. This results in a process crash instead of raising a recoverable exception.
• issues/122746

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: 60

Summarized Issues:

• Build and Compilation Issues: Multiple issues report build failures and compilation errors related to TensorFlow's source code and dependencies. These include challenges with clang on Windows, missing headers for TensorRT 10, Bazel and Clang 19 errors on Ubuntu 24.10, and undeclared identifiers in libtensorflow builds, all causing build interruptions or requiring configuration updates.
  • issues/14425, issues/63952, issues/66215, issues/66473, issues/68360, issues/90459

• Segmentation Faults and Crashes in TensorFlow Ops: Several TensorFlow operations cause segmentation faults or crashes due to invalid inputs or missing validation. Examples include crashes in AudioSpectrogram with negative parameters, Conv3DBackpropInputV2 with malformed input sizes, MaxPool with oneDNN optimizations, QuantizeAndDequantizeV3 with negative axis, and HistogramFixedWidth with infinite values. These faults often result from missing error handling and lead to process aborts instead of user-friendly exceptions.
  • issues/63352, issues/94114, issues/94115, issues/94116, issues/94122, issues/94127, issues/108664, issues/122748

• GPU and CUDA Related Failures: Multiple issues describe GPU-related errors including invalid argument errors during multi-GPU training, CUDA plugin duplicate registration warnings, GPU crashes due to illegal memory access, and GPU device busy errors causing crashes. These problems affect training stability, GPU utilization, and compatibility with CUDA versions, often resulting in aborted processes or degraded performance.
  • issues/53140, issues/65988, issues/76456, issues/78784, issues/85604, issues/113073, issues/113147, issues/113149, issues/87366

• Memory Leaks and Performance Regressions: There are reports of memory leaks during training and compilation, as well as performance regressions in TensorFlow Lite and TensorFlow training pipelines. Memory leaks occur in TensorFlow Probability variable compilation and during GPU training with CUDA 12.0, while performance drops are noted when switching data feeding methods and in TFLite Dense layer latency.
  • issues/73457, issues/87414, issues/69702, issues/109156

• TensorFlow Lite Specific Issues: Several issues concern TensorFlow Lite including missing or incorrect framework files on iOS, incomplete access to classes on iOS, thread utilization problems on Raspberry Pi5, and bugs in int8 quantization causing crashes or accuracy loss. These problems affect model conversion, deployment, and inference performance on various platforms.
  • issues/63952, issues/110160, issues/112769, issues/117052, issues/118185

• Logging and Import Errors: Issues include duplicated log messages caused by TensorFlow adding logging handlers without checking existing ones, and import failures due to conflicts with dependencies like rules_python, resulting in attribute errors and failed resource loading. These affect usability and debugging experience.
  • issues/75288, issues/78190

• TensorFlow API and Operation Bugs: Various bugs affect TensorFlow operations and APIs such as incorrect output shapes after freezing models, AutoGraph failures with Python match statements, incorrect byte order handling in tensor proto creation, and invalid shape errors in loss functions. These issues cause unexpected behavior, crashes, or incorrect results during model building and execution.
  • issues/43839, issues/55789, issues/60458, issues/89245

• Documentation and Type Checking Problems: There are stale or inaccurate documentation references, such as outdated Java/JNI packaging info, and false-positive static type errors reported by Pylance/Pyright due to TensorFlow's Python type definitions. These issues hinder developer understanding and tooling support.
  • issues/119584, issues/105851

• Model Saving and Weight Initialization Issues: Problems include unexpectedly large file sizes when saving model weights compared to saving the full model without optimizer, and inconsistent Keras lazy weight initialization producing different initial weights across execution modes despite fixed seeds. These affect reproducibility and storage efficiency.
  • issues/91498, issues/120627

• Multi-Worker and Distributed Training Bugs: A bug in MultiWorkerMirroredStrategy with RING communication and Adam optimizer causes divergence in loss and gradient norms between compiled and reference setups due to incorrect hardcoded batch size during XLA tracing, leading to inconsistent training results across workers.
  • issues/120986

• TensorFlow Lite Model Loading and Verification: Loading malformed TensorFlow Lite models without schema verification causes misaligned memory accesses and crashes on strict-alignment platforms, indicating a need for improved validation to prevent undefined behavior and SIGBUS errors.
  • issues/121217

• Integration of Third-Party Libraries: Integration of the KDNN library into TensorFlow 2.15.0 replaces the sigmoid operator with a custom operator to improve execution performance, with plans to extend this approach to other operators, demonstrating efforts to enhance TensorFlow's computational efficiency.
  • issues/118587

• TensorFlow Operation Fatal Errors on Malformed Inputs: Several TensorFlow operations such as ResourceApplyGradientDescent and SparseSegmentSumWithNumSegments cause fatal internal errors or GPU crashes on malformed inputs instead of raising recoverable exceptions, leading to process aborts and reduced robustness.
  • issues/113073, issues/113147, issues/113149

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: 30

Key Open Pull Requests

1. Fix XLA DenseBincount negative input validation for runtime tensors: This pull request fixes the XLA DenseBincount kernel to properly handle negative inputs in runtime (non-constant) tensors by masking out negative indices and preventing incorrect results during JIT compilation, thereby extending the previous compile-time constant validation to cover the common runtime use case without overwriting valid bins.

• URL: pull/122706

• Associated Commits: 579a3, 15f5a, ae044

2. Fix RangeOp abort on extreme float ranges before allocation: This pull request fixes an issue where calling tf.range with extreme floating-point values could cause the RangeOp to infer an excessively large output size that fits in int64 but leads to multi-terabyte memory allocations and process aborts under AddressSanitizer by adding pre-allocation checks that reject range outputs of 1 TiB or more with an InvalidArgument error, and mirrors these checks in shape inference and related operations, accompanied by new C++ and Python tests to prevent such excessive allocations.

• URL: pull/122811

• Associated Commits: cb7e6, 3f7f0, 3fbfe

3. Add descriptive message about your cFix: Resolve TypeError during import from sysconfig.get_path returning Nonehanges: This pull request fixes critical TypeErrors and AttributeErrors during import by adding checks for None values returned from sysconfig.get_path, safely handling path concatenation for PEP 420 namespace packages, and ensuring all can be modified whether defined as a list or tuple.

• URL: pull/122954

• Associated Commits: bd0cc, 22f02, b037c

Other Open Pull Requests

• XLA compiler bug fixes and improvements: Multiple pull requests address various bugs and compilation issues in the XLA compiler, including fixing memory corruption during zero-dimension StridedSlice operations, handling dynamic dimensions in tf.squeeze, resolving function lookup failures in repeated tf.while_loop bodies, and correcting integer trip count calculations in simplify-while-loops. These fixes improve the stability and correctness of XLA compilation and execution for TensorFlow operations involving dynamic shapes and loops.
  • pull/122980, pull/123015, pull/123016, pull/123069

• Validation enhancements for parameters in loss functions and activations: Several pull requests add or improve validation checks for parameters such as label_smoothing in binary_crossentropy and categorical_crossentropy, the alpha parameter in tf.nn.leaky_relu and ELU layers, and the penalty number in regularizers. These changes include range checks, tensor_util guards, and runtime assertions to ensure parameter values are within valid bounds, enhancing robustness and preventing invalid inputs.
  • pull/122716, pull/122717, pull/122771, pull/122772, pull/122773, pull/122714, pull/122715

• Parameter validation for model components and callbacks: Pull requests add validation to ensure that parameters such as groups in convolutional layers and patience in the EarlyStopping callback are greater than zero or non-negative. These checks prevent invalid configurations and improve error handling in model building and training workflows.
  • pull/122712, pull/122713

• Segmentation fault and crash fixes in core TensorFlow ops: Fixes include preventing segmentation faults in scatter_nd operations by adding early return checks for empty updates, adding shape-consistency checks to prevent out-of-bounds reads when restoring partitioned tensors, and limiting attribute values in tf.raw_ops.AsString to avoid crashes. These changes improve stability and correctness in tensor operations and checkpoint handling.
  • pull/122959, pull/123020, pull/123102

• Fixes for JIT/XLA compilation and gradient computation issues: One pull request fixes a JIT/XLA compilation failure in tf.sequence_mask by ensuring integer dtypes in internal operations, while another fixes crashes in gradient computations of tf.linalg.det and tf.linalg.slogdet on singular matrices by using SVD-based methods. These fixes ensure correct compilation and stable gradient calculations in edge cases.
  • pull/123112, pull/122823

• Grappler optimization correctness improvements: A pull request fixes incorrect results from Grappler's ArithmeticOptimizer by removing saturating or non-injective float32 ops from the list of monotonic functions safe for ArgMax/ArgMin optimization, preventing invalid transformations and ensuring correctness.
  • pull/122826

• Cache warning message clarification: The cache truncation warning message was updated to explicitly mention the effect of using batch(drop_remainder=True) on input pipelines, clarifying that elements may be dropped downstream causing the cache to never reach EOF and be deleted.
  • pull/122904

• Fix for oneDNN AvgPooling3DGrad core dump: A core dump issue in the AvgPooling3DGrad operation within TensorFlow's oneDNN integration was fixed, and a new CPU-only test was added to verify the fix, improving reliability of pooling gradient computations.
  • pull/122995

• Fix for tf.experimental.numpy.flip axis parameter handling: The tf.experimental.numpy.flip function was fixed to correctly support tuple and list inputs for the axis parameter by normalizing axes and using a helper function, ensuring behavior consistent with NumPy and adding regression tests.
  • pull/123027

• Validation in RaggedCross operation: A validation check was added to ensure the sparse input's indices matrix has the same number of rows as the values vector, preventing out-of-bounds reads when building row splits, similar to existing SparseCross checks.
  • pull/123107

• Security improvement by pinning GitHub Actions: Two GitHub Actions used in write-scoped workflows were pinned from mutable tags to specific full commit SHAs to prevent security risks from tag repointing, ensuring immutable references without changing behavior or versions.
  • pull/123116

• New fuzzing target for TFLite model loading and interpreter: A new fuzzing target was added to exercise the TFLite model loading and interpreter invocation path, aiming to detect potential memory-safety issues in previously unfuzzed code within OSS-Fuzz.
  • pull/123123

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: 24

Key Closed Pull Requests

1. Fix compute_average_loss scaling divergence during XLA tracing by leveraging replica context: This pull request fixes a critical bug causing optimization divergence during XLA tracing in distributed multi-worker training by modifying compute_average_loss to reliably determine the global batch size scaling factor through a new helper that prioritizes the XLA-safe per-thread replica context over the eager strategy stack, thereby preventing incorrect constant folding of the batch size and ensuring consistent gradient scaling across execution contexts.

• URL: pull/121366

• Associated Commits: e02af, c250e, 64410, ea9c9, d9bf8, e5a3a, 67dfe, f1e83, bf18f, 6001d, 3396a

• Associated Commits: e02af, c250e, 64410, ea9c9, d9bf8, e5a3a, 67dfe, f1e83, bf18f, 6001d, 3396a

2. Fix Pylance/Pyright false type errors for Tensor/Variable (#105851): This pull request fixes false type errors reported by Pylance and Pyright for Tensor and Variable classes in TensorFlow by correcting the return type annotations of tf.constant(), adding type stubs for dynamically-added operators in both Tensor and Variable classes to improve static type checking, and wrapping these stubs in TYPE_CHECKING guards to maintain dynamic behavior while providing accurate type information.

• URL: pull/110383

• Associated Commits: f1bef, 05fb6, 21c4b, 74227, 1f124, 3b5ac, cd5e6

• Associated Commits: f1bef, 05fb6, 21c4b, 74227, 1f124, 3b5ac, cd5e6

3. Fix integer overflow in pad output shape computation: This pull request fixes an integer overflow vulnerability in the pad::ResizeOutputTensor() function by changing the output dimension computation to use 64-bit integers with bounds checking against INT32_MAX, preventing heap-buffer-overflow writes caused by the unchecked sum of input size and padding values that could previously lead to memory corruption when loading malicious TensorFlow Lite models.

• URL: pull/115456

• Associated Commits: 87737, 3cecd, 75efc, 5b554

• Associated Commits: 87737, 3cecd, 75efc, 5b554

Other Closed Pull Requests

• Integer overflow and buffer overflow fixes in TensorFlow Lite kernels: Multiple pull requests address critical integer overflow vulnerabilities by adding validation to ensure dimensions and padding values are non-negative and fit within 32-bit integer bounds before copying or computing shapes. These changes prevent heap-buffer-overflow writes caused by unchecked narrowing from int64 to int32 and improve kernel robustness by early error handling.
• pull/115460, pull/115455

• Integer overflow prevention in Conv2D GEMM dimension calculations: A pull request fixes a critical integer overflow issue by promoting multiplicands to int64_t before multiplication, preventing process crashes caused by int32 overflow when input dimension products exceed INT32_MAX.
• pull/112776

• Validation improvements in TensorFlow Lite operations: Pull requests add validation to ensure input tensors have expected shapes or values, such as verifying the size tensor in resize_bilinear contains exactly two elements and that axis values in the Roll kernel are within valid ranges. These validations prevent out-of-range accesses and align behavior with similar operations.
• pull/116205, pull/121292

• Memory management and safety improvements in TensorFlow Lite kernels: A pull request replaces manual memory management with safer C++ constructs in the CTC Beam Search Decoder, LSH Projection, and Elementwise kernels, deferring allocations until after parameter validation and adding type tracking to ensure correct deallocation. This improves resource lifecycle safety without affecting public APIs or performance.
• pull/119695

• Python code style and exception handling improvements: Pull requests replace generic Exception raises with more specific built-in exceptions like ValueError and RuntimeError, and apply Pythonic best practices such as replacing bare except: with except Exception: and using idiomatic expressions to improve code readability and adherence to standards without changing behavior.
• pull/117895, pull/117894

• Documentation updates and clarifications: Several pull requests update docstrings and documentation to clarify behavior, correct inaccuracies, or remove outdated content, including clarifications on tf.random.set_seed, tf.boolean_mask, tf.signal.idct, sumSquare function examples, and removal of obsolete Java/JNI README sections.
• pull/122101, pull/117948, pull/121215, pull/121716, pull/123068

• Bug fixes preventing crashes and improving error handling: Pull requests add validation checks to prevent fatal crashes caused by invalid input shapes or values, such as in MklAvgPoolingGradOp::Compute, AvgPool3DGrad kernel, and AudioSpectrogramOp constructor, replacing crashes with structured error handling and proper exceptions.
• pull/120177, pull/109326, pull/122635, pull/119354

• Kernel fixes for numerical stability and correctness: Pull requests fix the softmax kernel by zeroing out invalid padded channels and the TFLite tf.nn.sigmoid operation by clamping inputs to prevent NaN outputs caused by numerical instability and integer overflow during range reduction.
• pull/108244, pull/119354

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.