Weekly GitHub Report for Tensorflow - 2024-10-14 12:00:50
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.
I. Issues
1.1 Open Issues
Open Issues This Week: 22
Summarized Issues:
- Loss Function Issues: TensorFlow versions 2.16.2 and 2.17.0 have a bug where using a dictionary at the output of a pipeline and model causes the loss function to return None. This results in a "no loss function" error during model training. The issue disrupts the training process and needs a fix.
- Optimizer and Strategy Issues: The
MirroredStrategyin TensorFlow 2.16.2 gets stuck at the optimizer level when callingapply_gradientswith bothtf.keras.optimizers.legacy.Adamandtf.keras.optimizers.Adam. This issue prevents the optimizer from functioning correctly. It needs to be addressed to ensure smooth training.
- Keras Functional API Issues: A bug in the
fit()function of the Keras Functional API with atf.data.Datasetfor validation data results in anAttributeError. This error is due to aNoneTypeobject having no attributeitems. It disrupts the model training process.
- Bitwise Operation Discrepancies: The
tf.bitwise.left_shiftfunction in TensorFlow version 2.18.0-dev20240728 shows different behaviors between CPU and GPU executions. The results differ when shifting negative values. This discrepancy needs to be resolved for consistent results.
- Sparse Tensor and CTC Loss Issues: Using a sparse tensor to calculate CTC loss in TensorFlow2 results in faster training and reduced GPU memory usage. However, it causes the training network to deteriorate and produce NaN values when the batch size is increased beyond 1. This issue needs to be addressed for stable training.
- ImportError Issues: An ImportError occurs when attempting to run TensorFlow, related to a failure in loading the
_pywrap_tensorflow_internalDLL. This issue prevents TensorFlow from running correctly and needs a fix.
- Gradient Calculation Issues: In TensorFlow version 2.12.1, the
math_ops.sparse_segment_meanandmath_ops.sparse_segment_sumfunctions return random gradient values instead of zero when provided with empty tensors. This bug affects the accuracy of gradient calculations.
- Shape Mismatch Errors: A shape mismatch error occurs when running a TensorFlow model on version 2.17.0, which does not happen on version 2.11.0. This discrepancy needs to be resolved to ensure consistent model performance across versions.
- Memory Mapping Errors: A ValueError occurs when initializing the MoveNet model in TensorFlow 2.15.0 on Google Colab due to a memory mapping error. The error message indicates "Mmap of '41' at offset '0' failed with error '22'". This issue needs to be fixed for successful model initialization.
- Unsupported GPU Operations: TensorFlow's
tf.gatherandtf.nn.embedding_lookupfunctions do not support usingtf.stringtype indices on the GPU. This limitation should be documented to inform users about the unsupported operations.
- TensorFlow Lite Crashes: A user experiences a crash after upgrading TensorFlow Lite from version 2.14.0 to 2.16.1. They request access to the unstripped shared object (.so) files to locate the crash stack using the symbol table. This issue needs to be addressed to prevent crashes.
- TensorFlow Lite Library Size: A user seeks advice on reducing the size of a TensorFlow Lite shared library (.so file) built for a BeagleBone Black device running Ubuntu 22.04.4 LTS. The current 4 MB file size is too large for their requirements. This issue needs a solution for size optimization.
- Compilation Errors with GPU Support: A user encounters a compilation error while manually compiling TensorFlow with GPU support. The error is related to the inability to find a registered compiler for the CUDA platform, raising concerns about operations potentially falling back to CPU execution.
- Custom LayerCAM Issues: A
ValueErroroccurs when attempting to create a submodel using a custom LayerCAM saliency function in TensorFlow 2.17. The error message indicates that the layer 'sequential' has never been called and thus has no defined output. This issue needs to be resolved for successful submodel creation.
- Support for 16KB Page Sizes: A request for support for 16KB page sizes in TensorFlow Lite on x86, x86_64, and armeabi-v7a architectures is made to improve performance on Android 15. This support is needed to leverage the new page sizes.
- TimeDistributed Layer Issues: A
ValueErroroccurs because theTimeDistributedlayer in TensorFlow is not being initialized with atf.keras.layers.Layerinstance but instead with aKerasTensorobject. This issue needs to be fixed for proper layer initialization.
- SavedModel Export Issues: A
FAILED_PRECONDITIONerror occurs when running a SavedModel exported from atf.Modulethat contains a Keras model as an instance variable. This issue does not occur in TensorFlow v2.15.0 but does in v2.17.0, indicating a regression that needs to be fixed.
- RNN Quantization Issues: Users are unable to quantize a Recurrent Neural Network (RNN) with customized cells using TensorFlow. This results in kernel crashes and segmentation faults during the conversion process, which needs to be addressed for successful quantization.
- Low-Level TPU Interaction: A user seeks detailed information and resources on low-level TPU interaction and the libtpu.so API. They aim to build an automatic differentiation library without relying on high-level frameworks like TensorFlow, JAX, or PyTorch-XLA.
- OpenSSL Substitution: A proposal is made to conditionally substitute boringSSL with the system-installed OpenSSL for building TensorFlow on IBM Power architecture. This is due to the lack of support for Power in the latest versions of boringSSL.
- SSE4.1 Instruction Issues: TensorFlow version 2.9.3 was compiled to use SSE4.1 instructions that are not available on the user's machine. This results in an "Aborted (core dumped)" error when attempting to use the library on Ubuntu 22.04.1 LTS, which needs to be resolved for compatibility.
1.2 Top 5 Active Issues:
We consider active issues to be issues that have generated much discussion in the issue's comments.
-
cuDNN, cuFFT, and cuBLAS Errors: This issue is about errors encountered when running TensorFlow on a GPU, specifically related to the inability to load cuDNN, cuFFT, and cuBLAS libraries, which results in slower training times for RNNs and FFNNs. The user is running TensorFlow on WSL2 with Ubuntu 22, CUDA version 11.8, and cuDNN version 8.7, and despite the GPU being recognized, the performance is not optimal due to these errors.
- The comments discuss various attempts to resolve the issue, including reinstalling TensorFlow with the
tensorflow[and-cuda]package, downgrading TensorFlow to version 2.13, and trying different CUDA and cuDNN versions. Some users report partial success, while others continue to face the same errors. The issue persists across different setups, including native Ubuntu installations and WSL2, and there is a consensus that the errors do not prevent GPU usage but may affect performance. The discussion also touches on potential fixes and workarounds, such as using nightly builds of TensorFlow or adjusting NUMA settings, but a definitive solution is still pending. - Number of comments: 130
- The comments discuss various attempts to resolve the issue, including reinstalling TensorFlow with the
-
how to download or install .so file for tflite conversion with gpu delegate: This issue is about a user seeking guidance on how to download or install the
.sofile required for TensorFlow Lite (TFLite) conversion with a GPU delegate. The user has provided detailed system information, code snippets, and specific problems encountered during the conversion process, including errors and requests for further assistance.- The comments section includes detailed troubleshooting steps, such as installing specific versions of Bazel, configuring Android NDK and SDK, and addressing errors related to missing files and compatibility issues. Despite multiple attempts and suggestions, the user continues to face challenges in successfully building and using the GPU delegate, leading to ongoing discussions and requests for further help.
- Number of comments: 55
-
cuBLAS Error in 2.14.0: This issue reports a bug in TensorFlow version 2.14.0, where users encounter errors related to the registration of cuDNN, cuFFT, and cuBLAS factories when importing TensorFlow on various platforms, including Ubuntu and Windows with WSL2. The errors suggest that the factories for these CUDA plugins are being registered multiple times, leading to warnings and potential performance issues.
- The comments section reveals that multiple users are experiencing the same issue across different operating systems and configurations, with some suggesting downgrading TensorFlow to version 2.9 or 2.13 as a temporary workaround, while others report that the errors do not significantly impact performance.
- Number of comments: 53
-
TensorFlow Profiler did not work correctly.: This issue involves the TensorFlow Profiler not functioning correctly when measuring the performance of large-scale deep learning models on CPU/GPU, as the results are not displayed in TensorBoard. The user suspects that the current version of the TensorFlow Profiler does not save all necessary files for measurement results.
- Multiple users reported similar issues with the TensorFlow Profiler, with some suggesting potential fixes like manually copying specific files, while others noted that the problem persisted across different TensorFlow versions. The discussion also included requests for directory structures and code examples to debug the issue, and suggestions to ensure compatibility between TensorFlow, TensorBoard, and the profiler plugin versions.
- Number of comments: 44
-
TF-TRT Warning: Could not find TensorRT: This issue is about a user experiencing difficulties with TensorFlow not being able to find TensorRT on their Ubuntu 22.04 system, despite having installed the necessary drivers and CUDA version. The user has tried various troubleshooting steps, including uninstalling and reinstalling TensorFlow, but continues to encounter the error, which is hindering their progress in a Machine Learning course.
- Multiple users report similar issues with TensorRT not being found, even after installing it via different methods. Suggestions include ensuring TensorRT is installed, setting the correct environment variables, and creating symbolic links for the TensorRT libraries. Some users found success by following specific installation guides or adjusting their Docker configurations, while others continue to face challenges.
- Number of comments: 41
1.3 Top 5 Quiet Issues:
We consider quiet issues to be issues that have been opened in this project for the longest time. The team should work together to get these issues resolved and closed as soon as possible.
-
coreml deleagate not support ResizeBilinear layer: This issue pertains to the lack of support for the ResizeBilinear layer with specific parameters (align_corners == false, half_pixel_center == true) in the CoreML delegate when running a TensorFlow Lite model on an iPhone 12. The user notes that while the ResizeBilinear layer with different parameters (align_corners == false, half_pixel_center == false) is supported by the ANE engine, the current configuration leads to performance issues due to the unsupported version.
- Open for 363 days, 13 hours, 01 minutes
-
Bazel@HEAD is breaking TensorFlow: This issue describes a problem where the latest version of Bazel is causing build failures in TensorFlow, specifically during the linking process of certain TensorFlow Lite components. The error logs indicate that there are undefined symbols related to the
tsl::Flagandtsl::Flagsclasses, which are causing the build to fail.- Open for 363 days, 07 hours, 10 minutes
-
Tensorflowlite flex delegate not loaded automatically when interpreter created even though tensorflowlite_flex.dll provided. (C++ on windows): This issue pertains to the TensorFlow Lite Flex delegate not being automatically loaded when an interpreter is created in a C++ application on Windows, despite the presence of the necessary
tensorflowlite_flex.dllfile. The problem arises because anullptris being passed to thegetProcAddress()function instead of a handle totensorflowlite_flex.dll, resulting in an error message indicating that the Flex delegate is not applied before inference.- Open for 362 days, 16 hours, 18 minutes
-
quantized range of fake_quant_with_min_max_args is -2num_bits + 1 to 2 num_bits.: This issue pertains to a bug in the
fake_quant_with_min_max_argsfunction in TensorFlow, where the quantized range is incorrectly calculated as-2**num_bits + 1to2 ** num_bits, making it impossible to represent the quantized weights usingnum_bitsintegers. The user has implemented this function in numpy and is seeking guidance on how to represent the weights using a signed type.- Open for 358 days, 07 hours, 12 minutes
-
Segmentation fault when running tensorflow.python.framework.kernels.get_registered_kernels_for_op: This issue reports a segmentation fault occurring when the function
tensorflow.python.framework.kernels.get_registered_kernels_for_opis called with aNoneargument in TensorFlow version 2.11.0. The problem is identified as a bug and has been open for 357 days, with the user providing a code snippet to reproduce the error and relevant log output indicating a TensorRT library loading warning followed by the segmentation fault.- Open for 357 days, 22 hours, 35 minutes
1.4 Closed Issues
Closed Issues This Week: 24
Average Issue Close Time (This Week): 30.21 days
Summarized Issues:
- Discrepancies in Model Outputs: This issue describes a discrepancy in the outputs produced by running the same quantized TensorFlow model in two different environments—TensorFlow and TFLiteMicro—on the same test input. The user is seeking the source code for the fully_connected kernel in TensorFlow to compare it with the TFLiteMicro implementation. The goal is to identify the cause of the difference.
- Performance Issues with TensorFlow Lite: This issue pertains to the poor performance of an int8 quantized MobileNetV3 model using TensorFlow Lite. The user is inquiring about alternative quantization methods and the potential future support for logarithmic scale quantization. The performance issues are significant enough to warrant exploration of different quantization techniques.
- Compilation and Build Errors: Multiple issues describe various compilation and build errors encountered across different platforms and configurations. These include errors in highwayhash/sip_hash.cc on Windows, CMake errors on Ubuntu, missing 'construct_at' members in the Eigen library, and build errors on CentOS 9 Stream.
- GPU Recognition and Utilization Issues: Several issues involve TensorFlow not recognizing or utilizing GPUs correctly. These include TensorFlow Lite GPU delegate and NNAPI failing on a Vivo x100 device, TensorFlow v2.12.0 not recognizing the GPU on Ubuntu 22.04, and TensorFlow failing to load GPU libraries in a Jupyter Notebook accessed through a remote VSCode connection.
- TensorFlow Lite Model Issues: This issue involves encountering an error when running TensorFlow Lite models on an ESP32C3. The error includes failing to get registration from the op code CUSTOM and AllocateTensors() failing, with detailed descriptions of the models, conversion code, and troubleshooting steps provided.
- Installation and Module Errors: Various issues describe problems encountered during the installation and module import processes. These include SSL certificate verification failures on a Mac, "Module not found" errors on Windows 10, and TensorFlow not recognizing the Keras module in Visual Studio Code.
- Security Vulnerability Visibility: This issue concerns the lack of publication and visibility of CVE-2023-33976, a security vulnerability affecting TensorFlow versions below 2.13.0. The vulnerability was not initially reported on the MITRE website or the TensorFlow security advisories, leading to it being unnoticed by most security tools.
- TensorFlow Profiler Issues: This issue pertains to the TensorFlow Profiler not functioning correctly on Colab with various TensorFlow and Python versions. The lack of responses to the problem has been a significant concern for users relying on the profiler for performance analysis.
- Compatibility with New Numpy Version: This issue pertains to a request for upgrading the TensorFlow library to support the newly released Numpy version 2.x. The current TensorFlow version is incompatible with the new Numpy version, causing issues for users who have updated their Numpy library.
- Import Errors in Keras Module: This issue describes import errors encountered when using certain functions in the
python.kerasmodule of TensorFlow version 2.14 and above. The errors are due to the removal of the__version__attribute from the__init__.pyfile, affecting functions likehdf5_format.save_weights_to_hdf_5_group().
1.5 Issue Discussion Insights
This section will analyze the tone and sentiment of discussions within this project's open issues within the past week 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 issues from the past week.
II. Pull Requests
2.1 Open Pull Requests
Open Pull Requests This Week: 3
Pull Requests:
- Example Usage Documentation: This topic covers the addition of example usage documentation for specific functions in the TensorFlow project. The pull request focuses on the functions get_lib(), get_build_info(), and get_include() within the systemconfig.py file. This documentation aims to help users understand how to utilize these functions effectively.
- CUDA and cuDNN Support: This topic involves adding configurations to TensorFlow’s WORKSPACE file to include CUDA and cuDNN support. Additionally, a new script is introduced to verify CUDA and nvcc installation. This addresses the issue of TensorFlow failing to find the CUDA compiler, ensuring smoother setup for users.
- Data Type Support in tf.range: This topic focuses on enhancing the TensorFlow library by adding support for float16 and bfloat16 data types to the
tf.rangefunction. The pull request addresses an issue tracked in the TensorFlow repository. This enhancement allows for more flexible data type usage within thetf.rangefunction.
2.2 Closed Pull Requests
Closed Pull Requests This Week: 4
Summarized Pull Requests:
- Typographical errors and rephrasing in documentation: This pull request addresses several typographical errors and rephrases sentences in the "on-device-training overview.ipynb" file. The changes aim to improve the readability and accuracy of the documentation. The provided GitHub link details the specific modifications made.
- Quantized Argmin operation for TensorFlow Optimizing Converter (toco): This pull request aims to enable the quantized Argmin operation for the TensorFlow Optimizing Converter (toco). The addition of this operation is intended to enhance the functionality and performance of the converter. The title and description of the pull request provide further details on the implementation.
- Testing ROCm Software compatibility: This pull request involves testing the compatibility and functionality of ROCm Software version 6.1.0 with tensorflow-rocm version 2.14.0.600. The tests are designed to ensure that the new ROCm version works seamlessly with TensorFlow. The results and findings from these tests are documented in the pull request.
- Fixing missing symbols in TensorFlow Lite C API on Windows: This pull request addresses the issue of missing symbols in the TensorFlow Lite C API on Windows. By ensuring that the preprocessor define
TFL_STATIC_LIBRARY_BUILDis undefined for the C API target, the pull request allows symbols to be exported correctly. This fix is crucial for the proper functioning of the TensorFlow Lite C API on Windows platforms.
2.3 Pull Request Discussion Insights
This section will analyze the tone and sentiment of discussions within this project's open pull requests within the past week 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 pull requests from the past week.
III. Commits
3.1 Commits
Commits This Week: 289
Summarized Commits:
- Automated Code Changes: Numerous commits involve automated code changes, indicating routine updates and maintenance tasks. These changes are often associated with PiperOrigin-RevId references and do not introduce new functionality but ensure the codebase remains up-to-date and consistent.
- LLVM Integration: Several commits integrate the LLVM project at specific commits, updating the usage to align with changes made in those revisions. This ensures the project remains compatible with the latest LLVM updates and improvements.
- Forward Compatibility and GraphDef Updates: Multiple commits update the forward compatibility horizon and GraphDef versions, ensuring the project remains compatible with future changes and maintains up-to-date graph definitions.
- Error Handling and Logging Improvements: Commits focus on reducing error logging, avoiding unnecessary logs during successful operations, and enhancing error messages for better debugging. These changes improve the clarity and efficiency of error handling in the project.
- Coordination Service Enhancements: Changes include reducing error logging in the coordination service and avoiding logging during successful shutdowns, which streamline the coordination service's operation and reduce noise in logs.
- LLVM and StableHLO Integration: Commits integrate the LLVM project and StableHLO components at specific revisions, ensuring the project aligns with the latest updates and improvements from these dependencies.
- Protocol Buffer and Dependency Updates: Updates to operations-related protocol buffer text files and the removal of unnecessary dependencies indicate ongoing maintenance to keep the project clean and efficient.
- Performance Optimizations: Various commits introduce performance optimizations, such as removing redundant checks, optimizing execution paths, and improving memory handling. These changes result in significant performance improvements across different benchmarks.
- New Features and Enhancements: New features include the introduction of the
original_valuefield in HLO graphs, support for new operations and conversions, and enhancements to existing functionalities like themapfunction andThunkExecutor.
- Test and Build Improvements: Commits introduce new tests, update existing ones, and address build issues to ensure the project remains robust and reliable. These changes include enabling tests for specific components and addressing test timeout issues.
- Code Cleanup and Refactoring: Numerous commits focus on code cleanup, refactoring, and removing obsolete or unused code. These changes improve the maintainability and readability of the codebase.
- Bug Fixes and Issue Resolutions: Several commits address and resolve bugs, such as pipelining issues, segmentation faults, and typographical errors. These fixes ensure the project remains stable and functional.
- Thread and Memory Management: Changes include introducing upper limits on intra-operation threads, enabling NUMA-aware memory allocation, and optimizing memory handling for better performance and efficiency.
- Sharding and Partitioning Enhancements: Commits enhance sharding and partitioning functionalities, such as propagating sharding information, optimizing partitioning logic, and introducing new sharding strategies.
- GPU and CPU Backend Improvements: Various commits focus on optimizing the GPU and CPU backends, including introducing new kernels, optimizing execution paths, and addressing specific backend issues.
- Documentation and Comment Updates: Updates to documentation and comments improve the clarity and understanding of the codebase, ensuring developers have accurate and helpful information.
- Compatibility and API Updates: Changes include updating the code to align with new API standards, such as replacing deprecated NumPy APIs and ensuring compatibility with new versions of dependencies.
- New Utility Functions and Attributes: Commits introduce new utility functions and attributes, such as the
IndexingMapAttrand functions for sorting JSON strings, enhancing the project's functionality and flexibility.
- Pipeline and Fusion Optimizations: Optimizations to pipeline tests, fusion logic, and execution paths result in improved performance and efficiency, particularly for GPU operations.
- Configuration and Build System Updates: Internal configuration changes and updates to the build system ensure the project remains easy to build and configure, addressing issues like build failures and dependency management.
- Error Detection and Verification Enhancements: Enhancements to error detection and verification processes, such as the ifrt-verify-donation pass, improve the project's robustness and reliability.
- New Tests and Benchmarks: The introduction of new tests and benchmarks ensures the project's performance and functionality are thoroughly evaluated, providing insights into areas for further improvement.
- Removal of Obsolete Features: The removal of obsolete features and dependencies, such as the Tesla patch and
device_tracerfunctionality, keeps the project lean and focused on current requirements.
- Cross-Platform Support: Updates to support different platforms, such as enabling Gloo support for MacOS and addressing Windows compilation issues, ensure the project remains accessible and functional across various environments.
- Optimization of Specific Operations: Commits focus on optimizing specific operations, such as the
Thunk::OkExecuteEventfunction and theKernelThunkcomponent, resulting in performance improvements for targeted scenarios.
- Enhanced Debugging and Traceability: The introduction of verbose logging and utility functions for debugging, such as generating compact 2D maps of occupied heap memory, enhances the project's traceability and ease of debugging.
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, or created at least 1 pull request in the past month.
| Contributor | Commits | Pull Requests | Issues |
|---|---|---|---|
| TensorFlower Gardener | 1235 | 0 | 0 |
| GitHub | 18 | 0 | 0 |
| tensorflow-jenkins | 0 | 8 | 0 |
| 372046933 | 0 | 1 | 2 |
| guillaume-tgl | 0 | 0 | 3 |
| eyalhir74 | 0 | 0 | 3 |
| libofei2004 | 0 | 0 | 3 |
| Ilia Sergachev | 2 | 0 | 0 |
| lhutton1 | 0 | 2 | 0 |
| SsomsakTH | 0 | 1 | 1 |
| abhinav-mcw | 0 | 2 | 0 |
| hguandl | 0 | 0 | 2 |
| t-kalinowski | 0 | 0 | 2 |
| arthurflor23 | 0 | 0 | 2 |
| dryglicki | 0 | 0 | 2 |
| siy415 | 0 | 0 | 2 |
| KnightGOKU | 0 | 0 | 2 |
| A-fliga | 0 | 0 | 2 |
| 6eanut | 0 | 0 | 2 |
| pradeep10kumar | 0 | 0 | 2 |
| mayurmunshi | 0 | 0 | 2 |
| Brett Taylor | 1 | 0 | 0 |
| Pol Dellaiera | 1 | 0 | 0 |
| drupol | 0 | 1 | 0 |
| vimalrajv | 0 | 1 | 0 |
| akhilgoe | 0 | 1 | 0 |
| RahulSundarMCW | 0 | 1 | 0 |
| alifbasha22 | 0 | 1 | 0 |
| tilakrayal | 0 | 1 | 0 |
| sergachev | 0 | 1 | 0 |
| gzmkl | 0 | 1 | 0 |
| yashpratap914 | 0 | 1 | 0 |
| quicbrtal | 0 | 1 | 0 |
| shadchin | 0 | 1 | 0 |
| Theorems-lab | 0 | 1 | 0 |
| eddie-santos | 0 | 1 | 0 |
| abhinavph21 | 0 | 1 | 0 |
| mshr-h | 0 | 1 | 0 |
| humayunh-amd | 0 | 1 | 0 |
| LakshmiKalaKadali | 0 | 1 | 0 |
| nehalmr | 0 | 1 | 0 |
| roll-123 | 0 | 1 | 0 |
| yayale1 | 0 | 0 | 1 |
| MohannadAbuIssa | 0 | 0 | 1 |
| Yulee3542 | 0 | 0 | 1 |
| a-sajjad72 | 0 | 0 | 1 |
| KeondoPark | 0 | 0 | 1 |
| coco-boy | 0 | 0 | 1 |
| dauso | 0 | 0 | 1 |
| BGigotSDS | 0 | 0 | 1 |
| risheek-mittal | 0 | 0 | 1 |
| hanan454 | 0 | 0 | 1 |
| wscJayasooriya | 0 | 0 | 1 |
| maximd1 | 0 | 0 | 1 |
| dhruv2103 | 0 | 0 | 1 |
| devapriyas2001 | 0 | 0 | 1 |
| nvn234 | 0 | 0 | 1 |
| AmarOk1412 | 0 | 0 | 1 |
| benjaminreynoso | 0 | 0 | 1 |
| Leli1024 | 0 | 0 | 1 |
| zhanghuicuc | 0 | 0 | 1 |
| prasad-mali07 | 0 | 0 | 1 |
| pulkitagarawal | 0 | 0 | 1 |
| Irayanbu05 | 0 | 0 | 1 |
| jiannanWang | 0 | 0 | 1 |
| Morehman27 | 0 | 0 | 1 |
| andy-tai | 0 | 0 | 1 |
| ewwll | 0 | 0 | 1 |
| Yongle-Fu | 0 | 0 | 1 |
| BREBION-Mathis | 0 | 0 | 1 |
| therooler | 0 | 0 | 1 |
| Ligeirinho00 | 0 | 0 | 1 |
| 333Random333 | 0 | 0 | 1 |
| SExpert12 | 0 | 0 | 1 |
| deepeshfujitsu | 0 | 0 | 1 |
| Tarun0000 | 0 | 0 | 1 |
| HashemZn-04 | 0 | 0 | 1 |
| Nayana-ibm | 0 | 0 | 1 |
| bernsny24 | 0 | 0 | 1 |
| tranvantungit | 0 | 0 | 1 |
| berndporr | 0 | 0 | 1 |
| Vinaygoudasp7 | 0 | 0 | 1 |
| jaskarannagi19 | 0 | 0 | 1 |
| VicB18 | 0 | 0 | 1 |
| 3cktorcrypto | 0 | 0 | 1 |
| Made-Jaya | 0 | 0 | 1 |
| Deepu777yt | 0 | 0 | 1 |
| poltomo | 0 | 0 | 1 |
| OlgasAcc | 0 | 0 | 1 |
| Allan1974 | 0 | 0 | 1 |
| arteen1000 | 0 | 0 | 1 |
| wonyoungmin | 0 | 0 | 1 |
| gajendrahatt | 0 | 0 | 1 |
| shkarupa-alex | 0 | 0 | 1 |
| JVD9kh96 | 0 | 0 | 1 |
| noahewolfe | 0 | 0 | 1 |
| justinvyu | 0 | 0 | 1 |
| mseDPYU4 | 0 | 0 | 1 |
| Jacob-yen | 0 | 0 | 1 |
| Andrew-XQY | 0 | 0 | 1 |
| SCH227 | 0 | 0 | 1 |
| Apprisco | 0 | 0 | 1 |
| Ayman250 | 0 | 0 | 1 |
| azhar47-sk | 0 | 0 | 1 |
| ThomasRichtsfeld | 0 | 0 | 1 |
| jmmelen | 0 | 0 | 1 |
| Rosie-93 | 0 | 0 | 1 |
| MikeyD-rbg | 0 | 0 | 1 |
| tom2002965 | 0 | 0 | 1 |
| intelav | 0 | 0 | 1 |
| pmdaye | 0 | 0 | 1 |
| grep-mb | 0 | 0 | 1 |
| just-sabyr | 0 | 0 | 1 |
| wangzhen0518 | 0 | 0 | 1 |
| yjiangling | 0 | 0 | 1 |
| LuvolwethuTokwe | 0 | 0 | 1 |
| tanwang2020 | 0 | 0 | 1 |
| ConnorLynchLV8 | 0 | 0 | 1 |
| AndreiMoraru123 | 0 | 0 | 1 |
| nbNobody | 0 | 0 | 1 |
| saad-koukous | 0 | 0 | 1 |
| bash-s | 0 | 0 | 1 |
| Senantq | 0 | 0 | 1 |
| juvelop17 | 0 | 0 | 1 |
| vixhead | 0 | 0 | 1 |
| ivansoban | 0 | 0 | 1 |
| Voivio | 0 | 0 | 1 |
| ghsanti | 0 | 0 | 1 |
| notlober | 0 | 0 | 1 |
| npanpaliya | 0 | 0 | 1 |
| bhavesh-hirpara | 0 | 0 | 1 |