Weights & Biases
Data & MLThe AI developer platform. Use Weights & Biases to train and fine-tune models, and manage models from experimentation to production.
Release History
v0.25.0Breaking3 fixes3 featuresThis release drops support for Python 3.8 and introduces new features for the W&B LEET TUI, including multi-run workspace and config editing, alongside several critical bug fixes.
v0.24.21 fix1 featureThis release introduces support for Federated Auth (JWT) in wandb.Api() and fixes an issue with refreshing expired artifact download URLs.
v0.24.14 fixes1 featureThis release fixes several regressions introduced in version 0.24.0, particularly around data synchronization and login host configuration, and adds a new method for downloading run history exports.
v0.24.0Breaking6 fixes2 featuresThis release removes the deprecated `wandb.beta.workflows` module, introduces signal forwarding for agents, and includes several fixes related to artifact ID generation, API calls, and run logging presentation.
v0.23.1Breaking9 fixes6 featuresThis release adds regex filtering and chart inspection to the LEET TUI, expands the automations API, introduces a beta_history_scan method, deprecates anonymous mode, and includes several performance and bug fixes.
v0.23.05 fixes3 featuresThis release adds an experimental wandb beta leet command, expands the Registry API with member management and a read‑only id property, and includes several bug fixes around artifact handling, upload progress, and VSCode notebook support.
v0.22.37 fixes2 featuresThis release adds configurable console log chunking settings and per_page support for Registry API queries, expands API key length support, and includes numerous bug fixes improving stability, artifact handling, and Jupyter notebook path inference.
v0.22.22 fixesThis release includes bug fixes for missing `output.log` uploads and for excessive data uploads when `run.save()` is called repeatedly on unchanged files.
v0.22.11 fix3 featuresThis release adds checksum caching for artifact downloads, lazy loading for run listings, and a storage_region option for artifacts, while fixing timeout handling in artifact existence checks.
v0.22.0Breaking3 fixesThis release removes macOS 10 support and includes several bug fixes, notably improved run‑resumption handling, API compatibility with older servers, and deadlock prevention in sweep command scripts.
v0.21.41 featureIntroduces DSPy integration for tracking metrics, logging predictions, and saving DSPy programs as W&B artifacts.
v0.21.31 fixThis release updates the click dependency to >=8.0.1 and fixes an extraneous environment-variable warning message.
v0.21.2Breaking2 fixes2 featuresThis update introduces new run-metric limits and a Weave integration, makes linking artifacts raise errors instead of returning `None` (a breaking change), and fixes config dataclass serialization and artifact-link performance.
v0.21.1Breaking8 fixes7 featuresThis release adds new summary options, paginated sweeps, pattern filtering for file listings, and faster large‑artifact uploads, while deprecating the beta workflows module and removing `__bool__` from registry iterators.
v0.21.0Breaking6 fixes3 featuresThis release removes the legacy service implementation and related settings, adds new process‑tracking and custom Vega chart‑preset features, deprecates the `.length` attribute, and includes several bug fixes.
v0.20.11 fixFixed a regression in wandb.Image() handling of NumPy arrays with values in [0,1] introduced in version 0.20.0.
v0.20.0Breaking7 fixes8 featuresThis release adds new Table log modes, expands Image support, introduces several new parameters, and removes deprecated service‑disable and docker features, with corresponding breaking changes and bug fixes.
v0.19.116 fixes11 featuresThis release adds extensive SDK registry management, new artifact linking features, multipart downloads, project ID access, a full W&B Automations API, and Nvidia GPU profiling, while fixing several bugs related to logging, configuration paths, artifact naming, pydantic compatibility, initialization permissions, and file deletion.
v0.19.103 fixes4 featuresThis release introduces new init behavior, artifact history step access, HTML handling improvements, deprecates several Run methods, removes a quiet flag from the IPython magic, and includes several bug fixes.
v0.19.9Breaking5 fixes2 featuresThis release adds a default reinit option and a new artifact URL scheme, deprecates boolean reinit values, introduces several breaking changes (logger propagation, Metaflow dependency, stricter per_page handling, updated artifact API), and includes multiple bug fixes.
v0.19.81 fix1 featureThis release fixes file upload issues with special-character paths and increases the system monitor sampling interval to 15 seconds.
v0.19.73 fixes2 featuresThis release introduces a Registry search API and improves wandb.init logging, while fixing several bugs and updating the moviepy dependency constraint.
v0.19.64 fixes1 featureThis release adds Prometheus API support for Nvidia DCGM GPU metrics and updates ECC counter handling, while fixing several artifact‑related performance and correctness bugs.
v0.19.53 fixes1 featureThis release adds a new `wandb login --base-url` alias, fixes offline mode network requests, expands artifact name sanitization, and improves run rewind handling, while temporarily disabling per‑core CPU stats.
v0.19.42 fixesThis release improves TPU monitoring by fixing device count and duty cycle reporting and adding support for non-consecutive TPU device IDs.
v0.19.31 fixThis release fixes a regression affecting wandb.Settings updates within the wandb.integration.metaflow module.
Common Errors
ModuleNotFoundError2 reportsThe "ModuleNotFoundError" in weights-biases (wandb) usually indicates that a required wandb dependency or a related package like gymnasium is not installed in your Python environment. To fix this, identify the missing module from the error message (e.g., 'wandb.errors', 'gymnasium') and install it using pip: `pip install wandb` or `pip install gymnasium` or `pip install` for the specific module shown in the error. Ensure you're installing into the correct Python environment for your project.
FileNotFoundError1 reportFileNotFoundError in Weights & Biases often occurs when syncing offline runs because artifact files created using `Artifact.new_file` or similar methods are not saved to disk before the syncing process begins. Ensure that you explicitly call `artifact.save()` after creating or modifying the artifact and before starting the `wandb sync` process to persist the files and resolve the error. Additionally, verify the file paths if the error persists even after saving.
AttributeError1 reportThe "AttributeError" in weights-biases often arises when outdated or conflicting versions of wandb or its dependencies are installed, particularly in environments like Kaggle notebooks. To fix it, try upgrading wandb to the latest version using `pip install wandb --upgrade` and restarting your kernel or runtime; if the the issue persists, consider specifying the version with `pip install wandb=={known_good_version}` or reinstalling the wandb package and its dependencies to ensure a clean installation and resolve potential conflicts.
CommError1 reportCommError in wandb often arises from network issues or race conditions in the wandb backend during initialization or finalization. Ensure a stable internet connection and consider adding explicit synchronization mechanisms (like locks) around crucial wandb operations such as `wandb.init()` and `wandb.finish()` if running in a multithreaded/multiprocess environment. Check for firewall restrictions or proxy settings potentially interfering with wandb's communication.
TypeError1 reportTypeError in weights-biases often arises from attempting to apply string operations or access methods meant for strings on non-string data types within a summary object. To fix this, ensure that all data being processed as strings are definitively strings by using `str()` to explicitly convert them before applying string operations or accessing string-specific methods. This will prevent unexpected type mismatches during summary processing.
LookupError1 reportThe "LookupError: ContextVar 'shell_parent'" in Weights & Biases (wandb) often arises in interactive environments like Jupyter notebooks when wandb's internal process tries to access the shell's parent process information that isn't readily available. To resolve this, explicitly initialize wandb within a defined context; a common fix is to wrap your `wandb.init()` call within an `if __name__ == '__main__':` block, ensuring it's only executed when the script is run directly (not when imported). Alternatively, try setting the environment variable `WANDB_NOTEBOOK_NAME` to bypass the shell parent lookup, for example, `os.environ["WANDB_NOTEBOOK_NAME"] = "my_notebook.ipynb"`.
Related Data & ML Packages
An Open Source Machine Learning Framework for Everyone
🤗 Transformers: the model-definition framework for state-of-the-art machine learning models in text, vision, audio, and multimodal models, for both inference and training.
Tensors and Dynamic neural networks in Python with strong GPU acceleration
scikit-learn: machine learning in Python
Flexible and powerful data analysis / manipulation library for Python, providing labeled data structures similar to R data.frame objects, statistical functions, and much more
Streamlit — A faster way to build and share data apps.
Subscribe to Updates
Get notified when new versions are released