Flower 1.25.0

Thanks to our contributors

We would like to give our special thanks to all the contributors who made the new version of Flower possible (in git shortlog order):

Chong Shen Ng, Daniel J. Beutel, Heng Pan, Javier, Mohammad Naseri, Soumik Sarker, William Lindskog, Yan Gao, sarahhfalco

What's new?

• Track compute time and network traffic for runs (#6241, #6242, #6243, #6244, #6245, #6249, #6266, #6267, #6268, #6269, #6270, #6271, #6272, #6273, #6274, #6275, #6276, #6279)

  Flower now records compute time and network traffic for a run. The run detail view shown by flwr list --run-id <run-id> displays traffic exchanged between SuperLink and SuperNode, as well as compute time used by ServerApp and ClientApp.

• Refactor flwr new to pull apps from the Flower platform (#6251, #6252, #6258, #6259, #6260, #6263, #6265, #6283, #6284, #6285, #6292, #6294, #6302)

  Refactors flwr new to fetch Flower apps directly from the Flower platform (see the usage reference). This introduces new and updated quickstart examples (including NumPy and FlowerTune LLM), renames and updates existing examples, aligns CI to run against platform-backed examples, and updates related documentation and benchmark instructions.

• Migrate examples to the Message API and remove outdated Docker Compose as well as Tensorflow Privacy examples (#6232, #6233, #6238, #6297, #6304)

  Migrates the scikit-learn, Vertical FL, and Opacus examples to the Message API, with the Vertical FL example also updated to use flwr-datasets. The outdated Docker Compose and Tensorflow Privacy examples are removed.

• Improve CLI output with human-readable durations (#6277, #6296)

  Updates the Flower CLI to display durations in a more human-friendly format (xd xh xm xs), automatically selecting appropriate units instead of the previous HH:MM:SS format.

• Update examples and baselines (#6234, #6256, #6257, #6264, #6280, #6281, #6286, #6287, #6288, #6290, #6291, #6293)

• Improve documentation (#6229, #6230, #6255, #6262)

• Update CI/CD configuration (#6168, #6246, #6295, #6305)

• General improvements (#6056, #6085, #6176, #6235, #6236, #6254, #6278, #6299)

  As always, many parts of the Flower framework and quality infrastructure were improved and updated.

Incompatible changes

• Remove bundled templates from flwr new (#6261)

  Removes the templates previously bundled with the Flower wheel now that flwr new pulls apps from the Flower platform. The --framework and --username options are deprecated as part of this change.

Flower 1.24.0

Thanks to our contributors

We would like to give our special thanks to all the contributors who made the new version of Flower possible (in git shortlog order):

Charles Beauville, Chong Shen Ng, Daniel J. Beutel, Daniel Nata Nugraha, Heng Pan, Javier, Patrick Foley, Robert Steiner, Yan Gao

What's new?

• Add Python 3.13 support (#6116, #6119, #6132)

  flwr now supports Python 3.13, with CI and dataset tests updated accordingly and ray bumped to ensure compatibility.

• Improve the flwr list view (#6117)

  flwr list now shows fewer details by default, while providing expanded information when using the --run-id flag.

• Extend federation commands and internal APIs (#6067, #6068, #6078, #6082, #6086, #6087, #6088, #6090, #6091, #6092, #6093, #6094, #6098, #6103, #6105, #6121, #6122, #6143, #6152, #6153, #6154, #6158, #6165, #6167)

  These updates refine and extend the existing federation concept. flwr federation show and flwr federation list now provide clearer visibility into SuperNodes and runs in the default federation.

• Introduce unified app heartbeat mechanism (#6204, #6206, #6209, #6210, #6212, #6213, #6214, #6215, #6218, #6219, #6221, #6224, #6225, #6226, #6227)

  Introduces a unified heartbeat mechanism for app processes, preventing hangs when an app process crashes without responding. The new system enables flwr-serverapp and flwr-simulation processes to exit more quickly when a run is stopped by the flwr stop command.

• Fix bugs (#6188, #6171, #6175, #6207)

  Resolves issues causing occasional missing or unregistered object IDs on lower-powered machines, prevent the flwr-serverapp process from hanging after being stopped via the CLI, and correct the finished_at timestamp and initial heartbeat interval for runs.

• Improve import performance (#6102)

• Improve internal SQLite database performance (#6178, #6195)

• Update CI workflows and development tooling (#5242, #6053, #6080, #6089, #6108, #6129, #6130, #6131, #6135, #6138, #6142, #6144, #6156, #6181, #6187, #6189)

• Update documentation (#6115, #6081, #6110, #6137, #6146, #6169, #6179, #6228)

• General improvements (#6077, #6083, #6084, #6095, #6097, #6100, #6101, #6109, #6114, #6123, #6127, #6139, #6140, #6150, #6151, #6157, #6159, #6160, #6162, #6164, #6172, #6173, #6174, #6180, #6190, #6191, #6192, #6193, #6196, #6197, #6198, #6199, #6200, #6202, #6203, #6205, #6208, #6211, #6222, #6223)

  As always, many parts of the Flower framework and quality infrastructure were improved and updated.

Incompatible changes

• Drop Python 3.9 support (#6118, #6136, #6147)

  flwr now requires Python 3.10 as the minimum supported version, with baselines and development scripts updated accordingly.

• Bump protobuf to 5.x (#6104)

  Upgrades protobuf to >=5.29.0, ensuring flwr uses the latest gRPC stack and remains compatible with TensorFlow 2.20+. Note that this version is incompatible with TensorFlow versions earlier than 2.18.

• Deprecate flwr.server.utils.tensorboard (#6113)

  The flwr.server.utils.tensorboard function is now deprecated, and a slow import issue occurring when tensorflow is installed has been resolved.

Flower 1.23.0

Thanks to our contributors

We would like to give our special thanks to all the contributors who made the new version of Flower possible (in git shortlog order):

Adam Tupper, Alan Yi, Alireza Ghasemi, Charles Beauville, Chong Shen Ng, Daniel Anoruo, Daniel J. Beutel, Daniel Nata Nugraha, Heng Pan, Javier, Patrick Foley, Robert Steiner, Rohat Bozyil, Yan Gao, combe4259, han97901, maviva

What's new?

• Enable dynamic SuperNode management via Flower CLI (#5953, #5954, #5955, #5962, #5968, #5974, #5977, #5980, #5981, #5985, #5986, #5987, #5988, #5991, #5998, #6003, #6004, #6006, #6009, #6023, #6028, #6029, #6037, #6064, #6070)

• Add migration guide for OpenFL to Flower (#5975)

  Adds a migration guide to support OpenFL users as the project approaches archival. The guide explains how to transition existing OpenFL setups to Flower, providing step-by-step migration instructions.

• Add quantum federated learning example with PennyLane (#5852)

  Introduces a new example demonstrating quantum federated learning using PennyLane. This example showcases how Flower can be integrated with quantum machine learning workflows—Flower is going quantum!

• Replace flwr ls with flwr list (keep alias for compatibility) (#5973)

  The old flwr ls command remains available as an alias.

• Migrate examples and tutorials to Message API (#5950, #5957, #5963, #5966)

  Migrates the remaining examples and tutorials, including the 30-minute Flower tutorial, whisper, quickstart-pandas, and federated-kaplan-meier-fitter, to the new Message API for improved consistency and maintainability.

• Refactor SuperNode lifecycle (#6051, #6052, #6060, #6061, #6063, #6069, #6073)

  Refactors the SuperNode lifecycle to align with the new management flow, streamlining SuperNode registration, activation, deactivation, and unregistration.

• Add deployment guide for multi-cluster OpenShift setups (#6001)

• Introduce file-based ObjectStore (#6040, #6036, #6008, #6042)

• Improve documentation (#5936, #5937, #5943, #5949, #5956, #5958, #5972, #5976, #5983, #5996, #5999, #6010, #6018, #6030)

• Update dependencies and CI (#5932, #5941, #5944, #5964, #6014, #6020, #6021, #6022, #6024, #6026, #6032, #6035, #6055, #6065)

• Bugfix (#5979)

• General improvements (#5773, #5938, #5939, #5942, #5948, #5951, #5959, #5984, #5989, #5992, #6007, #6011, #6033, #6038, #6041, #6046, #6047, #6048, #6050, #6054, #6057, #6058, #6074, #6075)

  As always, many parts of the Flower framework and quality infrastructure were improved and updated.

Incompatible changes

• Remove CSV-based SuperNode authentication (#5997)

• Rename user authentication to account authentication (#5965, #5969)

  Renames "user authentication" to "account authentication" across the framework for improved clarity and consistency. This change also updates the YAML key from auth_type to authn_type to align with authz_type.

• Deprecate --auth-supernode-public-key flag (#6002, #6076)

  The --auth-supernode-public-key flag in flower-supernode is deprecated and no longer in use. The public key is now automatically derived from the --auth-supernode-private-key, simplifying configuration and reducing redundancy.

Flower 1.22.0

Thanks to our contributors

We would like to give our special thanks to all the contributors who made the new version of Flower possible (in git shortlog order):

Charles Beauville, Chong Shen Ng, Daniel J. Beutel, Heng Pan, Javier, Patrick Foley, William Lindskog, William Lindskog-Munzing, Yan Gao

What's new?

• Migrate all strategies to Message API (#5845, #5850, #5851, #5867, #5884, #5894, #5904, #5905, #5908, #5913, #5914, #5915, #5917, #5919, #5920, #5931)

  Migrates and implements all federated learning strategies to support the new Message API. Strategies updated or introduced include FedAvg, FedOpt and its variants (FedAdam, FedYogi, FedAdagrad), FedProx, Krum, MultiKrum, FedAvgM, FedMedian, FedTrimmedAvg, QFedAvg, and Bulyan. Differential privacy strategies were also migrated, including both fixed and adaptive clipping mechanisms on the server and client sides.

• Migrate flwr new templates to Message API (#5901, #5818, #5893, #5849, #5883)

  All flwr new templates have been updated to use the Message API. The PyTorch template based on the legacy API is retained and explicitly marked as legacy for those who prefer or require the older approach. A new template for XGBoost is introduced.

• Revamp main tutorials to use the Message API (#5861)

  The primary tutorial series has been updated to showcase the Message API. The revamped content improves alignment with recent architectural changes and enhances learning clarity. See the updated tutorial: Get started with Flower.

• Upgrade tutorials and how-to guides to the Message API (#5862, #5877, #5891, #5895, #5896, #5897, #5898, #5906, #5912, #5916, #5921, #5922, #5923, #5924, #5927, #5928, #5925)

  All framework tutorials and how-to guides have been fully migrated to the Message API. This includes quickstarts for JAX, TensorFlow, PyTorch Lightning, MLX, FastAI, Hugging Face Transformers, and XGBoost, along with comprehensive updates to guides covering strategy design, differential privacy, checkpointing, client configuration, evaluation aggregation, and stateful client implementation. These changes ensure all learning resources are up-to-date, aligned with the current architecture, and ready for developers building on the Message API.

• Migrate and update examples to support the Message API (#5827, #5830, #5833, #5834, #5839, #5840, #5841, #5860, #5868, #5869, #5875, #5876, #5879, #5880, #5882, #5887, #5888, #5889, #5892, #5907, #5911, #5930)

  Migrates a wide range of examples to the new Message API, ensuring consistency with recent framework updates. Examples updated include quickstarts (e.g., TensorFlow, PyTorch Lightning, Hugging Face, MONAI, FastAI, MLX), advanced use cases (e.g., FlowerTune for ViT and LLMs, FedRAG, FL-VAE), and specialized scenarios (e.g., XGBoost, tabular data, embedded devices, authentication, and custom mods). Enhancements also include updated variable naming, model-saving logic, readme improvements, and import path corrections for better usability and alignment with the new API.

• Introduce experimental flwr pull command (#5863)

  The flwr pull Flower CLI command is the foundation for future functionality allowing for the retrieval of artifacts generated by a ServerApp in a remote SuperLink.

• Improve CI/CD workflows (#5810, #5842, #5843, #5854, #5856, #5857, #5858, #5859, #5865, #5874, #5900, #5815)

• General improvements (#5844, #5847, #5870, #5872, #5873, #5881, #5886, #5890, #5902, #5903, #5909, #5910, #5918)

  As always, many parts of the Flower framework and quality infrastructure were improved and updated.

Flower 1.21.0

Thanks to our contributors

We would like to give our special thanks to all the contributors who made the new version of Flower possible (in git shortlog order):

Charles Beauville, Chong Shen Ng, Daniel J. Beutel, Daniel Nata Nugraha, Dimitris Stripelis, Evram, Heng Pan, Javier, Robert Steiner, Yan Gao

What's new?

• Introduce Message API strategies (#5710, #5766, #5770, #5771, #5774, #5779, #5787, #5794, #5798, #5804, #5807, #5813, #5824, #5831, #5838)

  Introduces a new abstract base class Strategy that operate on Message replies, mirroring the design of those operating on FitIns/FitRes or EvaluateIns/EvaluteRes while providing more versatility on the type of payloads that can be federated with Flower. The first batch of Message-based strategies are: FedAvg, FedOpt, FedAdam, FedAdagrad, FedYogi, and fixed-clipping Differential Privacy strategies. More will follow in subsequent releases. A migration guide has been added to help users transition their existing Flower Apps operating on the original Strategy and NumPyClient abstractions to the Message API.

• Introduce Flower SuperExec (#5659, #5674, #5678, #5680, #5682, #5683, #5685, #5696, #5699, #5700, #5701, #5702, #5703, #5706, #5713, #5726, #5731, #5734, #5735, #5737, #5751, #5759, #5811, #5828)

  SuperExec is a new component responsible for scheduling, launching, and managing app processes (e.g., ServerApp, ClientApp) within the Flower deployment runtime. It is automatically spawned when running a SuperLink or SuperNode in subprocess mode (default). This also introduces a token-based mechanism that improves security by assigning a unique token to each app execution. Supporting changes include new RPCs, protocol updates, plugin abstractions, and Docker image support for SuperExec. For more details, refer to the updated Flower architecture explainer. Documentation has been revised to reflect the introduction of Flower SuperExec, including guides and tutorials such as quickstart with Docker, GCP deployment, and network communication to consistently use SuperExec.

• Update quickstart-pytorch to use Message API (#5785, #5786, #5802)

  The quickstart-pytorch tutorial has been migrated to the Message API, using the new FedAvg strategy and the new flwr new template.

• New PyTorch template with Message API (#5784)

  A new PyTorch template using the Message API is now available through flwr new.

• Add OpenShift deployment guide for Flower (#5781)

  Introduces a guide for deploying Flower on Red Hat OpenShift, including setup steps and configuration examples.

• Improve Helm documentation (#5711, #5733, #5748, #5758, #5765, #5816)

  Helm guide has been enhanced with additional configuration details and updated formatting. Changes include adding a parameters section, documenting how to set a custom secretKey, updating TLS instructions for version 1.20, introducing audit logging configuration, and using SuperExec.

• Improve documentation (#5159, #5655, #5668, #5692, #5723, #5738, #5739, #5740, #5753, #5764, #5769, #5775, #5782, #5788, #5795, #5809, #5812, #5817, #5819, #5825, #5836)

  Restructures the tutorial series, removes flower-simulation references, and updates versioned docs to use the correct flwr versions. The framework documentation homepage now defaults to the latest stable release instead of the main branch.

• Re-export user-facing API from flwr.*app (#5814, #5821, #5832, #5835)

  The following classes are now re-exported:

  • From flwr.serverapp: ServerApp, Grid
  • From flwr.clientapp: ClientApp, arrays_size_mod, message_size_mod
  • From flwr.app: Array, ArrayRecord, ConfigRecord, Context, Message, MetricRecord, RecordDict

  Importing these from flwr.server, flwr.client, or flwr.common is deprecated. Please update your imports to use flwr.serverapp, flwr.clientapp, or flwr.app instead to ensure forward compatibility.

• Add --health-server-address flag to Flower SuperLink/SuperNode/SuperExec (#5792)

• Update CI/CD workflows and dependencies (#5647, #5650, #5651, #5656, #5712, #5714, #5747, #5754, #5755, #5796, #5806, #5808, #5829)

• General improvements (#5622, #5660, #5673, #5675, #5676, #5686, #5697, #5708, #5719, #5720, #5722, #5736, #5746, #5750, #5757, #5776, #5777, #5789, #5805, #5797, #5820)

  As always, many parts of the Flower framework and quality infrastructure were improved and updated.

Incompatible changes

• Rename Exec API to Control API (#5663, #5665, #5667, [#5671](https://github.com/a...

Flower 1.20.0

Thanks to our contributors

We would like to give our special thanks to all the contributors who made the new version of Flower possible (in git shortlog order):

Charles Beauville, Chong Shen Ng, Daniel J. Beutel, Daniel Nata Nugraha, Dimitris Stripelis, Heng Pan, Javier, Kumbham Ajay Goud, Robert Steiner, William Lindskog, Yan Gao

What's new?

• Send/receive arbitrarily large models (#5552, #5550, #5600, #5611, #5614, #5551)

  Flower 1.20 can send and receive arbitrarily large models like LLMs, way beyond the 2GB limit imposed by gRPC. It does so by chunking messages sent and received. The best part? This happens automatically without the user having to do anything.

• Implement object-based messaging between SuperNode and ClientApp (#5540, #5577, #5581, #5582, #5583, #5584, #5585, #5586, #5587, #5589, #5590, #5592, #5595, #5597, #5598, #5599, #5602, #5604, #5605, #5606, #5607, #5609, #5613, #5616, #5624, #5645)

  Redesigns the messaging system to enable object-based communication between the SuperNode and ClientApp, replacing the previous message-coupled design. Introduces new RPCs and enhances the ClientAppIo and Fleet APIs to faciliate better object storage in SuperNode and decouple ObjectStore from Message, improving maintainability and extensibility. Several refactorings improve modularity, naming consistency, and model weight streaming.

• Refactor SuperNode to use NodeState exclusively (#5535, #5536, #5537, #5541, #5542, #5610, #5628)

  Refactors SuperNode to rely solely on NodeState for managing all information, decoupling internal components for improved maintainability and clearer state handling. RPCs of the ClientAppIo API have been refactored accordingly, laying the groundwork for future concurrent ClientApps support.

• Enforce maximum size limit for FAB files (#5493)

  Limits the size of FAB files to a maximum of 10MB to prevent oversized artifacts. Developers can reduce FAB size by excluding unnecessary files via the .gitignore file in the Flower app directory.

• Add CatBoost federated learning quickstart example (#5564)

  This example shows how to use CatBoost with Flower for federated binary classification on the Adult Census Income dataset. It applies a tree-based bagging aggregation method. View the example for more details.

• Fix Windows path issue in FAB builds (#5608)

  Updates the way FAB files represent relative paths to their internal files to ensure consistency across different operating systems. This fixes an issue where a FAB built on Windows would fail integrity checks when run on UNIX-based systems (e.g., Ubuntu).

• Add explainer for pyproject.toml configuration (#5636)

  Adds a guide explaining how to configure a Flower app using its pyproject.toml file. The documentation is available here.

• Improve flwr new templates with TOML comments and README links (#5635)

  Adds comments to the generated pyproject.toml and a new section in the README.md, both linking to the TOML explainer.

• Warn when running Ray backend on Windows and update simulation guide (#5579)

  Logs a warning when using the Ray backend for simulation on Windows. Updates the simulation guide to include a corresponding note about limited Windows support.

• Add Helm deployment guide (#5637)

  The documentation now includes a comprehensive guide for deploying Flower SuperLink and SuperNode using Helm charts. For full instructions, refer to the Helm Guide.

• Add docs for user authentication and audit logging (#5630, #5643, #5649)

  Introduces documentation for configuring user authentication (User Authentication Guide) and audit logging (Audit Logging Guide) in Flower.

• Support gRPC health check by default (#5591)

• Bugfixes (#5567, #5545, #5534)

• Improve CI/CD (#5560, #5544, #5531, #5532, #5547, #5578)

• Improve and update documentation (#5558, #5603, #5538, #5626, #5566, #5553, #5588, #5549, #5618, #5612, #5646)

• General improvements (#5543, #5594, #5623, #5615, #5629, #5571, #5617, #5563, #5620, #5619, #5546, #5601, #5641, #5555, #5533, #5548, #5557, #5565, #5554, #5621, #5644, #5576, #5648)

  As always, many parts of the Flower framework and quality infrastructure were improved and updated.

Incompatible changes

• Remove non-grpc-bidi transport support from deprecated start_client (#5593)

  Drops support for non-grpc-bidi transport in the deprecated start_client API. Pleaes use flower-supernode instead.

Flower 1.19.0

Thanks to our contributors

We would like to give our special thanks to all the contributors who made the new version of Flower possible (in git shortlog order):

Adam Tupper, Andrej Jovanović, Charles Beauville, Chong Shen Ng, Daniel J. Beutel, Daniel Nata Nugraha, Dimitris Stripelis, Haoran Jie, Heng Pan, Javier, Kevin Ta, Mohammad Naseri, Ragnar, Robert Steiner, William Lindskog, ashley09, dennis-grinwald, sukrucildirr

What's new?

• Revamp messaging system with content-addressable object model (#5513, #5477, #5424, #5379, #5353, #5372, #5507, #5364, #5517, #5514, #5342, #5393, #5508, #5504, #5335, #5341, #5430, #5308, #5487, #5509, #5438, #5369, #5354, #5486, #5380, #5496, #5399, #5489, #5446, #5432, #5456, #5442, #5462, #5429, #5497, #5435, #5371, #5450, #5384, #5488, #5434, #5425, #5475, #5458, #5494, #5449, #5492, #5426, #5445, #5467, #5474, #5527)

  Introduces a content-addressable messaging system that breaks messages into a tree of uniquely identified, SHA256-hashed objects. This model allows objects to be pushed and pulled efficiently, avoiding redundant uploads and enabling scalable message streaming and broadcasting. Core enhancements include the new InflatableObject abstraction and ObjectStore for storing and retrieving message content, with Array, Message, and *Record classes now inherit from InflatableObject. New utilities and RPCs facilitate recursive object handling, ID recomputation avoidance, and safe deletion. The framework's servicers, REST, and gRPC layers were refactored to integrate this system, improving real-world deployment scalability and communication efficiency.

• Improve user authorization and access control (#5428, #5505, #5506, #5422, #5510, #5421, #5420, #5448, #5447, #5503, #5501, #5502, #5511)

  Improves user authorization feature that integrates with the existing authentication protocol. When authentication is enabled, commands like flwr ls, flwr log, and flwr stop are restricted to displaying or affecting only the runs submitted by the authenticated user. This is enforced using the Flower Account ID. Additionally, fine-grained access control can be enforced for CLI operations based on assigned roles (RBAC).

• Add Floco baseline for personalized federated learning (#4941)

  Introduces Floco, a method that enhances both personalized and global model performance in non-IID cross-silo federated learning. It trains a shared solution simplex across clients, promoting collaboration among similar clients and reducing interference from dissimilar ones. Learn more in Floco Baseline Documentation.

• Add FEMNIST support to FedProx baseline (#5290)

  Adds FEMNIST dataset to FedProx with preprocessing matching the original paper—subsampling 'a'-'j' and assigning 5 classes per device. More details: FedProx Baseline Documentation

• Upgrade StatAvg baseline to new Flower App format (#4952)

  The StatAvg baseline is updated to use the new Flower App format. Changes include removing Hydra, switching to pyproject.toml configs, using ClientApp and ServerApp, and saving results via a custom Server class. More details: StatAvg Baseline Documentation.

• Add guide for running Flower on Google Cloud Platform (#5327)

  The documentation now includes a detailed guide on deploying and running Flower on Google Cloud Platform (GCP). It provides step-by-step instructions for managing Flower workloads in a GCP environment. For more information, refer to the official guide.

• Implement ServerApp heartbeat monitoring (#5228, #5370, #5358, #5332, #5322, #5324, #5230, #5325)

  Adds heartbeat support to ServerApp processes, enabling the SuperLink to detect crashed or terminated processes and mark them as finished:failed when no final status is received.

• Extend NodeState to improve SuperNode state management (#5470, #5473, #5402, #5521)

  Extends the NodeState interface and implementation to manage all SuperNode state.

• Refactor SuperNode for improved robustness and maintainability (#5398, #5397, #5443, #5410, #5411, #5469, #5419)

  Ongoing refactoring of SuperNode improves modularity, simplifies client execution, removes gRPC bidirectional streaming and unused code, and centralizes connection logic. These changes align SuperNode's behavior more closely with SuperLink to make Flower the best platform for robust enterprise deployments..

• Restructure Flower (#5465, #5476, #5460, #5409, #5408, #5396, #5389, #5392, #5461)

  Reorganizes infrastructure code into dedicated submodules to improve maintainability and clarify the separation from user-facing components.

• Improve CI/CD workflows (#5498, #5265, #5266, #5328, #5500, #5346, #5318, #5256, #5298, #5257, #5483, #5440, #5304, #5313, [#5381](https://github.com/adap/...

Flower 1.18.0

Thanks to our contributors

We would like to give our special thanks to all the contributors who made the new version of Flower possible (in git shortlog order):

Alan Silva, Andrej Jovanović, Charles Beauville, Chong Shen Ng, Chunhui XU, Daniel J. Beutel, Daniel Nata Nugraha, Dimitris Stripelis, Guanheng Liu, Gustavo Bertoli, Heng Pan, Javier, Khoa Nguyen, Mohammad Naseri, Pinji Chen, Robert Steiner, Stephane Moroso, Taner Topal, William Lindskog, Yan Gao

What's new?

• Add support for Python 3.12 (#5238)

  Python 3.12 is officially supported (Python 3.12 was in preview support since Flower 1.6). Python 3.13 support continues to be in preview until all dependencies officially support Python 3.13.

• Enable TLS connection for flwr CLI using CA certificates (#5227, #5237, #5253, #5254)

  flwr CLI now supports secure TLS connections to SuperLink instances with valid CA certificates. If no root certificates are provided, the CLI automatically uses the default CA certificates bundled with gRPC.

• Add --version and -V flags to display flwr version (#5236)

  Users can run flwr --version or flwr -V to print the current Flower version. The update also adds -h as a shorthand for CLI help.

• Use Hugging Face flwrlabs datasets in FlowerTune templates (#5205)

  FlowerTune templates switch to use datasets hosted under the flwrlabs organization on Hugging Face.

• Upgrade FedBN baseline to support flwr CLI (#5115)

  Refactors the FedBN baseline to use the new Flower CLI, removes Hydra, migrates configs, enables result saving, adds run instructions, and ensures stateful clients.

• Fix bug in Shamir's secret sharing utilities affecting Secure Aggregation (#5252)

  Refactors Shamir's secret sharing utilities to fix a bug impacting Secure Aggregation. Thanks to Pinji Chen and Guanheng Liu for their contributions.

• Ensure backward compatibility for RecordDict (#5239, #5270)

  The RecordDict (formerly RecordSet) now maintains full backward compatibility. Legacy usages of RecordSet and its properties are supported, with deprecation warnings logged when outdated references are used. Users are encouraged to transition to the updated RecordDict interface promptly to avoid future issues.

• Refactor and optimize CI/CD for repository restructuring (#5202, #5176, #5200, #5203, #5210, #5166, #5214, #5212, #5209, #5199, #5204, #5201, #5191, #5167, #5248, #5268, #5251)

  Improves CI/CD workflows to align with repository changes. Updates issue templates, fixes Docker and docs jobs, enhances script compatibility, adds checks, and bumps tool versions to streamline development and deployment.

• Improve and clean up documentation (#5233, #5179, #5216, #5211, #5217, #5198, #5168, #5215, #5169, #5171, #5240, #5259)

  Removes outdated content, redundant CLI flags, and unnecessary sections; updates Docker READMEs and virtual environment setup guide; and syncs translation source texts.

• General Improvements (#5241, #5180, #5226, #5173, #5219, #5208, #5158, #5255, #5264, #5272)

  As always, many parts of the Flower framework and quality infrastructure were improved and updated.

Incompatible changes

• Restructure repository (breaking change for contributors only) (#5206, #5194, #5192, #5185, #5184, #5177, #5183, #5207, #5267, #5274)

  Restructures the Flower repository by moving all framework-related code, configs, and dev tools into the framework/ subdirectory. This includes relocating all files under src/, dev scripts, pyproject.toml and other configs. Contributor documentation has been updated to reflect these changes.

Flower 1.17.0

Thanks to our contributors

We would like to give our special thanks to all the contributors who made the new version of Flower possible (in git shortlog order):

Aline Almeida, Charles Beauville, Chong Shen Ng, Daniel Hinjos García, Daniel J. Beutel, Daniel Nata Nugraha, Dimitris Stripelis, Heng Pan, Javier, Robert Steiner, Yan Gao

What's new?

• Allow registration of functions for custom message types (#5093)

  Enables support for custom message types in ServerApp by allowing the message_type field to be set as "<action_type>.<action_name>", where <action_type> is one of train, evaluate, or query, and <action_name> is a valid Python identifier. Developers can now register handler functions for these custom message types using the decorator @app.<action_type>("<action_name>"). For example, the my_echo_fn function is called when the ServerApp sends a message with message_type set to "query.echo", and the get_mean_value function is called when it's "query.mean":

  app = ClientApp()
  
  @app.query("echo")
  def my_echo_fn(message: Message, context: Context):
      # Echo the incoming message
      return Message(message.content, reply_to=message)
  
  @app.query("mean")
  def get_mean_value(message: Message, context: Context):
      # Calculate the mean value
      mean = ...  # Replace with actual computation
  
      # Wrap the result in a MetricRecord, then in a RecordDict
      metrics = MetricRecord({"mean": mean})
      content = RecordDict({"metrics": metrics})
  
      return Message(content, reply_to=message)

• Rename core Message API components for clarity and consistency (#5140, #5133, #5139, #5129, #5150, #5151, #5146, #5152)

  To improve clarity and ensure consistency across the Message API, the following renamings have been made:

  • Driver → Grid
  • RecordSet → RecordDict
  • ParametersRecord → ArrayRecord
  • MetricsRecord → MetricRecord
  • ConfigsRecord → ConfigRecord

  Backward compatibility is maintained for all the above changes, and deprecation notices have been introduced to support a smooth transition.

• Enable seamless conversions between ArrayRecord/Array and NumPy/PyTorch types (#4922, #4920)

  One-liner conversions are now supported between Array and numpy.ndarray or torch.Tensor, and between ArrayRecord (formerly ParametersRecord) and PyTorch state_dict or a list of numpy.ndarray. This simplifies workflows involving model parameters and tensor data structures. Example usage includes ArrayRecord(model.state_dict()) and array_record.to_torch_state_dict(). Refer to the ArrayRecord and Array documentation for details.

• Revamp message creation using Message constructor (#5137, #5153)

  Revamps the Message creation workflow by enabling direct instantiation via the Message(...) constructor. This deprecates the previous APIs and simplifies message creation:

  • Driver.create_message(...) → Message(...)
  • <some_message>.create_reply(...) → Message(..., reply_to=<some_message>)

• Stabilize low-level Message API (#5120)

  With all the changes above, the stability of the low-level Message API has been significantly improved. All preview feature warnings have been removed, marking the completion of its transition out of experimental status.

• Add node availability check to reduce wait time (#4968)

  Adds a node availability check to SuperLink. If the target SuperNode is offline, SuperLink automatically generates an error reply message when the ServerApp attempts to pull the reply. This mechanism helps avoid unnecessary delays in each round caused by waiting for responses from unavailable nodes.

• Enable extensible event logging for FleetServicer and ExecServicer (#4998, #4997, #4951, #4950, #5108)

  Introduces the necessary hooks and infrastructure to support RPC event logging for FleetServicer and ExecServicer. This enables advanced auditing and observability of RPC calls made by flwr CLI users and SuperNodes, when appropriate event log plugins are available.

• Add CareQA benchmark for medical LLM evaluation (#4966)

  Adds the CareQA dataset as a new benchmark for evaluating medical knowledge in LLMs. CareQA consists of 5,621 QA pairs from official Spanish healthcare exams (2020–2024), translated to English and covering multiple disciplines. This enhances the diversity of datasets used in the Flower Medical LLM Leaderboard.

• Fix docstrings and improve handling of Ray nodes without CPU resources (#5155, #5076, #5132)

  Fixes inaccurate or outdated docstrings in RecordDict and the FlowerClient used in FlowerTune templates, improving documentation clarity. Also adds handling for Ray nodes that report zero CPU resources, preventing potential runtime issues.

• Improve documentation and examples (#5162, #5079, #5123, #5066, #5143, #5118, #5148, #5134, #5080, #5160, #5069, #5032)

• Update CI/CD (#5125, #5062, #5056, #5048, #5065, #5061, #5057, #5064, #5144)

• General Improvements (#5074, #5126, #5122, #5149, #5157)

  As always, many parts of the Flower framework and quality infrastructure were improved and updated.

Flower 1.16.0

Thanks to our contributors

We would like to give our special thanks to all the contributors who made the new version of Flower possible (in git shortlog order):

Alan Silva, Andrej Jovanović, Charles Beauville, Chong Shen Ng, Daniel J. Beutel, Dimitris Stripelis, Heng Pan, Javier, Kevin Ta, Li Shaoyu, Mohammad Naseri, Taner Topal, Yan Gao

What's new?

• Enhance RecordSet and Array for improved usability (#4963, #4980, #4918)

  RecordSet now supports dictionary-like access, allowing interactions similar to built-in Python dictionaries. For example, instead of recordset.parameters_records["model"], users can simply use recordset["model"]. This enhancement maintains backward compatibility with existing recordset.*_records properties.

  Additionally, the Array class now accepts numpy.ndarray instances directly in its constructor, enabling instantiation with a NumPy array via Array(your_numpy_ndarray).

• Support function-specific Flower Mods for ClientApp (#4954, #4962)

  Flower Mods can now be applied to individual functions within the ClientApp rather than affecting the entire application. This allows for more granular control. The documentation has been updated to reflect these changes — please refer to How to Use Built-in Mods for details.

• Introduce @app.lifespan() for lifecycle management (#4929, #4986)

  ServerApp and ClientApp now support @app.lifespan(), enabling custom enter/exit handlers for resource setup and cleanup. Throughout the entire FL training, these handlers in ClientApp may run multiple times as instances are dynamically managed.

• Add FedRAG example (#4955, #5036, #5042)

  Adds a FedRAG example, integrating Federated Learning with Retrieval Augmented Generation (RAG). This approach allows Large Language Models (LLMs) to query distributed data silos without centrally aggregating the corpora, enhancing performance while preserving data privacy.

• Upgrade FedProx baseline to a Flower App (#4937)

  Updates FedProx to the a Flower App by removing Hydra, migrating configs to pyproject.toml, using ClientApp and ServerApp, integrating flwr-datasets with DistributionPartitioner, enabling result saving, and updating README.md. This baseline now supports flwr run.

• Migrate framework to Message-based system (#4959, #4993, #4979, #4999)

  The Flower framework has been fully migrated from a TaskIns/TaskRes-based system to a Message-based system, aligning with the user-facing Message class. This includes adding validator functions for Message, introducing LinkState methods that operate on Message, updating LinkState to use Message-only methods, and removing the Task-related code entirely.

• Introduce event logging extension points (#4948, #5013)

  Begins implementing an event logging system for SuperLink, allowing RPC calls to be logged when enabled. These changes introduce initial extension points.

• Increase default TTL and message size (#5011, #5028)

  The default TTL for messages is now 12 hours (up from 1 hour), and the gRPC message size limit has increased from 512MB to 2GB. TTL sets a hard limit on the time between the ServerApp sending an instruction and receiving a reply from the ClientApp.

• Improve documentation (#4945, #4965, #4994, #4964, #4991, #5014, #4970, #4990, #4978, #4944, #5022, #5007, #4988, #5053)

• Update CI/CD (#4943, #4942, #4953, #4985, #4984, #5025, #4987, #4912, #5049)

• Bugfixes (#4969, #4974, #5017, #4995, #4971, #5037, #5038)

• General Improvements (#4947, #4972, #4992, #5020, #5018, #4989, #4957, #5000, #5012, #5001)

  As always, many parts of the Flower framework and quality infrastructure were improved and updated.

Incompatible changes

• Remove deprecated CLI commands (#4855)

  Removes deprecated CLI commands: flower-server-app, flower-superexec, and flower-client-app. These commands are no longer available in the framework.

• Bump minimum Python and cryptography versions (#4946)

  Bumps the minimum Python version from 3.9 to 3.9.2 and updates the cryptography package from 43.0.1 to 44.0.1. This change ensures compatibility with the latest security updates and features.