Name | Modified | Size | Downloads / Week |
---|---|---|---|
Parent folder | |||
pfl-0.3.0.tar.gz | 2025-05-28 | 165.6 kB | |
pfl-0.3.0-py3-none-any.whl | 2025-05-28 | 224.4 kB | |
pfl 0.3.0 source code.tar.gz | 2025-05-28 | 2.7 MB | |
pfl 0.3.0 source code.zip | 2025-05-28 | 2.9 MB | |
README.md | 2025-05-28 | 1.4 kB | |
Totals: 5 Items | 5.9 MB | 0 |
v0.3.0
New features
- Implemented
MLXModel
to support training with MLX (#80). - Implemented Horovod-compatible
Barrier
class (#100). - Implemented
JointMechanism
to enable applying different DP noise on subsets of the statistics (#105). - Implemented
JointPrivacyAccountant
for joint privacy accounting in the case of usingJointMechanism
(#111). - Implemented policy-based model checkpointing callbacks (#106).
Tasks completed
- Added MLX image classification example (#80).
- Added MLX language model example (#82).
- Added notebook tutorial for training with MLX (#81).
- Updated notebooks to work on Colab (#95).
- Added CITATION.cff (#99).
- Don't hardcode for CUDA 11.8 (#108).
Bug fixes
- Fixed bug in
PyTorchFederatedDataset
where it sometimes hang (#98). - Fix
PyTorchSeedScope
for non-CPU random states (#100). - Respect
CUDA_VISIBLE_DEVICES
if it is set for process (#100). - Fixed algorithm states not working properly when restored from
Saveable
(#103). - Fixed edge case for having
HyperParams
insideHyperParams
(#112).