v0.7.7

What's Changed

• Use latest setuptools by @linshokaku in #805
• Support torch 2.2 by @take-cheeze in #806
• Update PyTorch installation in nightly test workflow by @linshokaku in #808
• Bump black from 23.3.0 to 24.3.0 in /tests by @dependabot in #811
• [onnx] Also check validation error in model load by @take-cheeze in #810
• Fix CuPy version installed in CI by @kmaehashi in #817
• Allow running nightly CI against pull requests by @kmaehashi in #816
• ignore warnings inside the nightly version of torch by @linshokaku in #815
• Upload wheel by @linshokaku in #776
• Avoid using pkg_resources by @asi1024 in #824
• bump version to v0.7.7 by @linshokaku in #823
• Add long_description to package by @kmaehashi in #825

New Contributors

• @dependabot made their first contribution in #811

Full Changelog: v0.7.6...v0.7.7

v0.7.6

Enhancements and bug-fixes

• ShardedSnapshot by @linshokaku in #799
  • e.g. FSDP model support for state_dict/load_state_dict
• Accumulate Extensions by @linshokaku in #761
  • Provides an aggregate extension of log values in average/min/max/std, etc.
  • Provides aggregation of values between processes in a distributed environment
• Running pytest in an MPI environment by @linshokaku in #763
• Append mode ChromeTracer by @linshokaku in #801
• Change output file destination in pytest to a temporary directory by @linshokaku in #802
• Fix Logic.consume_options() by @linshokaku in #769
• fix test by @linshokaku

See the list of merged pull-requests for the details.

v0.7.5

Enhancements and bug-fixes

This release includes the following enhancements and bug fixes:

• Add CI for PyTorch 2.1 by @take-cheeze in #785
• Support chrome tracing in ppe.profiler by @emcastillo in #742
• Add more type annotation of np.zeros by @take-cheeze in #786
• Trigger for using user custom functions by @linshokaku in #789
• ppe.compile Enable forward only and custom decompositions by @emcastillo in #740
• Add interface for register custom ops in torch.ops.ppe by @emcastillo in #796
• Fix some test by @linshokaku @asi1024

See the list of merged pull-requests for the details.

v0.7.4

What's Changed

• Support for ubuntu 18.04 has been discontinued due to EOL by @linshokaku in #765
• Fixed an issue where tf32 was implicitly used for tests related to onnx by @take-cheeze in #768
• pfto float8 support by @niboshi in #781
• Fixed possible file corruption when running DistributedSnapshot.on_error() by @linshokaku in #780

New Contributors

• @niboshi made their first contribution in #781

Full Changelog: v0.7.3...v0.7.4

v0.7.3

Enhancements and bug-fixes

This release includes the following enhancements and bug fixes:

• Add distributed option to the create_evaluator function to support creating DistributedEvaluator (#675, #733)

See the list of merged pull-requests for the details.

v0.7.2

Enhancements and bug-fixes

This release includes the following enhancements and bug fixes:

• Add wait_for_first_optimizer_step option to LRScheduler extension that prevents stepping LR schedulers before stepping optimizers. #737
• Fix IntervalTrigger trigger to correctly handle an epoch number given in float. #738
• Fix Snapshot extension to only trigger the cleanup process in the saver rank. #732
• Improve ppe.cuda.stream to propagate PyTorch's device ID to CuPy. #745
• Add Quick Start documentation. #743
• Improved API Reference. #731

See the list of merged pull-requests for the details.

v0.7.1

Enhancements and bug-fixes

This release includes the following enhancements and bug fixes:

• Fix CUDA profiling with ppe.profiler to respect the current device #726

See the list of merged pull-requests for the details.

v0.7.0

Enhancements and bug-fixes

This release includes the following enhancements and bug fixes:

• ppe.compile API to produce computational graphs with joint backward and optimizer steps #714
• Snapshot support for objects with state_dict passed in the options kwarg to Trainer and Manager #702
• Evaluator with float16 support #698
• Multiple fixes to pfto #669 #684 #687 #715
• Fixed a problem in which autoload of snapshot could not be read when a Writer is specified. #692
• Fixed a problem that caused model train/eval mode switching after Evaluator execution to not work correctly. #697

No compatibility changes

• Support for PyTorch 1.9 and Python 3.7 has been removed from the code base #706

See the list of merged pull-requests for the details.

v0.6.7

Enhancements and bug-fixes

This release includes the following enhancements and bug fixes:

• Use state_dict to decide name of initializers #670
• Fix autocast options consumption with trainer #668
• Various PFTO fixes

See the list of merged pull-requests for the details.

v0.6.6

Enhancements and bug-fixes

This release includes the following enhancements and bug fixes:

• Support PyTorch 2.0 #645
• Add a Logic class for clousure mode optimization #657
• Enable new PyTorch autocast options #649

See the list of merged pull-requests for the details.