feat: lr scheduler#151

Open

Chamberlain0w0 wants to merge 17 commits into

masterfrom

feat/lr_scheduler

Commits on Jun 18, 2026

refactor(optimizer): hoist learning_rate_ to Optimizer base and add lr accessors, passthrough SetLearningRate/GetLearningRate, and add initial_learning_rate and it's accessors

kinorw
authored and
Chamberlain0w0
committed
feat(lr_scheduler): add LRScheduler abstract base class, ConstantLR, StepLR, LinearLR, LambdaLR and SequentialLR

kinorw
authored and
Chamberlain0w0
committed
refactor(lr_scheduler): apply template method pattern to LRScheduler base class, add factory method Create<T>() with two-phase init and update all tests to use Create<T>() factory method.

kinorw
authored and
Chamberlain0w0
committed
feat(lr_scheduler): add ChainedScheduler composite strategy, and add closed and chained form, adjust LinearLR、SequentialLR

kinorw
authored and
Chamberlain0w0
committed
feat(lr_scheduler): add scheduler factory for CLI integration

kinorw
authored and
Chamberlain0w0
committed
refactor(examples): add scheduler placeholder and use runtime lr in logs, and integrate scheduler into training loop

kinorw
authored and
Chamberlain0w0
committed
style: apply clang-format to all legacy files

littleotherut
authored and
Chamberlain0w0
committed
refactor: rename current_lr_ to recover_lr_ and update related methods, add validation tests for learning rate schedulers

kinorw
authored and
Chamberlain0w0
committed
fix: adapt to megatron-style arguments
Chamberlain0w0
committed
style: fix setter/getter name
Chamberlain0w0
committed
fix: add lr_scheduler test group in config
Chamberlain0w0
committed
fix: resolve comments
Chamberlain0w0
committed
refactor: reorganize unit tests to comply with ctest infra
Chamberlain0w0
committed
fix: save runtime lr in checkpoints
Chamberlain0w0
committed
fix: support checkpoint of lr schedulers
Chamberlain0w0
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

feat: lr scheduler#151

feat: lr scheduler#151
Chamberlain0w0 wants to merge 17 commits into
masterfrom
feat/lr_scheduler

Commits on Jun 18, 2026

refactor(optimizer): hoist learning_rate_ to Optimizer base and add lr accessors, passthrough SetLearningRate/GetLearningRate, and add initial_learning_rate and it's accessors

feat(lr_scheduler): add LRScheduler abstract base class, ConstantLR, StepLR, LinearLR, LambdaLR and SequentialLR

refactor(lr_scheduler): apply template method pattern to LRScheduler base class, add factory method Create<T>() with two-phase init and update all tests to use Create<T>() factory method.

feat(lr_scheduler): add ChainedScheduler composite strategy, and add closed and chained form, adjust LinearLR、SequentialLR

feat(lr_scheduler): add scheduler factory for CLI integration

refactor(examples): add scheduler placeholder and use runtime lr in logs, and integrate scheduler into training loop

style: apply clang-format to all legacy files

refactor: rename current_lr_ to recover_lr_ and update related methods, add validation tests for learning rate schedulers

fix: adapt to megatron-style arguments

style: fix setter/getter name

fix: add lr_scheduler test group in config

fix: resolve comments

refactor: reorganize unit tests to comply with ctest infra

fix: save runtime lr in checkpoints

fix: support checkpoint of lr schedulers

Commits on Jun 23, 2026

fix: remove last_lr, refactor some sanity checks

fix: remove save/load lr_scheduler bool flags