fix(callbacks): Defer step/time-triggered ModelCheckpoint saves until validation metrics are available #21106
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
littlebullGit
requested review from
Borda,
ethanwharris,
justusschock,
lantiga and
tchaton
as code owners
github-actions
bot
added
pl
Codecov Report❌ Patch coverage is Additional details and impacted files@@ Coverage Diff @@
## master #21106 +/- ##
=======================================
Coverage 87% 87%
=======================================
Files 269 269
Lines 23520 23545 +25
=======================================
+ Hits 20508 20542 +34
+ Misses 3012 3003 -9 |
|
@Borda , I cannot see the error in the two failed jobs. Can you help me or point me to what the error is ? |
these tests are optional for now as the same are failing also on master |
Borda
approved these changes
Aug 22, 2025
… validation metrics are available Root cause: - With `every_n_train_steps` (or `train_time_interval`), checkpoints could save at train batch end before validation ran, so the monitored val metric was missing/stale and `best_model_score` was incorrect. (Refs Lightning-AI#20919) Fix: - In [src/lightning/pytorch/callbacks/model_checkpoint.py:ModelCheckpoint.on_train_batch_end]: - Defer saves when the monitored key is missing from [trainer.callback_metrics] - If on the last train batch and not saving at train-epoch-end, defer only when validation will run next: - `trainer.enable_validation` is True - `trainer.num_val_batches` > 0 - `trainer.check_val_every_n_epoch` schedule matches the upcoming epoch - Perform deferred saves in [on_validation_end], ensuring fresh validation metrics are used. - Allow zero `timedelta` for `train_time_interval` and broadcast the time-trigger decision across ranks. - Do not defer when monitoring a train metric or when no validation is scheduled. Tests: - Repro (previously failing, now passing): - [tests/tests_pytorch/callbacks/test_model_checkpoint_step_interval_val_metric.py] - Additional validations: - [tests/tests_pytorch/callbacks/test_model_checkpoint_additional_cases.py] - [tests/tests_pytorch/callbacks/test_model_checkpoint_edge_cases.py] Outcome: - `best_model_score` matches the validation metric after the epoch. - Step/time-interval checkpointing behaves correctly without premature or skipped saves.
littlebullGit
force-pushed
the
fix/20919-checkpoint-step-val-metric
branch
from
094b278 to
6c1554a
Compare
justusschock
approved these changes
Aug 28, 2025
Borda
added a commit
that referenced
this pull request
Sep 3, 2025
… validation metrics are available (#21106) * fix(callbacks): defer step/time-triggered ModelCheckpoint saves until validation metrics are available Root cause: - With `every_n_train_steps` (or `train_time_interval`), checkpoints could save at train batch end before validation ran, so the monitored val metric was missing/stale and `best_model_score` was incorrect. (Refs #20919) Fix: - In [src/lightning/pytorch/callbacks/model_checkpoint.py:ModelCheckpoint.on_train_batch_end]: - Defer saves when the monitored key is missing from [trainer.callback_metrics] - If on the last train batch and not saving at train-epoch-end, defer only when validation will run next: - `trainer.enable_validation` is True - `trainer.num_val_batches` > 0 - `trainer.check_val_every_n_epoch` schedule matches the upcoming epoch - Perform deferred saves in [on_validation_end], ensuring fresh validation metrics are used. - Allow zero `timedelta` for `train_time_interval` and broadcast the time-trigger decision across ranks. - Do not defer when monitoring a train metric or when no validation is scheduled. Tests: - Repro (previously failing, now passing): - [tests/tests_pytorch/callbacks/test_model_checkpoint_step_interval_val_metric.py] - Additional validations: - [tests/tests_pytorch/callbacks/test_model_checkpoint_additional_cases.py] - [tests/tests_pytorch/callbacks/test_model_checkpoint_edge_cases.py] Outcome: - `best_model_score` matches the validation metric after the epoch. - Step/time-interval checkpointing behaves correctly without premature or skipped saves. * test: disable logger in model checkpoint tests to avoid side effects * chlog --------- Co-authored-by: Jirka B <[email protected]> (cherry picked from commit b1cc925)
lantiga
pushed a commit
that referenced
this pull request
Sep 5, 2025
… validation metrics are available (#21106) * fix(callbacks): defer step/time-triggered ModelCheckpoint saves until validation metrics are available Root cause: - With `every_n_train_steps` (or `train_time_interval`), checkpoints could save at train batch end before validation ran, so the monitored val metric was missing/stale and `best_model_score` was incorrect. (Refs #20919) Fix: - In [src/lightning/pytorch/callbacks/model_checkpoint.py:ModelCheckpoint.on_train_batch_end]: - Defer saves when the monitored key is missing from [trainer.callback_metrics] - If on the last train batch and not saving at train-epoch-end, defer only when validation will run next: - `trainer.enable_validation` is True - `trainer.num_val_batches` > 0 - `trainer.check_val_every_n_epoch` schedule matches the upcoming epoch - Perform deferred saves in [on_validation_end], ensuring fresh validation metrics are used. - Allow zero `timedelta` for `train_time_interval` and broadcast the time-trigger decision across ranks. - Do not defer when monitoring a train metric or when no validation is scheduled. Tests: - Repro (previously failing, now passing): - [tests/tests_pytorch/callbacks/test_model_checkpoint_step_interval_val_metric.py] - Additional validations: - [tests/tests_pytorch/callbacks/test_model_checkpoint_additional_cases.py] - [tests/tests_pytorch/callbacks/test_model_checkpoint_edge_cases.py] Outcome: - `best_model_score` matches the validation metric after the epoch. - Step/time-interval checkpointing behaves correctly without premature or skipped saves. * test: disable logger in model checkpoint tests to avoid side effects * chlog --------- Co-authored-by: Jirka B <[email protected]> (cherry picked from commit b1cc925)
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Defer step/time-triggered ModelCheckpoint saves until validation metrics are available
Fixes #20919
Root cause
every_n_train_steps(ortrain_time_interval), checkpoints could save at train-batch end before validation ran. The monitored validation metric was missing/stale, sobest_model_scorecould be incorrect.Fix
trainer.enable_validationis Truetrainer.num_val_batches> 0trainer.check_val_every_n_epochschedule matches the upcoming epochtimedeltafortrain_time_intervaland broadcast the time-trigger decision across ranks viatrainer.strategy.broadcast.Tests
Outcome
best_model_scorematches the latest validation metric.📚 Documentation preview 📚: https://pytorch-lightning--21106.org.readthedocs.build/en/21106/