fix: Skip val_check_interval validation when limit_val_batches=0

[ManasVardhan]

Fixes #21553

When limit_val_batches=0 (validation disabled), the val_check_interval sanity check is now skipped since validation won't run anyway.

Changes

• In _FitLoop.setup_data(), added an early check: when trainer.limit_val_batches == 0, set val_check_batch = inf and skip all interval validation logic.
• Added regression test confirming val_check_interval > limit_train_batches no longer raises when validation is disabled.

---

📚 Documentation preview 📚: https://pytorch-lightning--21560.org.readthedocs.build/en/21560/

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 79%. Comparing base (8e805f9) to head (fc89bb4).
⚠️ Report is 2 commits behind head on master.
✅ All tests successful. No failed tests found.

❗ There is a different number of reports uploaded between BASE (8e805f9) and HEAD (fc89bb4). Click for more details.

HEAD has 726 uploads less than BASE

Flag	BASE (8e805f9)	HEAD (fc89bb4)
cpu	198	33
python	18	3
lightning_fabric	54	0
pytest	99	0
python3.10	18	3
lightning	90	15
python3.12	54	9
python3.12.7	54	9
python3.13	18	3
python3.11	36	6
pytorch2.3	9	3
pytest-full	99	33
pytorch2.1	18	6
pytorch2.8	18	6
pytorch_lightning	54	18
pytorch2.5.1	9	3
pytorch2.7	9	3
pytorch2.2.2	9	3
pytorch2.4.1	9	3
pytorch2.9	9	3
pytorch2.6	9	3

Additional details and impacted files

@@            Coverage Diff            @@
##           master   #21560     +/-   ##
=========================================
- Coverage      87%      79%     -8%     
=========================================
  Files         270      267      -3     
  Lines       24077    24018     -59     
=========================================
- Hits        20865    18961   -1904     
- Misses       3212     5057   +1845     

[ManasVardhan]

Addressed both review comments:

1. Moved comments inside the if-elif blocks
2. Removed the extra blank line in the test docstring

Thanks for the review @deependujha!

[deependujha]

thanks. Can you update the changelog too.

[ManasVardhan]

Updated:

• Added changelog entry under [unreleased] > Fixed
• Collapsed test docstring to single line

Thanks for the review!

[taha-yassine]

Hey @ManasVardhan, thanks for picking this up so quickly. I was planning to submit a PR as well. Since I opened the original issue and proposed the solution, would you be open to adding me as a co-author?

[ManasVardhan]

Hey @ManasVardhan, thanks for picking this up so quickly. I was planning to submit a PR as well. Since I opened the original issue and proposed the solution, would you be open to adding me as a co-author?

Hi Taha, yes absolutely.

[ManasVardhan]

Re: why not None instead of float("inf"):

val_check_batch is compared against float("inf") in training_epoch_loop.py (lines 424 and 438) to determine validation scheduling. Using None would break those comparisons. float("inf") ensures the validation check condition is never satisfied, which is the correct behavior when validation is disabled.

[deependujha]

not sure with your reasoning.

[ManasVardhan]

@deependujha Re: why float("inf") instead of None - val_check_batch is compared against float("inf") downstream in training_epoch_loop.py (line 535: is_infinite_dataset = self.trainer.val_check_batch == float("inf")), and when it's not inf, there's an assert self.trainer.val_check_batch is not None on line 556 before using it in a modulo operation. Setting it to None would risk hitting that assertion. float("inf") follows the existing convention for "no batch-based validation check" (same value used for iterable datasets without a defined length).

[deependujha]

thanks for the clarification, but i believe, the line won't even get executed.

pytorch-lightning/src/lightning/pytorch/loops/training_epoch_loop.py

Lines 523 to 536 in a3a28b3

	def _should_check_val_epoch(self) -> bool:
	return self.trainer.enable_validation and (
	self.trainer.check_val_every_n_epoch is None
	or (self.trainer.current_epoch + 1) % self.trainer.check_val_every_n_epoch == 0
	)
	def _should_check_val_fx(self, data_fetcher: _DataFetcher) -> bool:
	"""Decide if we should run validation."""
	if not self._should_check_val_epoch():
	return False
	# val_check_batch is inf for iterable datasets with no length defined
	is_infinite_dataset = self.trainer.val_check_batch == float("inf")
	is_last_batch = self.batch_progress.is_last_batch

_should_check_val_fx calls _should_check_val_epoch, and it calls self.trainer.enable_validation, which will return False, and hence, the function should_check_val_fx will return False (line 532), and those inf checks & assertion won't even run.

pytorch-lightning/src/lightning/pytorch/trainer/trainer.py

Lines 1353 to 1360 in a3a28b3

	@property
	def enable_validation(self) -> bool:
	"""Check if we should run validation during training."""
	return (
	self.fit_loop.epoch_loop.val_loop._data_source.is_defined()
	and is_overridden("validation_step", self.lightning_module)
	and self.limit_val_batches > 0
	)

• limit_val_batches is already set to 0, so will return False.

[ManasVardhan]

Good catch @deependujha! You're right, when limit_val_batches=0, enable_validation returns False so none of those inf checks or assertions ever run. None would totally work here too.

I went with float("inf") just because that's what the rest of the codebase uses for "skip this", but honestly it doesn't matter much in this case. Happy to switch to None if you think that reads better - lmk!

[deependujha]

thanks, but I believe the current changes are enough. Thanks to both of you for the great work. Should be landed soon.

[ManasVardhan]

CI is all green -- ready to merge whenever you get a chance! Thanks for the review 🙏