fix NaN loss not caught by fast-fail check#84
Merged
karpathy merged 2 commits intokarpathy:masterfrom Mar 11, 2026
Merged
Conversation
`train_loss_f > 100` silently passes on NaN because IEEE 754 NaN comparisons always return False. When an agent experiment produces NaN (e.g. from an aggressive LR change), the run wastes the full 5-minute budget instead of failing fast. `not (x <= 100)` catches both >100 and NaN with no added complexity. Co-Authored-By: Claude Opus 4.6 <[email protected]>
sunnypatneedi
added a commit
to sunnypatneedi/autoresearch-muon
that referenced
this pull request
Mar 10, 2026
Fixes adopted from karpathy/autoresearch PRs: - karpathy#84: NaN loss bypasses fast-fail (IEEE 754: NaN > 100 is False). Fix: `not x <= 100`. Applied to both train.py and train_mlx.py. - karpathy#83: ParquetFile handles never closed, causing FD exhaustion on multi-epoch training. Fix: try/finally with pf.close(). - karpathy#107: Save pre-eval checkpoint so eval OOM/crash doesn't lose the entire training run. Removed on successful eval. - karpathy#93: MFU off-by-one: warmup skips 11 steps (0-10), not 10. - karpathy#70: Loss only reported last microstep, not average across grad accumulation. Fix: accumulate loss += detach() / grad_accum_steps. - karpathy#53: Debug checkpoint on loss explosion with step/loss metadata for post-mortem analysis (train.py only, merged into karpathy#84 fix). - karpathy#62: Input validation for --num-shards and --download-workers. Co-Authored-By: Claude Opus 4.6 <[email protected]>
sunnypatneedi
added a commit
to sunnypatneedi/autoresearch-muon
that referenced
this pull request
Mar 10, 2026
Fixes adopted from karpathy/autoresearch PRs: - karpathy#84: NaN loss bypasses fast-fail (IEEE 754: NaN > 100 is False). Fix: `not x <= 100`. Applied to both train.py and train_mlx.py. - karpathy#83: ParquetFile handles never closed, causing FD exhaustion on multi-epoch training. Fix: try/finally with pf.close(). - karpathy#107: Save pre-eval checkpoint so eval OOM/crash doesn't lose the entire training run. Removed on successful eval. - karpathy#93: MFU off-by-one: warmup skips 11 steps (0-10), not 10. - karpathy#70: Loss only reported last microstep, not average across grad accumulation. Fix: accumulate loss += detach() / grad_accum_steps. - karpathy#53: Debug checkpoint on loss explosion with step/loss metadata for post-mortem analysis (train.py only, merged into karpathy#84 fix). - karpathy#62: Input validation for --num-shards and --download-workers. Co-Authored-By: Claude Opus 4.6 <[email protected]>
svlandeg
approved these changes
Mar 10, 2026
Collaborator
svlandeg
left a comment
svlandeg
left a comment
There was a problem hiding this comment.
Oh interesting, I didn't know this about direct nan comparisons! But I can confirm that if not math.nan <= 100 would hit the fast fail, while if math.nan > 100 wouldn't. So this fix is valid.
But wouldn't it be nicer & less error-prone to just check explicitely?
if train_loss is math.nan or train_loss > 100:
Contributor
Author
|
Great suggestion — updated to an explicit check now:\n\n\n\nAlso added accordingly. |
Contributor
Author
|
(Follow-up) Code snippet got stripped in previous comment by shell quoting. The update is now: if math.isnan(train_loss_f) or train_loss_f > 100. |
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
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.
The fast-fail check
train_loss_f > 100silently passes on NaN because IEEE 754 NaN comparisons always return False. When an agent experiment produces NaN (e.g. from an aggressive LR change), the run wastes the full 5-minute budget training on garbage instead of aborting immediately.not (x <= 100)is the standard negated-comparison idiom that catches both >100 and NaN, with no added complexity.