Fix ZeroDivisionError in SFT during record_train_metrics#3581
Merged
copybara-service[bot] merged 1 commit intomainfrom Apr 10, 2026
Merged
Fix ZeroDivisionError in SFT during record_train_metrics#3581copybara-service[bot] merged 1 commit intomainfrom
copybara-service[bot] merged 1 commit intomainfrom
Conversation
Codecov Report❌ Patch coverage is
📢 Thoughts on this report? Let us know! |
igorts-git
force-pushed
the
igorts/sft-timing-fix
branch
from
be0941c to
3c3bfff
Compare
igorts-git
force-pushed
the
igorts/sft-timing-fix
branch
from
3c3bfff to
4440993
Compare
igorts-git
requested review from
A9isha,
NicoGrande,
NuojCheng,
RissyRan,
SurbhiJainUSC,
abhinavclemson,
aireenmei,
bvandermoon,
dipannita08,
gagika,
gobbleturk,
hengtaoguo,
jesselu-google,
jiangjy1982,
khatwanimohit,
richjames0,
shralex,
suexu1025 and
vipannalla
as code owners
| train_step: int, | ||
| train_loss: float, | ||
| step_time: float, | ||
| step_time: float = 0.0, # No longer provided. See https://github.com/google/tunix/pull/1289. |
Collaborator
There was a problem hiding this comment.
We can also modify the hooks interface in Tunix and remove step_time. Can be done as a follow up.
SurbhiJainUSC
approved these changes
Apr 10, 2026
copybara-service bot
pushed a commit
to google/tunix
that referenced
this pull request
Apr 10, 2026
…_step_end` function. MaxText we change this flag to optional in AI-Hypercomputer/maxtext#3581. This will allow MaxText to be compatible with older versions of Tunix. Eventually we will delete it there as well. PiperOrigin-RevId: 897791378
copybara-service bot
pushed a commit
to google/tunix
that referenced
this pull request
Apr 10, 2026
…_step_end` function. In MaxText we changed this flag to optional in AI-Hypercomputer/maxtext#3581. This will allow MaxText to be compatible with older versions of Tunix. Eventually we will delete it there as well. PiperOrigin-RevId: 897791378
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.
Description
This PR fixes a ZeroDivisionError that occurs in the MaxText MetricLogger during Tunix-based SFT runs.
The Problem
As the result of Tunix PR1289. The Tunix library's internal training loop now hard-codes the step time to zero when calling the
on_train_step_endhook.The MaxText SFT hook passes the zero value to the
metric_logger.The MetricLogger divides FLOPS and tokens values by the passed zero
step_time.The Fix
This PR modifies SFTTrainingHooks in src/maxtext/trainers/post_train/sft/hooks.py to independently track time intervals in between calls to
on_train_step_endand using this time interval for reporting step time metrics.FIXES: b/497665859
Tests
Manual invocation of SFT training with and without the fix with Tunix installed from main/HEAD.
Checklist
Before submitting this PR, please make sure (put X in square brackets):
gemini-reviewlabel.