Skip to content

Fix ZeroDivisionError in SFT during record_train_metrics #3581

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
copybara-service merged 1 commit into from
Apr 10, 2026
+9 −2
Merged

Fix ZeroDivisionError in SFT during record_train_metrics #3581

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
11 changes: 9 additions & 2 deletions src/maxtext/trainers/post_train/sft/hooks.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -25,6 +25,7 @@
from typing_extensions import override

import jax
import time
import jax.numpy as jnp

from flax import nnx
Expand Down Expand Up @@ -54,6 +55,7 @@ def __init__(self, config, mesh, learning_rate_schedule, goodput_recorder):
self.metadata = {}
self.train_metadata = defaultdict(float)
self.eval_metadata = defaultdict(float)
self.step_start_time = 0.0

@override
def on_train_start(self, train_ctx: peft_trainer.PeftTrainer):
Expand All @@ -79,6 +81,7 @@ def on_train_start(self, train_ctx: peft_trainer.PeftTrainer):
)

self.metadata["first_train_step"] = train_ctx.train_steps
self.step_start_time = time.perf_counter()

@override
def on_train_end(self, train_ctx: peft_trainer.PeftTrainer): # pylint: disable=unused-argument
Expand Down Expand Up @@ -109,22 +112,26 @@ def on_train_step_end(
train_ctx: peft_trainer.PeftTrainer,
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.
Copy link
Copy Markdown
Collaborator

@SurbhiJainUSC SurbhiJainUSC Apr 10, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

We can also modify the hooks interface in Tunix and remove step_time. Can be done as a follow up.

):
"""Called at the end of training step.
This hook is called by Tunix after the step counter has been incremented for logging purposes.
Therefore, using `train_step - 1` to refer to the state of the previous step counter.
However, we will use the current `train_step` value to record metrics in this hook to be
consistent with Tunix's metric logging convention.
"""

assert train_step - 1 in self.train_metadata, (
"SFTTrainingHooks.on_train_step_start() must be called before" " SFTTrainingHooks.on_train_step_end()"
)

if self.metadata["first_train_step"] == train_step - 1:
max_utils.print_mem_stats("After params initialized")

# Use our own timing since Tunix passes 0.0
current_time = time.perf_counter()
step_time = current_time - self.step_start_time
self.step_start_time = current_time

metrics = {
"scalar": {
"learning/loss": train_loss,
Expand Down
Loading