Skip to content

[Tunix] Remove utils.time_measure from the training loop. #1118

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
Mar 9, 2026
Merged

[Tunix] Remove utils.time_measure from the training loop. #1118

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
163 changes: 83 additions & 80 deletions tunix/sft/peft_trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -622,95 +622,98 @@ def train(
train_iterator = iter(train_ds)
index = 0
last_step_completion_time = time.perf_counter()
with utils.time_measure("Train loop"):
while True:
self._prof.maybe_activate(self._iter_steps)
with jax.profiler.StepTraceAnnotation(
"train", step_num=self._iter_steps
while True:
self._prof.maybe_activate(self._iter_steps)
with jax.profiler.StepTraceAnnotation(
"train", step_num=self._iter_steps
):
train_example = None
if self.data_hooks:
train_example = self.data_hooks.load_next_train_batch(self)
else:
try:
train_example = next(train_iterator)
if not self.is_managed_externally:
# TODO(mridulsahu): Add support to restore the iterator state
# instead of skipping the already trained examples.
if index < self._iter_steps:
# Skip the examples that are already trained.
index += 1
continue
index += 1
except StopIteration:
pass

if train_example is None:
break

# Stop training if max_steps is reached.
if (
self.config.max_steps is not None
and self._train_steps >= self.config.max_steps
):
train_example = None
if self.data_hooks:
train_example = self.data_hooks.load_next_train_batch(self)
else:
try:
train_example = next(train_iterator)
if not self.is_managed_externally:
# TODO(mridulsahu): Add support to restore the iterator state
# instead of skipping the already trained examples.
if index < self._iter_steps:
# Skip the examples that are already trained.
index += 1
continue
index += 1
except StopIteration:
pass

if train_example is None:
break

# Stop training if max_steps is reached.
if (
self.config.max_steps is not None
and self._train_steps >= self.config.max_steps
):
break
break

train_example = self._prepare_inputs(train_example)
train_example = sharding_utils.shard_input(
train_example, self.config.data_sharding_axis
)
train_example = self._prepare_inputs(train_example)
train_example = sharding_utils.shard_input(
train_example, self.config.data_sharding_axis
)

self._throttler.wait_for_next()
if self.training_hooks:
self.training_hooks.on_train_step_start(self)

with self._perf_tracer.span(
"peft_train_step", pxla.thread_resources.env.physical_mesh.devices
) as span:
train_loss, aux = train_step(train_example)
span.device_end([train_loss])

current_time = time.perf_counter()
step_time_delta = current_time - last_step_completion_time
last_step_completion_time = current_time

self._throttler.add_computation(train_loss)
self._buffered_train_metrics = self._buffer_metrics(
self._buffered_train_metrics,
loss=train_loss,
step=self._train_steps,
step_time_delta=step_time_delta,
self._throttler.wait_for_next()
if self.training_hooks:
self.training_hooks.on_train_step_start(self)

with self._perf_tracer.span(
"peft_train_step", pxla.thread_resources.env.physical_mesh.devices
) as span:
train_loss, aux = train_step(train_example)
span.device_end([train_loss])

current_time = time.perf_counter()
step_time_delta = current_time - last_step_completion_time
last_step_completion_time = current_time

self._throttler.add_computation(train_loss)
self._buffered_train_metrics = self._buffer_metrics(
self._buffered_train_metrics,
loss=train_loss,
step=self._train_steps,
step_time_delta=step_time_delta,
)
# NB: put this after self._buffer_metrics is important
self._post_process_train_step(aux)
self._iter_steps += 1

if (
self._iter_steps
% self.config.get_with_default("gradient_accumulation_steps", 1)
== 0
):
self._train_steps += 1
self._write_train_metrics()

# Checkpoint frequency is configured by checkpointing_options.
self.checkpoint_manager.save(
self._train_steps,
self.model,
self.optimizer,
save_only_lora_params=self._lora_enabled,
custom_metadata=self.custom_checkpoint_metadata(),
)
# NB: put this after self._buffer_metrics is important
self._post_process_train_step(aux)
self._iter_steps += 1

if (
self._iter_steps
% self.config.get_with_default("gradient_accumulation_steps", 1)
== 0
eval_ds
and self._train_steps % self.config.eval_every_n_steps == 0
):
self._train_steps += 1
self._write_train_metrics()

# Checkpoint frequency is configured by checkpointing_options.
self.checkpoint_manager.save(
self._train_steps,
self.model,
self.optimizer,
save_only_lora_params=self._lora_enabled,
custom_metadata=self.custom_checkpoint_metadata(),
)

if (
eval_ds
and self._train_steps % self.config.eval_every_n_steps == 0
):
self._run_eval(eval_ds, eval_step)

self._prof.maybe_deactivate(self._iter_steps)
self._run_eval(eval_ds, eval_step)

self._prof.maybe_deactivate(self._iter_steps)

self._throttler.wait_for_all()
logging.info(
"Train loop finished in: %.4f seconds",
time.perf_counter() - last_step_completion_time,
)
if self.training_hooks:
self.training_hooks.on_train_end(self)
if not self.is_managed_externally:
Expand Down