Skip to content
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

storing & logging gradient norm in trainer #27326

Merged
merged 2 commits into from
Feb 19, 2024
Merged

storing & logging gradient norm in trainer #27326

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
fb2c2ce
report grad_norm during training
shijie-wu Nov 6, 2023
111823a
support getting grad_norm from deepspeed
shijie-wu Jan 17, 2024
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
24 changes: 18 additions & 6 deletions src/transformers/trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -195,6 +195,7 @@
from accelerate import __version__ as accelerate_version
from accelerate.utils import (
DistributedDataParallelKwargs,
DistributedType,
GradientAccumulationPlugin,
load_fsdp_model,
load_fsdp_optimizer,
Expand Down Expand Up @@ -1783,6 +1784,7 @@ def _inner_training_loop(
self._total_loss_scalar = 0.0
self._globalstep_last_logged = self.state.global_step
model.zero_grad()
grad_norm: Optional[float] = None

self.control = self.callback_handler.on_train_begin(args, self.state, self.control)

Expand Down Expand Up @@ -1900,19 +1902,27 @@ def _inner_training_loop(
# deepspeed does its own clipping

if is_sagemaker_mp_enabled() and args.fp16:
self.optimizer.clip_master_grads(args.max_grad_norm)
_grad_norm = self.optimizer.clip_master_grads(args.max_grad_norm)
elif self.use_apex:
# Revert to normal clipping otherwise, handling Apex or full precision
nn.utils.clip_grad_norm_(
_grad_norm = nn.utils.clip_grad_norm_(
amp.master_params(self.optimizer),
args.max_grad_norm,
)
else:
self.accelerator.clip_grad_norm_(
_grad_norm = self.accelerator.clip_grad_norm_(
model.parameters(),
args.max_grad_norm,
)

if (
is_accelerate_available()
and self.accelerator.distributed_type == DistributedType.DEEPSPEED
):
grad_norm = model.get_global_grad_norm()
else:
grad_norm = _grad_norm.item() if _grad_norm is not None else None

# Optimizer step
self.optimizer.step()
optimizer_was_run = not self.accelerator.optimizer_step_was_skipped
Expand All @@ -1926,7 +1936,7 @@ def _inner_training_loop(
self.state.epoch = epoch + (step + 1 + steps_skipped) / steps_in_epoch
self.control = self.callback_handler.on_step_end(args, self.state, self.control)

self._maybe_log_save_evaluate(tr_loss, model, trial, epoch, ignore_keys_for_eval)
self._maybe_log_save_evaluate(tr_loss, grad_norm, model, trial, epoch, ignore_keys_for_eval)
else:
self.control = self.callback_handler.on_substep_end(args, self.state, self.control)

Expand All @@ -1941,7 +1951,7 @@ def _inner_training_loop(
self.control.should_training_stop = True

self.control = self.callback_handler.on_epoch_end(args, self.state, self.control)
self._maybe_log_save_evaluate(tr_loss, model, trial, epoch, ignore_keys_for_eval)
self._maybe_log_save_evaluate(tr_loss, grad_norm, model, trial, epoch, ignore_keys_for_eval)

if DebugOption.TPU_METRICS_DEBUG in self.args.debug:
if is_torch_tpu_available():
Expand Down Expand Up @@ -2246,7 +2256,7 @@ def _issue_warnings_after_load(self, load_result):
f"There were unexpected keys in the checkpoint model loaded: {load_result.unexpected_keys}."
)

def _maybe_log_save_evaluate(self, tr_loss, model, trial, epoch, ignore_keys_for_eval):
def _maybe_log_save_evaluate(self, tr_loss, grad_norm, model, trial, epoch, ignore_keys_for_eval):
if self.control.should_log and self.state.global_step > self._globalstep_last_logged:
if is_torch_tpu_available():
xm.mark_step()
Expand All @@ -2260,6 +2270,8 @@ def _maybe_log_save_evaluate(self, tr_loss, model, trial, epoch, ignore_keys_for
tr_loss -= tr_loss

logs["loss"] = round(tr_loss_scalar / (self.state.global_step - self._globalstep_last_logged), 4)
if grad_norm is not None:
logs["grad_norm"] = grad_norm
logs["learning_rate"] = self._get_learning_rate()

self._total_loss_scalar += tr_loss_scalar
Expand Down
Loading