Skip to content

Add dynamic loss balancing for distillation feature and logit losses #4220

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

Open
ajkv-google wants to merge 1 commit into
base: main
Choose a base branch
from
+18 −5
Open

Add dynamic loss balancing for distillation feature and logit losses #4220

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
23 changes: 18 additions & 5 deletions src/maxtext/trainers/post_train/distillation/distillation_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -543,11 +543,10 @@ def compute_loss(

ce_teacher_sum = jnp.sum(ce_teacher_per_pos * mask)

base_logit_loss = (alpha * soft_loss_mean) + ((1.0 - alpha) * hard_loss_mean)

# Feature Loss Scaling (match Logit Loss magnitude)
feature_loss = jnp.array(0.0, dtype=jnp.float32)

if self.beta_feature > 0.0:
assert s_features is not None and t_features is not None
if self.layer_indices is not None:
s_features_sliced = jnp.take(s_features, self.layer_indices, axis=0)
t_features_sliced = jnp.take(t_features, self.layer_indices, axis=0)
Expand All @@ -558,9 +557,23 @@ def compute_loss(
s_features_sliced = s_features_sliced.astype(jnp.float32)
t_features_sliced = t_features_sliced.astype(jnp.float32)

feature_loss = beta_feature * self.feature_loss_fn(s_features_sliced, t_features_sliced, mask)
# Calculate raw feature loss
raw_feature_loss = self.feature_loss_fn(s_features_sliced, t_features_sliced, mask)

# Match magnitude: Scale raw feature loss to match soft loss mean
feature_mag_scale = jax.lax.stop_gradient(soft_loss_mean) / jnp.maximum(
jax.lax.stop_gradient(raw_feature_loss), 1e-8
)
feature_loss = raw_feature_loss * feature_mag_scale

# Combined Distillation Loss (Logits + Balanced Features)
kd_loss = soft_loss_mean + feature_loss

# KD Scaling (match Hard/Task Loss magnitude)
kd_mag_scale = jax.lax.stop_gradient(hard_loss_mean) / jnp.maximum(jax.lax.stop_gradient(kd_loss), 1e-8)
balanced_kd_loss = kd_loss * kd_mag_scale

total_loss = base_logit_loss + feature_loss
total_loss = hard_loss_mean + (alpha * balanced_kd_loss)

moe_lb_loss = jnp.array(0.0)
if student_output.moe_lb_loss is not None:
Expand Down
Loading