Skip to content

Fix duplicate peak learning rate in warmup schedule #3095

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
ChingTsai wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Fix duplicate peak learning rate in warmup schedule #3095

Show file tree
Hide file tree
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
2 changes: 1 addition & 1 deletion src/maxtext/utils/maxtext_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1376,7 +1376,7 @@ def schedule(step):
boundaries = []

if warmup_steps > 0:
warmup_schedule = optax.linear_schedule(init_value=0.0, end_value=lr, transition_steps=warmup_steps - 1)
warmup_schedule = optax.linear_schedule(init_value=0.0, end_value=lr, transition_steps=warmup_steps)
pieces.append(warmup_schedule)
boundaries.append(warmup_steps)

Expand Down
10 changes: 10 additions & 0 deletions tests/unit/maxtext_utils_test.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -748,6 +748,11 @@ def test_cosine_schedule(self):
# Warmup phase: 0 -> peak
self.assertAlmostEqual(float(schedule_fn(0)), 0.0, places=6)
self.assertAlmostEqual(float(schedule_fn(warmup_steps)), learning_rate, places=6)
# Ensure delta is constant
expected_slope = learning_rate / warmup_steps
for i in range(1, warmup_steps + 1):
current_lr = float(schedule_fn(i))
self.assertAlmostEqual(current_lr - float(schedule_fn(i - 1)), expected_slope, places=6)

# Cosine decay phase
lr_end = schedule_fn(learning_rate_schedule_steps - 1)
Expand Down Expand Up @@ -791,6 +796,11 @@ def test_wsd_schedule(self):
# Warmup phase: 0 -> peak
self.assertAlmostEqual(float(schedule_fn(0)), 0.0, places=6)
self.assertAlmostEqual(float(schedule_fn(warmup_steps)), learning_rate, places=6)
# Ensure delta is constant
expected_slope = learning_rate / warmup_steps
for i in range(1, warmup_steps + 1):
current_lr = float(schedule_fn(i))
self.assertAlmostEqual(current_lr - float(schedule_fn(i - 1)), expected_slope, places=6)

# Stable phase: constant at peak
self.assertAlmostEqual(float(schedule_fn(warmup_steps + 10)), learning_rate, places=6)
Expand Down
Loading