Skip to content

Conduct lr check only once #950

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
tianyu-l merged 7 commits into from
Mar 14, 2025
Merged

Conduct lr check only once #950

Changes from all commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
89b68db
[Scheduler] Ensure `warmup_stable_steps` is an integer
yzhangcs Mar 11, 2025
9f99a41
Conduct lr check only once
yzhangcs Mar 11, 2025
1127c74
Delete `decay_ratio` checks in config manager
yzhangcs Mar 12, 2025
5f189ed
Raise warning only once for abnormal decay ratio
yzhangcs Mar 12, 2025
a8ff29e
Pass warm/stable/decay steps as args of `linear_warmup_stable_decay`
yzhangcs Mar 13, 2025
55e76a1
Revert changes in logging
yzhangcs Mar 13, 2025
1994c63
Fix code formats
yzhangcs Mar 14, 2025
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
32 changes: 18 additions & 14 deletions torchtitan/components/optimizer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -7,7 +7,7 @@
import copy
import functools
import math
from typing import Any, Callable, Dict, Generic, List, TypeVar, Union
from typing import Any, Callable, Dict, Generic, List, TypeVar

import torch
import torch.nn as nn
Expand Down Expand Up @@ -392,14 +392,26 @@ def build_lr_schedulers(
"""
training_steps = job_config.training.steps
warmup_steps = int(job_config.lr_scheduler.warmup_steps)
lr_decay_ratio = job_config.lr_scheduler.decay_ratio
if job_config.lr_scheduler.decay_ratio is not None:
decay_steps = round(training_steps * job_config.lr_scheduler.decay_ratio)
if warmup_steps + decay_steps > training_steps:
logger.warning(
f"Warmup ({warmup_steps}) + decay ({decay_steps}) steps exceed "
f"total training steps ({training_steps}). "
f"Adjusting decay steps to {training_steps - warmup_steps}."
)
decay_steps = training_steps - warmup_steps
else:
decay_steps = training_steps - warmup_steps
stable_steps = training_steps - warmup_steps - decay_steps
lr_decay_type = job_config.lr_scheduler.decay_type
lr_min = job_config.lr_scheduler.lr_min

def linear_warmup_stable_decay(
current_step: int,
warmup_steps: int,
lr_decay_ratio: Union[float, None],
stable_steps: int,
decay_steps: int,
lr_decay_type: str,
lr_min: float,
):
Expand All @@ -419,15 +431,7 @@ def linear_warmup_stable_decay(
If `lr_min` is specified, the decay range is scaled from 1 to `lr_min`
to ensure the learning rate does not drop below this minimum value.
"""
if lr_decay_ratio is None:
warmup_stable_steps = warmup_steps
else:
warmup_stable_steps = training_steps * (1 - lr_decay_ratio)
if warmup_stable_steps < warmup_steps:
logger.warning(
f"The warmup steps should be less than or equal to the warmup-stable steps ({warmup_stable_steps}). "
f"Consider reducing either the decay ratio ({lr_decay_ratio}) or the warmup steps ({warmup_steps})."
)
warmup_stable_steps = warmup_steps + stable_steps
if current_step < warmup_steps:
# linear warmup
# 0-indexed step, hence + 1 adjustments
Expand All @@ -436,7 +440,6 @@ def linear_warmup_stable_decay(
elif current_step < warmup_stable_steps:
curr_adjustment = 1.0
else:
decay_steps = float(max(1, training_steps - warmup_stable_steps))
progress = float(current_step - warmup_stable_steps) / decay_steps

if lr_decay_type == "linear":
Expand All @@ -451,7 +454,8 @@ def linear_warmup_stable_decay(
lr_lambda = functools.partial(
linear_warmup_stable_decay,
warmup_steps=warmup_steps,
lr_decay_ratio=lr_decay_ratio,
stable_steps=stable_steps,
decay_steps=decay_steps,
lr_decay_type=lr_decay_type,
lr_min=lr_min,
)
Expand Down