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

Support adamw_torch_8bit #34993

Merged
merged 3 commits into from
Jan 21, 2025
Merged

Support adamw_torch_8bit #34993

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
83916ed
var
fzyzcjy Nov 28, 2024
d888a8f
more
fzyzcjy Nov 28, 2024
7898a18
test
fzyzcjy Nov 28, 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
14 changes: 11 additions & 3 deletions src/transformers/trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1606,7 +1606,10 @@ def optimizer_hook(param):
"gradient_clipping": float(optim_args.get("gradient_clipping", 1.0)),
}
)
elif args.optim == OptimizerNames.ADAMW_TORCH_4BIT:
elif args.optim in [
OptimizerNames.ADAMW_TORCH_4BIT,
OptimizerNames.ADAMW_TORCH_8BIT,
]:
if not is_torchao_available() or version.parse(importlib.metadata.version("torchao")) < version.parse(
"0.4.0"
):
Expand All @@ -1619,9 +1622,14 @@ def optimizer_hook(param):
"You need to have `torch>2.4` in order to use torch 4-bit optimizers. "
"Install it with `pip install --upgrade torch` it is available on pipy. Otherwise, you need to install torch nightly."
)
from torchao.prototype.low_bit_optim import AdamW4bit
from torchao.prototype.low_bit_optim import AdamW4bit, AdamW8bit

optimizer_cls = AdamW4bit
if args.optim == OptimizerNames.ADAMW_TORCH_4BIT:
optimizer_cls = AdamW4bit
elif args.optim == OptimizerNames.ADAMW_TORCH_8BIT:
optimizer_cls = AdamW8bit
else:
raise ValueError("Invalid optimizer")
optimizer_kwargs.update(adam_kwargs)
elif args.optim in [
OptimizerNames.SCHEDULE_FREE_ADAMW,
Expand Down
1 change: 1 addition & 0 deletions src/transformers/training_args.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -156,6 +156,7 @@ class OptimizerNames(ExplicitEnum):
ADAFACTOR = "adafactor"
ADAMW_ANYPRECISION = "adamw_anyprecision"
ADAMW_TORCH_4BIT = "adamw_torch_4bit"
ADAMW_TORCH_8BIT = "adamw_torch_8bit"
ADEMAMIX = "ademamix"
SGD = "sgd"
ADAGRAD = "adagrad"
Expand Down
7 changes: 7 additions & 0 deletions tests/trainer/test_trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4741,6 +4741,13 @@ def hp_name(trial):
default_adam_kwargs,
)
)
optim_test_params.append(
(
TrainingArguments(optim=OptimizerNames.ADAMW_TORCH_8BIT, output_dir="None"),
torchao.prototype.low_bit_optim.AdamW8bit,
default_adam_kwargs,
)
)


@require_torch
Expand Down