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

prepare deepspeed accomodate fp16 and bf16 #1728

Merged
merged 2 commits into from
Jun 17, 2024
Merged

prepare deepspeed accomodate fp16 and bf16 #1728

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
9e11707
prepare deepspeed accomodate fp16 and bf16
mnoukhov Jun 13, 2024
6767799
precommit
mnoukhov Jun 13, 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
8 changes: 6 additions & 2 deletions trl/trainer/ppov2_trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -190,8 +190,12 @@ def __init__(
self.eval_dataloader = accelerator.prepare(self.eval_dataloader)

if self.is_deepspeed_enabled:
self.reward_model = prepare_deepspeed(self.reward_model, args.per_device_train_batch_size)
self.ref_policy = prepare_deepspeed(self.ref_policy, args.per_device_train_batch_size)
self.reward_model = prepare_deepspeed(
self.reward_model, args.per_device_train_batch_size, args.bf16, args.fp16
)
self.ref_policy = prepare_deepspeed(
self.ref_policy, args.per_device_train_batch_size, args.bf16, args.fp16
)
else:
self.ref_policy = self.ref_policy.to(self.accelerator.device)
self.reward_model = self.reward_model.to(self.accelerator.device)
Expand Down
8 changes: 6 additions & 2 deletions trl/trainer/rloo_trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -175,8 +175,12 @@ def __init__(
self.eval_dataloader = accelerator.prepare(self.eval_dataloader)

if self.is_deepspeed_enabled:
self.reward_model = prepare_deepspeed(self.reward_model, args.per_device_train_batch_size)
self.ref_policy = prepare_deepspeed(self.ref_policy, args.per_device_train_batch_size)
self.reward_model = prepare_deepspeed(
self.reward_model, args.per_device_train_batch_size, args.bf16, args.fp16
)
self.ref_policy = prepare_deepspeed(
self.ref_policy, args.per_device_train_batch_size, args.bf16, args.fp16
)
self.deepspeed = self.model
else:
self.ref_policy = self.ref_policy.to(self.accelerator.device)
Expand Down
9 changes: 7 additions & 2 deletions trl/trainer/utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1001,7 +1001,9 @@ def forward(
)


def prepare_deepspeed(model: torch.nn.Module, per_device_train_batch_size: int):
def prepare_deepspeed(
model: torch.nn.Module, per_device_train_batch_size: int, fp16: bool = False, bf16: bool = False
):
"""
Prepares the model for training with DeepSpeed (both for stage 2 and 3), configuring the appropriate settings based on the model and
batch size.
Expand All @@ -1024,10 +1026,13 @@ def prepare_deepspeed(model: torch.nn.Module, per_device_train_batch_size: int):
config_kwargs["train_micro_batch_size_per_gpu"] = per_device_train_batch_size
config_kwargs = {
"train_micro_batch_size_per_gpu": config_kwargs["train_micro_batch_size_per_gpu"],
"bf16": {"enabled": True},
"prescale_gradients": False,
"wall_clock_breakdown": False,
}
if bf16:
config_kwargs["bf16"] = {"enabled": True}
elif fp16:
config_kwargs["fp16"] = {"enabled": True}
else:
if hasattr(model, "config"):
hidden_size = (
Expand Down
Loading