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

Jump to bottom

[plugin] add cast inputs option for zero #6003

Merged
merged 1 commit into from
Aug 15, 2024
Merged

[plugin] add cast inputs option for zero #6003

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
13 changes: 10 additions & 3 deletions colossalai/booster/plugin/low_level_zero_plugin.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -62,7 +62,9 @@ class OptimizerParamCheckState(enum.Enum):


class LowLevelZeroModel(ModelWrapper, AMPModelMixin):
def __init__(self, module: nn.Module, precision: str, overlap_allgather: bool = False) -> None:
def __init__(
self, module: nn.Module, precision: str, overlap_allgather: bool = False, cast_inputs: bool = True
) -> None:
super().__init__(module)
self.dtype = None
if precision == "fp16":
Expand All @@ -74,7 +76,7 @@ def __init__(self, module: nn.Module, precision: str, overlap_allgather: bool =
module = module.to(get_accelerator().get_current_device())
self.module = module
self.convert_fn = None
if self.dtype is not None:
if self.dtype is not None and cast_inputs:
self.convert_fn = partial(_convert_floating_point, dtype=self.dtype)
self.overlap_allgather = overlap_allgather
if overlap_allgather:
Expand Down Expand Up @@ -334,6 +336,7 @@ def __init__(
cpu_offload: bool = False,
master_weights: bool = True,
verbose: bool = False,
cast_inputs: bool = True,
) -> None:
super().__init__()
assert stage in (1, 2), f"LowLevelZeroPlugin only supports stage 1/2 training"
Expand All @@ -360,6 +363,7 @@ def __init__(
)
self.lora_enabled = False
self.verbose = verbose
self.cast_inputs = cast_inputs

# set class name with stage, for better error message
setattr(self.__class__, "__name__", f"LowLevelZeroPlugin_ZeRO-{stage}")
Expand Down Expand Up @@ -474,7 +478,10 @@ def configure(

if not isinstance(model, ModelWrapper):
model = LowLevelZeroModel(
model, self.precision, overlap_allgather=self.zero_optim_kwargs["overlap_allgather"]
model,
self.precision,
overlap_allgather=self.zero_optim_kwargs["overlap_allgather"],
cast_inputs=self.cast_inputs,
)

# TODO: Support Galore + ZeRO
Expand Down
Loading