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

Enabling bias_dropout_add_fused with no bias term #9676

Closed
wants to merge 2 commits into from
Closed

Enabling bias_dropout_add_fused with no bias term #9676

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
6bf0f92
Enabling bias_dropout_add_fused with no bias term
ashvinnihalani Jul 10, 2024
021cebe
Apply isort and black reformatting
ashvinnihalani Jul 10, 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
17 changes: 14 additions & 3 deletions nemo/collections/nlp/modules/common/megatron/fused_bias_dropout_add.py
View file
Open in desktop
Copy link

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Should we also update bias_dropout_add_fused_inference_ to check if bias exists? otherwise it seems the type would mismatch when doing x + bias

Original file line number Diff line number Diff line change
Expand Up @@ -48,12 +48,23 @@ def bias_dropout_add_fused_train_(
return bias_dropout_add(x, bias, residual, prob, True)


@torch.jit.script
def dropout_add_fused_train_(x: torch.Tensor, bias: torch.Tensor, residual: torch.Tensor, prob: float) -> torch.Tensor:
# type: (Tensor, None, Tensor, float) -> Tensor
return dropout_add(x, bias, residual, prob, True)


def bias_dropout_add_fused_train(x, bias, residual, prob):
# re-enable torch grad to enable fused optimization.
with torch.enable_grad():
args = _cast_if_autocast_enabled(x, bias, residual, prob)
with torch.cuda.amp.autocast(enabled=False):
return bias_dropout_add_fused_train_(*args)
if bias:
args = _cast_if_autocast_enabled(x, bias, residual, prob)
with torch.cuda.amp.autocast(enabled=False):
return bias_dropout_add_fused_train_(*args)
else:
args = _cast_if_autocast_enabled(x, residual, prob)
with torch.cuda.amp.autocast(enabled=False):
return dropout_add_fused_train_(*args)


@torch.jit.script
Expand Down
2 changes: 1 addition & 1 deletion nemo/collections/nlp/modules/common/megatron/transformer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -453,7 +453,7 @@ def _get_bias_droput_add_func(self, transformer_block_type='pre_ln', position_af
if transformer_block_type == 'normformer' and position_after == 'attention':
bias_dropout_add_func = get_dropout_add(self.training)
# Bias dropout add fused kernel
elif self.bias and self.bias_dropout_add_fusion:
elif self.bias_dropout_add_fusion:
if self.training:
bias_dropout_add_func = bias_dropout_add_fused_train
else:
Expand Down
Loading