Skip to content

Fix the bug of block attention residuals in AMP #719

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

Open
pkuzyc wants to merge 1 commit into
base: develop
Choose a base branch
from
+8 −0
Open

Fix the bug of block attention residuals in AMP #719

Show file tree
Hide file tree
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
3 changes: 3 additions & 0 deletions src/paddlefleet/transformer/block_attn_res.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -132,4 +132,7 @@ def forward(self, partial_block: Tensor, blocks: list[Tensor]) -> Tensor:
# Equivalent to einsum("n b s, n b s d -> b s d", weights, V)
h = (weights.unsqueeze(-1) * V).sum(axis=0)

if partial_block is not None and h.dtype != partial_block.dtype:
h = h.to(partial_block.dtype)

return h
5 changes: 5 additions & 0 deletions src/paddlefleet/transformer/transformer_layer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -568,6 +568,11 @@ def _forward_impl(
)

# Accumulate attn output into partial_block
if (
partial_block is not None
and partial_block.dtype != hidden_states.dtype
):
partial_block = partial_block.to(hidden_states.dtype)
partial_block = (
partial_block + hidden_states
if partial_block is not None
Expand Down
Loading