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

change llama/modeling.py to opt npu performence #8342

Merged
merged 7 commits into from
Apr 30, 2024
Merged

change llama/modeling.py to opt npu performence #8342

Changes from 1 commit
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
5db1b03
change llama/modeling.py to opt npu performence
wanghuancoder Apr 28, 2024
48a1338
update
wanghuancoder Apr 28, 2024
11a52e0
update
wanghuancoder Apr 28, 2024
f2252e8
Update modeling.py
Galaxy1458 Apr 29, 2024
198e58e
Merge branch 'develop' of https://github.com/PaddlePaddle/PaddleNLP i…
wanghuancoder Apr 30, 2024
e557b5a
Merge branch 'develop' of https://github.com/Galaxy1458/PaddleNLP int…
wanghuancoder Apr 30, 2024
672f4aa
add judge
wanghuancoder Apr 30, 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: 5 additions & 3 deletions paddlenlp/transformers/llama/modeling.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -246,7 +246,7 @@ def scaled_dot_product_attention(
attention_mask is None,
True,
False,
False,
is_casual_mask(attention_mask),
)[0]
else:
attn_output = F.scaled_dot_product_attention(
Expand Down Expand Up @@ -1704,8 +1704,10 @@ def forward(self, prediction_scores, masked_lm_labels):
_hcg = fleet.get_hybrid_communicate_group()
masked_lm_loss = ConcatSePMaskedLoss.apply(masked_lm_loss, axis=1, group=_hcg.get_sep_parallel_group())
# skip ignore_index which loss == 0
masked_lm_loss = masked_lm_loss[masked_lm_loss > 0]
loss = paddle.mean(masked_lm_loss)
# masked_lm_loss = masked_lm_loss[masked_lm_loss > 0]
# loss = paddle.mean(masked_lm_loss)
binary_sequence = paddle.where(masked_lm_loss > 0, paddle.ones_like(masked_lm_loss), paddle.zeros_like(masked_lm_loss))
loss = paddle.sum(masked_lm_loss * binary_sequence) / paddle.sum(binary_sequence)
Copy link
Collaborator

@wawltor wawltor Apr 29, 2024 •
edited
Loading

Choose a reason for hiding this comment

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

paddle.sum(binary_sequence) 中binary_sequence有可能全为0，导致sum后0；这里loss会有异常

之前的业务遇到过这个问题

Copy link
Collaborator

Choose a reason for hiding this comment

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

同时对GPU性能的影响是什么


return loss

Expand Down
Loading