Skip to content

Fix key_padding_mask when using attn_impl='flash' #163

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

Merged
merged 5 commits into from
Feb 14, 2023
Merged

Fix key_padding_mask when using attn_impl='flash' #163

Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
0a34f79
try1
abhi-mosaic Feb 14, 2023
5ea8570
try2
abhi-mosaic Feb 14, 2023
91f7f17
pyright, bugfix
abhi-mosaic Feb 14, 2023
54220b5
elif
abhi-mosaic Feb 14, 2023
b4728fe
add comment
abhi-mosaic Feb 14, 2023
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 examples/llm/src/mosaic_gpt.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -421,10 +421,17 @@ def forward(self,
seq_len=S,
key_padding_mask=key_padding_mask,
dtype=x.dtype)
if self.cfg.attn_impl == 'flash' and key_padding_mask is None:
# HazyResearch FlashMHA appears to use more memory when `key_padding_mask=None`
# in certain settings like MosaicGPT-7B. So we always provide a tensor.
# See https://github.com/mosaicml/examples/pull/163 for more details.
mod_key_padding_mask = torch.ones_like(input_ids, dtype=torch.bool)
elif self.cfg.attn_impl == 'triton':
mod_key_padding_mask = None
else:
mod_key_padding_mask = key_padding_mask
for block in self.transformer.blocks: # type: ignore
x = block(
x, None if self.cfg.attn_impl == 'triton' else key_padding_mask,
attn_mask)
x = block(x, mod_key_padding_mask, attn_mask)
x = self.transformer.ln_f(x) # type: ignore
# output embedding weight tied to input embedding
assert isinstance(self.transformer.wte, nn.Module) # pyright
Expand Down