Skip to content

[Attention] Remove slow setattr in MLA #14769

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 2 commits into from
Mar 13, 2025
+7 −2
Merged

[Attention] Remove slow setattr in MLA #14769

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
483e95f
fix slowattr
LucasWilkinson Mar 13, 2025
3669abf
format
LucasWilkinson Mar 13, 2025
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
9 changes: 7 additions & 2 deletions vllm/model_executor/layers/rotary_embedding.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -161,8 +161,13 @@ def forward_cuda(
) -> Tuple[torch.Tensor, torch.Tensor]:
from vllm import _custom_ops as ops

self.cos_sin_cache = self.cos_sin_cache.to(query.device,
dtype=query.dtype)
# __setattr__ in nn.Module (called by `self.cos_sin_cache = ...`)
# is expensive, so avoid calling it if possible
if self.cos_sin_cache.device != query.device or \
self.cos_sin_cache.dtype != query.dtype:
Comment on lines +166 to +167
Copy link
Member

Choose a reason for hiding this comment

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

I have been curious about this for a long time. when will this condition be false? i.e. the query has different device / dtype from the cos_sin_cache .

self.cos_sin_cache = self.cos_sin_cache.to(query.device,
dtype=query.dtype)

# ops.rotary_embedding()/batched_rotary_embedding()
# are in-place operations that update the query and key tensors.
if offsets is not None:
Expand Down