Skip to content

Force blksize=128 when head_dim=256 on ascend #4723

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
wanfengcxz wants to merge 1 commit into
base: main
Choose a base branch
from
+3 −4
Open

Force blksize=128 when head_dim=256 on ascend #4723

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
7 changes: 3 additions & 4 deletions lmdeploy/pytorch/engine/executor/base.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -239,16 +239,15 @@ def _adjust_block_size(self):
if self.cache_config.block_size != 64:
raise ValueError('Please set block_size to 64 for flash_mla.')
return
# Linear attention requires a kv block size of 128 on ascend.
# head_dim=256 requires block_size=128 on ascend.
# Other models keep the user-provided block size.
is_ssm = len(self.model_config.states_shapes) > 0
if (self.cache_config.device_type == 'ascend' and is_ssm and
if (self.cache_config.device_type == 'ascend' and self.model_config.k_head_dim == 256 and
(self.cache_config.block_size != 128 or self.cache_config.kernel_block_size != 128)):
logger.warning(
'Force `block_size=128` and `kernel_block_size=128` '
f'(was block_size={self.cache_config.block_size}, '
f'kernel_block_size={self.cache_config.kernel_block_size}) '
'for linear attention on ascend.')
'for head_dim=256 on ascend.')
self.cache_config.block_size = 128
self.cache_config.kernel_block_size = 128
return
Expand Down
Loading