Skip to content

Fix config.softmax_scale not being considered #3698

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
janEbert wants to merge 1 commit into
base: main
Choose a base branch
from
+11 −1
Open

Fix config.softmax_scale not being considered #3698

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
2 changes: 2 additions & 0 deletions megatron/core/transformer/attention.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -803,6 +803,8 @@ def flash_decode_and_prefill(
q = q.squeeze(1)
if getattr(self, "softmax_scale", None) is not None:
softmax_scale = self.softmax_scale
elif self.config.softmax_scale is not None:
softmax_scale = self.config.softmax_scale
else:
softmax_scale = q.shape[-1] ** -0.5
if HAVE_FA3:
Expand Down
10 changes: 9 additions & 1 deletion megatron/core/transformer/multi_latent_attention.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -125,7 +125,15 @@ def __init__(
self.qkv_up_checkpoint = None

mscale = _yarn_get_mscale(self.config.rotary_scaling_factor, self.config.mscale_all_dim)
self.softmax_scale = mscale * mscale / math.sqrt(self.q_head_dim)
self.softmax_scale = (
mscale
* mscale
* (
(1 / math.sqrt(self.q_head_dim))
if self.config.softmax_scale is None
else self.config.softmax_scale
)
)
self.cache_mla_latents = self.config.cache_mla_latents

if self.config.rope_type == "rope":
Expand Down
Loading