Skip to content

Fix QK norm reshape to match RMSNorm's expected 3D input shape #1221

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
brendanlong wants to merge 1 commit into
base: dev
Choose a base branch
from
Open

Fix QK norm reshape to match RMSNorm's expected 3D input shape #1221

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
9 changes: 5 additions & 4 deletions transformer_lens/components/abstract_attention.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -357,10 +357,11 @@ def _apply_qk_norm(
Returns:
Normalized tensor with same shape as input
"""
# Reshape from [batch, pos, head_index, d_head] to [batch * pos * head_index, d_head]
d_head = x.shape[-1]
x_normed = norm_module(x.reshape(-1, d_head))
return x_normed.reshape(x.shape)
# Reshape from [batch, pos, head_index, d_head] to [batch * pos, head_index, d_head]
# so it matches RMSNorm's expected 3D input shape (batch, pos, length)
batch, pos, n_heads, d_head = x.shape
x_normed = norm_module(x.reshape(batch * pos, n_heads, d_head))
return x_normed.reshape(batch, pos, n_heads, d_head)

def calculate_qkv_matrices(
self,
Expand Down
5 changes: 3 additions & 2 deletions transformer_lens/components/grouped_query_attention.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -197,8 +197,9 @@ def _apply_qk_norm(
Returns:
Normalized tensor with same shape as input
"""
# Reshape from [batch, pos, head_index, d_head] to [batch * pos * head_index, d_head]
# Reshape from [batch, pos, head_index, d_head] to [batch * pos, head_index, d_head]
# so it matches RMSNorm's expected 3D input shape (batch, pos, length)
batch, pos, n_heads, d_head = x.shape
x_reshaped = x.reshape(-1, d_head)
x_reshaped = x.reshape(batch * pos, n_heads, d_head)
x_normed = norm_module(x_reshaped)
return x_normed.reshape(batch, pos, n_heads, d_head)
Loading