Skip to content

fix(gqa): replace float division with integer division in GQA reshape #668

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
Sumu004 wants to merge 1 commit into
base: main
Choose a base branch
from
Open

fix(gqa): replace float division with integer division in GQA reshape #668

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
4 changes: 2 additions & 2 deletions gemma/gm/nn/_modules.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -241,7 +241,7 @@ def __call__(
# Reshape matrices to enable einsums over groups.
b, t, kg, h = query_scaled.shape
query_scaled = query_scaled.reshape(
(b, t, self.num_kv_heads, int(kg / self.num_kv_heads), h)
(b, t, self.num_kv_heads, kg // self.num_kv_heads, h)
)
logits = jnp.einsum('BTKGH,BSKH->BTKGS', query_scaled, key_proj)
b, t, k, g, s = logits.shape
Expand Down Expand Up @@ -285,7 +285,7 @@ def __call__(
# Reshape matrices to enable einsums over groups.
b, t, kg, h = probs.shape
probs = probs.reshape(
(b, t, self.num_kv_heads, int(kg / self.num_kv_heads), h)
(b, t, self.num_kv_heads, kg // self.num_kv_heads, h)
)
encoded = jnp.einsum('BTKGS,BSKH->BTKGH', probs, value_proj)
b, t, k, g, h = encoded.shape
Expand Down
4 changes: 2 additions & 2 deletions gemma/gm/nn/gemma3n/_modules.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -339,7 +339,7 @@ def __call__(
# Reshape matrices to enable einsums over groups.
b, t, kg, h = query_scaled.shape
query_scaled = query_scaled.reshape(
(b, t, self.num_kv_heads, int(kg / self.num_kv_heads), h)
(b, t, self.num_kv_heads, kg // self.num_kv_heads, h)
)
logits = jnp.einsum('BTKGH,BSKH->BTKGS', query_scaled, key_proj)
b, t, k, g, s = logits.shape
Expand Down Expand Up @@ -384,7 +384,7 @@ def __call__(
# Reshape matrices to enable einsums over groups.
b, t, kg, h = probs.shape
probs = probs.reshape(
(b, t, self.num_kv_heads, int(kg / self.num_kv_heads), h)
(b, t, self.num_kv_heads, kg // self.num_kv_heads, h)
)
encoded = jnp.einsum('BTKGS,BSKH->BTKGH', probs, value_proj)
b, t, k, g, h = encoded.shape
Expand Down
4 changes: 2 additions & 2 deletions gemma/gm/nn/gemma4/_modules.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -323,7 +323,7 @@ def __call__(
# Reshape matrices to enable einsums over groups.
b, t, kg, h = query_proj.shape
query_proj = query_proj.reshape(
(b, t, self.num_kv_heads, int(kg / self.num_kv_heads), h)
(b, t, self.num_kv_heads, kg // self.num_kv_heads, h)
)
logits = jnp.einsum('BTKGH,BSKH->BTKGS', query_proj, key_proj)
b, t, k, g, s = logits.shape
Expand Down Expand Up @@ -362,7 +362,7 @@ def __call__(
# Reshape matrices to enable einsums over groups.
b, t, kg, h = probs.shape
probs = probs.reshape(
(b, t, self.num_kv_heads, int(kg / self.num_kv_heads), h)
(b, t, self.num_kv_heads, kg // self.num_kv_heads, h)
)
encoded = jnp.einsum('BTKGS,BSKH->BTKGH', probs, value_proj)
b, t, k, g, h = encoded.shape
Expand Down
4 changes: 2 additions & 2 deletions gemma/research/t5gemma/modules.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -237,7 +237,7 @@ def __call__(
# Reshape matrices to enable einsums over groups.
b, t, kg, h = query_scaled.shape
query_scaled = query_scaled.reshape(
(b, t, self.num_kv_heads, int(kg / self.num_kv_heads), h)
(b, t, self.num_kv_heads, kg // self.num_kv_heads, h)
)
logits = jnp.einsum('BTKGH,BSKH->BTKGS', query_scaled, key_proj)
b, t, k, g, s = logits.shape
Expand Down Expand Up @@ -269,7 +269,7 @@ def __call__(
# Reshape matrices to enable einsums over groups.
b, t, kg, h = probs.shape
probs = probs.reshape(
(b, t, self.num_kv_heads, int(kg / self.num_kv_heads), h)
(b, t, self.num_kv_heads, kg // self.num_kv_heads, h)
)
encoded = jnp.einsum('BTKGS,BSKH->BTKGH', probs, value_proj)
b, t, k, g, h = encoded.shape
Expand Down