Skip to content

Extract _lora_call into a shared free function in lora.py (#20306) #20306

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
meta-codesync merged 1 commit into from
Jun 17, 2026
+15 −14
Merged

Extract _lora_call into a shared free function in lora.py (#20306) #20306

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: 9 additions & 0 deletions examples/models/llama/lora.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -69,3 +69,12 @@ def forward(
z = self.lora_a(self.dropout(x))
z = (self.alpha / self.rank) * self.lora_b(z)
return out + z


def lora_call(linear, x_in, lora_blob):
if lora_blob is not None:
key = getattr(linear, "_lora_key", None)
if key is not None and key in lora_blob:
a, b = lora_blob[key]
return linear(x_in, a, b)
return linear(x_in)
20 changes: 6 additions & 14 deletions examples/models/llama/static_attention.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -13,7 +13,7 @@
ForwardOptions,
register_attention,
)
from executorch.examples.models.llama.lora import LoRALinear
from executorch.examples.models.llama.lora import lora_call, LoRALinear
from executorch.examples.models.llama.model_args import ModelArgs
from executorch.examples.models.llama.norm import ScalelessRMSNorm
from executorch.examples.models.llama.rope import Rope
Expand Down Expand Up @@ -1014,14 +1014,6 @@ def from_attention_mha(

return instance

def _lora_call(self, linear, x_in, lora_blob):
if lora_blob is not None:
key = getattr(linear, "_lora_key", None)
if key is not None and key in lora_blob:
a, b = lora_blob[key]
return linear(x_in, a, b)
return linear(x_in)

def forward(
self,
x: torch.Tensor,
Expand All @@ -1044,7 +1036,7 @@ def forward(
# Default behavior (no blob, or no `_lora_key`) is unchanged.
_lora_blob = kwargs.get("__lora_io_blob__")

new_qs = [self._lora_call(wq, x, _lora_blob) for wq in self.wqs]
new_qs = [lora_call(wq, x, _lora_blob) for wq in self.wqs]

shared_kv = kwargs.get("shared_kv")
if shared_kv is not None:
Expand All @@ -1054,8 +1046,8 @@ def forward(
new_ks = []
new_vs = []
else:
new_ks = [self._lora_call(wk, x, _lora_blob) for wk in self.wks]
new_vs = [self._lora_call(wv, x, _lora_blob) for wv in self.wvs]
new_ks = [lora_call(wk, x, _lora_blob) for wk in self.wks]
new_vs = [lora_call(wv, x, _lora_blob) for wv in self.wvs]

if self.use_conv2d:

Expand Down Expand Up @@ -1092,7 +1084,7 @@ def from_conv2ds(ts):

if self.use_conv2d:
y = (
self._lora_call(
lora_call(
self.wo,
y.reshape(bsz, -1, 1, self.n_heads * self.head_dim).transpose(1, 3),
_lora_blob,
Expand All @@ -1101,7 +1093,7 @@ def from_conv2ds(ts):
.reshape(bsz, -1, self.dim)
)
else:
y = self._lora_call(self.wo, y, _lora_blob)
y = lora_call(self.wo, y, _lora_blob)

update = {"out_cache_state": out_cache_state}
if kv_to_share is not None:
Expand Down
Loading