Restore device runtime scales for compiled adapters (#963)

Co-authored-by: Rydén Johan <johan.ryden@bostad.uppsala.se>
This commit is contained in:
ZironicandRydén Johan authored and GitHub committed 2026-07-18 08:19:57 -06:00
1 parent cfdc9033a6
commit a0224793ce
7 files changed
+202 -16

No files matched your search

+3 -3
View File
@@ -212,7 +212,7 @@ class LokrModule(ToolkitModuleMixin, nn.Module):
if self.use_w2 and self.use_w1:
# use scale = 1
alpha = lora_dim
self.scale = float(alpha) / self.lora_dim
self._set_runtime_scale(float(alpha) / self.lora_dim)
self.register_buffer('alpha', torch.tensor(alpha)) # treat as constant
if self.use_w2:
@@ -251,7 +251,7 @@ class LokrModule(ToolkitModuleMixin, nn.Module):
(self.lokr_w2 if self.use_w2
else make_weight_cp(self.lokr_t2, self.lokr_w2_a, self.lokr_w2_b) if self.cp
else self.lokr_w2_a@self.lokr_w2_b),
self.scale
self._runtime_scale
)
if orig_weight is not None:
weight = weight.reshape(orig_weight.shape)
@@ -382,7 +382,7 @@ class LokrModule(ToolkitModuleMixin, nn.Module):
# scale folded into A (not applied to the reduction output) to avoid
# an inductor lowering bug under torch.compile
delta = torch.einsum('...qo,pq->...po', tmp, A * self.scale) # (..., out_l, out_k)
delta = torch.einsum('...qo,pq->...po', tmp, A * self._runtime_scale) # (..., out_l, out_k)
delta = delta.flatten(-2, -1)
if self.training and self.rank_dropout: