Merge pull request #16151 from AUTOMATIC1111/dora-fix
Possible fix of wrong scale in weight decomposition
This commit is contained in:
commit
372a8e0658
|
@ -204,10 +204,12 @@ class NetworkModule:
|
||||||
if ex_bias is not None:
|
if ex_bias is not None:
|
||||||
ex_bias = ex_bias * self.multiplier()
|
ex_bias = ex_bias * self.multiplier()
|
||||||
|
|
||||||
|
updown = updown * self.calc_scale()
|
||||||
|
|
||||||
if self.dora_scale is not None:
|
if self.dora_scale is not None:
|
||||||
updown = self.apply_weight_decompose(updown, orig_weight)
|
updown = self.apply_weight_decompose(updown, orig_weight)
|
||||||
|
|
||||||
return updown * self.calc_scale() * self.multiplier(), ex_bias
|
return updown * self.multiplier(), ex_bias
|
||||||
|
|
||||||
def calc_updown(self, target):
|
def calc_updown(self, target):
|
||||||
raise NotImplementedError()
|
raise NotImplementedError()
|
||||||
|
|
Loading…
Reference in New Issue