Merge pull request #15283 from AUTOMATIC1111/dora-weight-decompose
Use correct DoRA implementation
This commit is contained in:
commit
df8c09bcb3
|
@ -153,7 +153,7 @@ class NetworkModule:
|
|||
self.scale = weights.w["scale"].item() if "scale" in weights.w else None
|
||||
|
||||
self.dora_scale = weights.w.get("dora_scale", None)
|
||||
self.dora_mean_dim = tuple(i for i in range(len(self.shape)) if i != 1)
|
||||
self.dora_norm_dims = len(self.shape) - 1
|
||||
|
||||
def multiplier(self):
|
||||
if 'transformer' in self.sd_key[:20]:
|
||||
|
@ -170,10 +170,22 @@ class NetworkModule:
|
|||
return 1.0
|
||||
|
||||
def apply_weight_decompose(self, updown, orig_weight):
|
||||
orig_weight = orig_weight.to(updown)
|
||||
# Match the device/dtype
|
||||
orig_weight = orig_weight.to(updown.dtype)
|
||||
dora_scale = self.dora_scale.to(device=orig_weight.device, dtype=updown.dtype)
|
||||
updown = updown.to(orig_weight.device)
|
||||
|
||||
merged_scale1 = updown + orig_weight
|
||||
merged_scale1_norm = (
|
||||
merged_scale1.transpose(0, 1)
|
||||
.reshape(merged_scale1.shape[1], -1)
|
||||
.norm(dim=1, keepdim=True)
|
||||
.reshape(merged_scale1.shape[1], *[1] * self.dora_norm_dims)
|
||||
.transpose(0, 1)
|
||||
)
|
||||
|
||||
dora_merged = (
|
||||
merged_scale1 / merged_scale1.mean(dim=self.dora_mean_dim, keepdim=True) * self.dora_scale
|
||||
merged_scale1 * (dora_scale / merged_scale1_norm)
|
||||
)
|
||||
final_updown = dora_merged - orig_weight
|
||||
return final_updown
|
||||
|
|
Loading…
Reference in New Issue