Skip to content

Use correct DoRA implementation #15283

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
AUTOMATIC1111 merged 3 commits into from
Mar 16, 2024
Merged

Use correct DoRA implementation #15283

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
1792e19
Use correct implementation, fix device error
KohakuBlueleaf Mar 16, 2024
199c51d
linter
KohakuBlueleaf Mar 16, 2024
8dcb8fa
Merge branch 'dev' into dora-weight-decompose
AUTOMATIC1111 Mar 16, 2024
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
18 changes: 15 additions & 3 deletions extensions-builtin/Lora/network.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -153,7 +153,7 @@ def __init__(self, net: Network, weights: NetworkWeights):
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]:
Expand All @@ -170,10 +170,22 @@ def calc_scale(self):
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
Expand Down